Artificial Intelligence and Machine Learning in Cold Spray Additive Manufacturing: A Systematic Literature Review

Abstract

Due to its unique benefits over conventional subtractive manufacturing, additive manufacturing methods continue to attract interest in both academia and industry. One such method is called Cold Spray Additive Manufacturing (CSAM), a solid-state coating deposition technology to manufacture repair metallic components using a gas jet and powder particles. CSAM offers low heat input, stable phases, suitability for heat-sensitive substrates, and high deposition rates. However, persistent challenges include porosity control, geometric accuracy near edges and concavities, anisotropy, and cost sensitivities linked to gas selection and nozzle wear. Interdisciplinary research across manufacturing science, materials characterisation, robotics, control, artificial intelligence (AI), and machine learning (ML) is deployed to overcome these issues. ML supports quality prediction, inverse parameter design, in situ monitoring, and surrogate models that couple process physics with data. To demonstrate the impact of AI and ML on CSAM, this study presents a systematic literature review to identify, evaluate, and analyse published studies in this domain. The most relevant studies in the literature are analysed using keyword co-occurrence and clustering. Four themes were identified: design for CSAM, material analytics, real-time monitoring and defect analytics, and deposition and AI-enabled optimisation. Based on this synthesis, core challenges are identified as small and varied datasets, transfer and identifiability limits, and fragmented sensing. Main opportunities are outlined as physics-based surrogates, active learning, uncertainty-aware inversion, and cloud-edge control for reliable and adaptable ML use in CSAM. By systematically mapping the current landscape, this work provides a critical roadmap for researchers to target the most significant challenges and opportunities in applying AI/ML to industrialise CSAM.

1. Introduction

Design in additive manufacturing (AM) has evolved into physics-aware, co-designed workflows that couple digital design, process physics, and data-driven control. This builds on international standards that define AM’s layer-wise nature and its implications for tolerances, planning, and verification [1,2,3,4]. However, AM is a multidisciplinary field, combining mechanics, materials science, fluid and thermal physics, electronics, robotics, and computer science; therefore, quality depends on linked process conditions, parameters, and feedstock quality [2,5,6,7,8].

One such AM method that is receiving immense attention due to its benefits is called Cold Spray Additive Manufacturing (CSAM). It is a solid-state deposition technology that accelerates powders to supersonic velocities in heated gas and consolidates them by severe plastic deformation, oxide fracture, and mechanical interlocking without melting. Using de Laval or constant-diameter nozzles (CD) and gases such as N₂ and He (or blends), CSAM achieves dense deposits, phase retention, and compatibility with temperature-sensitive substrates [2,9]. Porosity control, geometric accuracy near edges and concavities, helium cost exposure, and nozzle wear remain limiting features [2,3,10]. Because CSAM is a continuous jetting process, dimensional control depends on path planning, spray angle, stand-off, traverse speed, and jet-footprint overlap. Bonding further requires exceeding material-dependent critical velocities that link feedstock state to nozzle gas conditions [5,6,11]. CSAM is multi-physics and multi-scale. Nozzle flow and particle acceleration set in-flight states; impact plasticity forms solid surfaces; track and layer build-up depend on kinematics and overlap. Final parts show porosity and residual stress. Physics-based models, computational fluid dynamics (CFD) for gas–particle flow and nozzle design, finite-element analysis for impact, analytical deposition maps, and geometric growth solvers, provide useful insights but are slow to compute and have only been tested in limited cases. This restricts wide use, reverse design, or real-time control [5,6,7,12].

Artificial intelligence (AI) and machine learning (ML) help address these gaps. Surrogate models speed up track/layer prediction and support parameter design as well as nozzle optimisation. Design filters estimate porosity or quality, vision and signal tools enable in-process defect detection, and physics-guided solvers combine analytical and CFD knowledge with flexible corrections [13,14,15,16]. Recent works show robot-mounted laser profilometry for layer-wise detect-and-correct, aero-acoustics for nozzle and feed anomalies, and thermal forecasting for feed-rate changes, pointing toward closed-loop control [17,18,19]. Most ML analysis for CSAM relies on small, varied datasets from single machines or mined from the literature. Random data splits overestimate accuracy and reduce generalisability. More robust evaluation is needed, using grouped or forward-looking splits, propagating uncertainty from reported ranges, and combining training across simulations and experiments [16,20,21]. Model transfer also breaks down under changes in methods, sensor, or material. Identifiability limits appear in inverse design (e.g., low sensitivity to stand-off), and purely data-driven one-shot predictors fail under new conditions or geometries [22,23,24]. Sensing is still fragmented, where thermal, sizing–particle tracking velocimetry (S-PTV), sound, and laser-scanning data are rarely combined into fast feedback systems [2,25,26].

To organise the landscape and provide a holistic overview of AI and ML on CSAM, a systematic literature review has been conducted in this study using structured searches and keyword clustering. ML and AI for CSAM are grouped into four main research domains: (RD1) design (printability, path planning, inverse mapping, nozzle optimisation); (RD2) material analysis (linking process to structure and properties); (RD3) real-time monitoring and defect detection (thermal, optical, sound, and geometric sensing); and (RD4) deposition modelling and AI-based optimisation (surrogates, inverse design, and workflow integration). Across these areas, four key strategies are highlighted: combined physics-and-data surrogates improved with targeted experiments; active learning to cover difficult regions; uncertainty-aware search that respects feasibility; and keep heavy CFD and FEA offline while using light predictors at the tool level [13,14,17].

This review also draws out implications for industrial use and sustainability. Costs and environmental impacts come from gas heating and helium use, wear of consumables, and wasted material from errors and porosity. ML supports improvements in all areas: CFD-based optimisation to reach required speeds at lower temperatures; cost-effective gas mixing; design filters that reduce failed builds; predictive geometry and porosity control that cut rework; and monitoring systems that detect wear and drift early [15,17,20,27]. A key missing element is an integrated, fast, and uncertainty-aware closed-loop system that combines sensing with robotic control, particularly for complex geometries. The research methodology for this systematic literature review has been described in Section 2, following by classification and discussion of the research domains in Section 3 and Section 4, respectively. Section 5 highlights the challenges and opportunities for AI/ML in CSAM, whereas Section 6 provides answers to the research questions. The conclusions of this work are outlined in Section 8.

2. Research Methodology

An extensive systematic literature review combined with text-mining analysis was undertaken to identify, evaluate, and synthesise studies in which artificial intelligence or machine learning techniques have been applied to cold-spray additive manufacturing up to April 2025, using the PRISMA-2020 guidelines [28]. The overall process adopted is summarised in Figure 1. Identification and evaluation of relevant publications followed the systematic-review procedure outlined by [29], which involves formulating guiding research questions, collecting potentially relevant studies, and critically selecting and evaluating the collected literature. Each primary study was examined against predetermined inclusion and exclusion criteria (

Table 1. Inclusion and exclusion criteria.

Inclusion Criteria	Exclusion Criteria
All English-language publications AI/ML applied to CSAM	No AI/ML Manufacturing process mismatch
All years included	Non-English works
No access limitations

).

For the final set of accepted papers, keyword co-occurrence and clustering analysis were performed using VOSviewer version 1.6.20, according to the VOSviewer method proposed by [30,31]. This bibliometric step provides a clear overview of the main research topics and directions in the application of AI and ML to CSAM.

2.1. Step 1: Potential Research Questions

Here are three high-level research questions that capture the core themes in the literature:

• RQ1: What types of CSAM problems are currently being addressed with AI/ML?
• RQ2: Which AI/ML models are used for the identified CSAM problems?
• RQ3: What benefits, limitations, and open challenges are reported when applying AI/ML to CSAM?

2.2. Step 2: Keywords Identification

To conduct the systematic literature review, keywords were selected carefully to identify studies at the intersection of cold spray additive manufacturing and machine learning (using Boolean operators: “AND”, “OR”).

[(“cold spray additive manufacturing” OR “cold spray 3D printing”) AND (“artificial intelligence” OR “machine learning” OR “supervised learning” OR “semi-supervised learning” OR “unsupervised learning” OR “reinforcement learning” OR “deep learning” OR “neural networks”)].

2.3. Step 3: Resources for Searching

Databases used for identifying the relevant papers include Scopus, IEEE Xplore, ACM Digital Library, ScienceDirect, SpringerLink, and Taylor & Francis Online. Together, these platforms provide the most extensive and reliable sources to cover the subject of this review, offering major collections of peer-reviewed journal articles, books, conference proceedings, and other scholarly publications. The search covered publications from 2014 to 2025 in all the aforementioned databases (last search: 30 April 2025). No study registers or literature sources were searched. The review protocol was not formally registered in any external registry.

2.4. Step 4: Inclusion and Exclusion Criteria

Table 1 summarises the inclusion and exclusion criteria guiding the article-selection process. Publications were first screened using their titles, abstracts, and keyword lists; those remaining were then evaluated in full-text form to confirm inclusion.

2.5. Step 5: Selecting Studies and Analysing Keyword Co-Occurrence

The systematic literature review was conducted in four stages, following the PRISMA 2020 guidelines [28]. First, an initial Boolean keyword search identified 386 candidate papers. Before screening, 3 duplicate records were removed. In the second stage, titles, abstracts and keywords of 383 records were screened against the exclusion criteria, leading to the removal of 160 records and 223 reports being sought for retrieval (reports not retrieved, n = 0). In the third stage, 223 full texts were assessed for eligibility. At this stage, a study had to

• State clearly defined CSAM problems.
• Justify why an AI/ML approach was suitable.
• Demonstrate that the proposed method was evaluated or validated.

Applying these criteria resulted in the exclusion of 204 reports (No AI/ML = 120; manufacturing process mismatch non-CSAM = 80; non-English = 4). In the fourth and final stage, 19 primary studies met all requirements and were included for synthesis. The above process is summarised in the PRISMA flow diagram (Figure 1).

Analysing how specific keywords appear together in the literature offers a practical way to highlight the main themes in a research field [32]. A keyword co-occurrence matrix C = [$C_{ij}$] was constructed, where $C_{ij}$ is the number of documents in which keywords $i$ and $j$ co-occur and $C_{ii}$ = 0. ${\omega }_i$ is the total number of documents containing keyword $i$. Following VOSviewer, pairwise similarity ($s_{ij}$) was computed using association strength:

$$s_{ij}= {\displaystyle \frac{C_{ij}}{{\omega }_i{\omega }_j}}.$$

(1)

This normalisation corrects marginal frequencies and is the default in VOSviewer [30]. Two-dimensional coordinates ${\mathit{x}}_i$ = ($x_{i1},x_{i2}$) ∈ $R^2$ were obtained using the VOS mapping technique by minimising a weighted sum of squared distances,

$$V (x_1,\dots ,x_N) ={\sum }_{\mathrm{i}<\mathrm{j}}{s}_{ij}{\parallel x_i-x_j\parallel }^2,$$

(2)

subject to the scale-fixing constraint that the average inter-item distance equals 1:

$${\displaystyle \frac{2}{n(n-1)}}{\sum }_{\mathrm{i}<\mathrm{j}}\parallel x_i-x_j\parallel = 1, \parallel x_i-x_j\parallel = \sqrt{{(x_{i1}-x_{j1})}^2+{(x_{i2}-x_{j2})}^2}.$$

(3)

These steps correspond to VOSviewer’s similarity to mapping pipeline [30].

For clustering, links weights were taken as $c_{ij}$ = $C_{ij}$ with node degrees $c_i$ = $\sum _{\mathrm{j}\ne \mathrm{i}}{c}_{ij}$ and total link weight m = ${\displaystyle \frac{1}{2}}$ $\sum _{\mathrm{i}}{c}_i$. Items were assigned to cluster labels $X_i$ ∈ {1, …, K} by maximising the weighted, resolution-parameterized modularity used in the unified VOS framework:

$$\widehat{V}(X_1,\dots ,X_N)= {\displaystyle \frac{1}{2m}}{\sum }_{\mathrm{i}<\mathrm{j}}\delta \left(X_i,X_j\right){\omega }_{ij}(C_{ij}- \mathsf{\gamma }{\displaystyle \frac{C_i{C}_j}{2m}}), {\omega }_{ij}= {\displaystyle \frac{2m}{C_i{C}_j}}, \mathsf{\gamma } > 0.$$

(4)

Here $\delta \left(X_i,X_j\right)$ = 1 if items i and j are in the same cluster and 0 otherwise; ${\omega }_{ij}$ is a pairwise weight; γ is the resolution parameter controlling typical cluster size; and K is the number of clusters. This objective is equivalent to minimising the unified energy with clustering distances, and it reduces to modularity when ${\omega }_{ij}$ = 1 and $\mathsf{\gamma }$ = 1 [31].

The ${\mathit{x}}_{\mathit{i}}$ are continuous map coordinates from (2) and (3) (used for visualisation), whereas $X_i$ are discrete cluster labels from (4). They are produced by related but distinct optimisations; in plots, items are positioned at ${\mathit{x}}_{\mathit{i}}$ and coloured by $X_i$. There is no direct algebraic mapping from $X_i$ to ${\mathit{x}}_{\mathit{i}}$ [30,31]. VOSviewer was used to generate keyword maps and clusters [30].

Figure 2 shows the keyword co-occurrence map obtained from the 19 papers that met the inclusion criteria. This is a label-only VOSviewer view (without showing links). The size of each keyword shows how often it appears and is usually linked to its total connections. The colours represent VOSviewer clusters (groups of related keywords), and the distance between labels shows how closely they are related (closer labels co-occur more often). VOSviewer settings: minimum keyword occurrences = 1; normalisation = Association Strength; clustering resolution = 1.00; minimum cluster size = 4; layout (default) = Attraction 2, Repulsion 0. In the VOSviewer layout, physical distance reflects how frequently two terms co-occur; therefore, the closeness of nodes and groups indicates how strongly the keywords are related. Moreover, a grey scale keyword co-occurrence map with description has been provided in the Appendix A.

Six colour-coded clusters emerge, indicating the main research focus areas within the field of AI/ML that connect with CSAM. As the VOSviewer output, physical distance in the map reflects how frequently two terms appear together, the closeness of each group shows how related the keywords are together. The red cluster includes phrases such as anomaly detection, powder-flow monitoring, thermal imaging and deep learning and therefore represents Monitoring & Defect Analytics; the green cluster collects materials terms (Cantor alloy, thermal spray, composites, gradient boosting), pointing to Material Analytics; a dark-blue cluster built around layer, profile, and Gaussian process reflects Deposition Geometry Modelling; the yellow cluster, contains spray distance, coating-profile prediction and related path-planning terms and is labelled Deposition Path & Profile Prediction; the purple cluster groups design-stage concepts such as additive manufacturing, tool-path optimisation and porosity, defining Design-for-CSAM (DfCSAM); finally, a light-blue cluster centred on neural network and AI-driven deposition control rounds out the map.

Although six clusters appear visually, the findings are reported in four research domains aligned with the CSAM workflow. Applying the ≥50% content rule, dark blue, yellow, and light blue clusters all map to RD4 (Deposition & AI Optimisation, reported together under RD4). The remaining clusters map to RD1 DfCSAM (purple), RD2 Material Analytics (green), and RD3 Real-Time Monitoring & Defect Analytics (Red). This organisation matches the clustered keywords to the CSAM phases from design through in-process control to deposition optimisation, as summarised in

Table 2. Clustering of selected publishing outcomes.

Cluster	Keywords Samples	Research Domain
Cluster 1 (purple)	additive manufacturing, toolpath optimisation, porosity, data driven	DfCSAM (RD1)
Cluster 2 (green)	cantor alloy, composites, polymers, metal deposition, thermal spray, gradient boosting	Material Analytics (RD2)
Cluster 3 (Red)	anomaly detection, powder-flow monitoring, thermal imaging, multi-stage model, deep learning, time-series forecasting	Real-Time Monitoring & Defect Analytics (RD3)
Cluster 4 (dark blue + yellow + light blue)	multi-layer profile, coating-profile prediction, spray distance, geometry/overlap, gaussian process, neural network	Deposition & AI Optimisation (RD4)

.

Keyword statistics supporting these thematic patterns are presented in

Table 3. The keywords clustering co-occurrence.

Keyword	RD (Cluster)	Occurrences	Total Link Strength
cold spray	RD3	11	51
additive manufacturing	RD1	7	35
machine learning	RD2	5	26
neural network	RD4	5	25
deep learning	RD3	4	14
model	RD4	3	19
thermal spray	RD2	3	12
profile	RD4	2	14
spray angle	RD4	2	14
trajectory	RD4	2	9
data-driven	RD1	2	7
data-efficient	RD4	1	9
geometry	RD4	1	9
limited data	RD4	1	9
multi-stage model	RD3	1	5

. “Cold spray” appears eleven times and has the highest total link strength (51), confirming this cluster (RD3) is a key research area. Across the 75 important keywords, RD 4 (geometry and AI-focused) shows the most prominence, as it includes 28 occurrences (almost 37% of the sample) and a total link strength of 155. Example items include “neural network” (5 occurrences, 25 links), “model” (3 occurrences, 19 links), and the paired geometry descriptors “profile” and “spray angle” (each 2 occurrences, 14 links). Even single occurrence terms such as “data-efficient,” “geometry,” “limited data,” and “trajectory” have link strengths of 9, showing ongoing interest in predictive models that work with less data.

In contrast, the design focused RD 1 cluster contains 11 keyword mentions “additive manufacturing” and “data-driven” and an aggregate link strength of 49, while the materials analytics RD 2 cluster has 14 hits “machine learning” and “thermal spray” with a combined strength of 62. Because the table lists only the top terms, these figures represent lower than actual totals; nevertheless, they indicate that current research gives more attention to in-process monitoring (RD 3) and Deposition & AI Optimisation (RD 4), whereas the systematic integration of AI/ML into design part (RD 1) and alloy development (RD 2) remains comparatively under explored.

2.6. Data Synthesis: A Priori Plan, Feasibility of Pooling, and Overview of Observed Patterns

At the planning stage, it was decided that a quantitative meta-analysis would only be performed if studies looked at similar outcomes, used comparable measures on similar scales, and reported uncertainty (or enough data to calculate it). Otherwise, a structured summary would be provided, organised by research area, data source, sensing, and algorithm type, with collection of study details and reported performance results.

After collecting the data, meta-analysis was not carried out because of major differences in research focus, tasks, and outcomes as well as variations in measures and scales with frequent lack of uncertainty estimates, differences in data sources and measurement methods, and variation in validation approaches and sample sizes. These factors prevented calculating a meaningful overall effect. The planned structured summary was therefore followed.

In terms of overall findings, performance depended heavily on context: results within the same setup were usually better than across different setups; imaging-based monitoring often showed good accuracy within the tested ranges; porosity-focused models were less consistent and more sensitive to how the dataset was built; and inverse-design studies often used forward models with limited outside validation. No claims are made about specific methods or algorithms beyond this neutral summary.

2.7. Bias Risk

Risk of bias was evaluated with the PROBAST tool [33], which reviews prediction-model studies across four domains and assigns Low, High, or Unclear judgments at both domain and study level. The first domain (participants and data source) examines dataset suitability and origin, including size and coverage of the explored process space, completeness or missing data, and whether different literature sources were combined. The second (predictors) assesses how input variables were defined and measured and whether any preprocessing, segmentation, or feature construction could introduce information leakage. The third (outcome) reviews the clarity and consistency of outcome definition and measurement and matches the stated modelling goal. The fourth (analysis and validation) evaluates sample size in relation to model complexity; separation of tuning from evaluation; carefulness of internal split design; presence or absence of external or temporal testing; handling of missing data; and clarity of performance reporting, including calibration where applicable. PROBAST rules were followed: overall risk of bias is High if any domain is High; overall is Low only if all four domains are Low; otherwise overall is Unclear. Inter-rater agreement was measured using percent agreement before resolving disagreements: title/abstract screening, 86% agreement; full-text eligibility, 83% agreement; risk-of-bias ratings, 84% agreement. Disagreements were resolved through discussion, and no automation tools were used. Across 19 studies, overall risk of bias was High in 7; the remaining studies did not meet criteria for Low risk across all domains and were therefore judged Unclear overall. Common concerns included limited or mixed datasets (especially literature-based), reliance on internal splits without external or temporal testing, barriers to reproducibility, measurement inconsistency, and potential context leakage when segmenting or combining data. Per-study domain-level judgments and justifications are presented below in

Table 4. Risk of bias (PROBAST) per-study judgments and justifications.

	Reference	Participants and Data Source	Predictors	Outcome	Analysis and Validation	Overall Risk of Bias
1	[21]	Unclear: Mixed FEM + experimental with small FEM pool (~50 impacts); experimental share not numerically specified.	Low: Inputs clearly defined/consistent across sources; no cross-paper aggregation.	Unclear: Outcome measured (SEM), but experimental reporting incomplete.	Unclear: Overfitting flagged for GPR.	Unclear
2	[16]	Unclear: 227 literature-derived cases with mixed sources.	High: Predictor definitions may vary across sources (7 inputs from many papers).	High: Porosity measured by different methods across papers.	Unclear: Validation details sparse; some algorithms underperform markedly.	High
3	[34]	Unclear: Purpose-built dataset, but only 4 lab states.	Unclear: potential leakage from segmentation with a single 80/20 split.	Unclear: Fault labels clear but not linked to deposition metrics.	Unclear: 80/20 internal split only; small Number; potential leakage/overfitting.	Unclear
4	[22]	Unclear: Reuses a single prior experimental dataset ([23]); limited coverage.	Low: Predictors from a single source reduce heterogeneity.	Unclear: Inverse-problem setup; sparse detail on outcome fidelity	Unclear: Forward model driven; inverse accuracy constrained by forward MLP.	Unclear
5	[20]	Unclear: 35-paper literature dataset with many missing values (84% complete).	High: Mixed-source predictors with imputation choices add bias risk.	High: Porosity from differing measurement techniques and operator bias.	Unclear: Many models; performance depends on handling; external generalisation limited.	High
6	[35]	Unclear—Mixed FEM/experimental; absolute sizes not stated.	Low—Study-defined predictors; materials fixed.	Low—Outcomes consistently measured within study	Unclear—Overfitting on penetration; GA tuned architecture only.	Unclear
7	[36]	Unclear: Pure MD-generated data.	Low: Predictors defined within simulation.	Unclear: Flattening ratio is geometric proxy; no adhesion outcomes.	Unclear: Optimisers compared but no physical validation.	Unclear
8	[12]	High: Very small dataset (~56; single material pair).	Low: Study-defined predictors, Single derived predictor (impact velocity).	Low: Outcomes measured consistently within study.	Unclear: Simple internal split; poor fit for flattening.	High
9	[19]	Unclear: 37 sprays; Controlled runs/material; limited size.	Low: Thermal time-series defined within study.	Low: Anomaly labels tied to thermal trends within study.	Unclear: Clear train/validation/test.	Unclear
10	[13]	Unclear: All-experimental overlaps (48 overlap profiles + references); single material setup.	Low: Study-defined geometric predictors including surface prior.	Low: Profile outcomes consistently measured.	Low: Strong comparative baselines; large improvement; clear test split.	Unclear
11	[37]	Unclear: Small training/test with some virtual data.	Low: Predictors defined within study.	Unclear: Virtual data may be biased; single-track only.	Unclear: reliance on virtual augmentation; varied spray/nozzle setups untested; model trained offline	Unclear
12	[24]	Low: ~330 multi-layer samples (flat/curved).	Unclear: High-dimensional inputs and sample-level splits could leak context; reporting insufficient to exclude it.	Low: Outcome definition clear and consistent.	Low: Train/validation/test split; strong RMSE/R2; acknowledges overfit risks.	Unclear
13	[14]	Unclear: Large, simulated set with limited experimental confirmation.	Low: Study-defined inputs.	Unclear: Simulation-dominant outcomes; edge-case errors noted.	Unclear: Good within-sim metrics; edge degradation; external validity unclears.	Unclear
14	[38]	Unclear: DoE within one setup; range not fully explained.	Unclear: Multimodal (optical/ Infrared) + process sensors well specified.	Low: Labels from lab tests.	Unclear: 5-fold Cross-Validation; very high F1/Acc; drift not yet handled.	Unclear
15	[39]	High: Very small mixed literature/experimental set (n = 35), with 80/20 split.	High: Mixed sources and limited features increase bias/leakage risk.	Unclear: Porosity target with limited detail.	Unclear: Multiple complex models on tiny sample number; overfitting risk high.	High
16	[23]	High: Small single-material dataset (48 tracks).	Low: Study-defined predictors.	Unclear: Accuracy degrades at off-normal angles; sensitivity suggests instability.	Unclear: No extra data splitting was performed; the train/test set was small, and the model’s accuracy went down at angles.	High
17	[17]	High: No unified quantitative dataset shared.	Unclear: Predictors referenced conceptually; not specified for a usable model.	Unclear: No quantitative outcomes reported.	High: Conceptual; no validation metrics/splits.	High
18	[40]	Unclear: Planning framework with limited experimental checks.	Low: Inputs from CAD/simulation defined within study.	Unclear: Outcomes are geometric errors; limited detail on measurement repeatability.	Unclear: Demonstrations only.	Unclear
19	[41]	High: Small single-material dataset; narrow process range.	Low: Distance/step predictors measured directly.	Low: Thickness measured consistently.	Unclear: Limited validation/reporting; narrow conditions.	High

.

2.8. Certainty of Evidence Approach

Certainty (confidence) in the body of evidence for each outcome was graded qualitatively (High, Moderate, Low, Very Low) using four criteria: risk of bias, assessed with the PROBAST tool at study and domain level; consistency of findings across studies; imprecision (sample size, uncertainty/variance reporting, and presence of external or temporal validation); and indirectness (applicability of data, tasks, and outcomes to CSAM). Ratings were downgraded when serious concerns were identified in any criteria. Across the 19 included studies, PROBAST overall risk of bias was ‘High’ in 7 and ‘Unclear’ in 12, which drove risk-of-bias downgrades at the outcome level. Considering all four criteria, certainty was ‘Moderate’ for geometry/profile prediction and monitoring and anomaly detection (consistent within-study performance but limited by small, single-site datasets and scarce external validation), and ‘Low’ for porosity prediction and inverse design and optimisation (high risk of bias from mixed literature-derived data, inconsistent measurement and metrics, limited or simulation-only validation, and small test samples). Outcome-level justifications and contributing studies are detailed in

Table 5. Certainty of evidence.

Outcome Domain	Certainty	Downgrades Applied (with Reasons)
Geometry/profile prediction (RD4)	Moderate	Evidence For 9 studies ([12]; [21]; [35]; [23]; [37]; [13]; [24]; [14]; [41]) shows, Risk of bias: downgraded (3/9 High; 6/9 Unclear; small single-site datasets; mixed FEM/experimental). Imprecision: downgraded (limited scope; very few external or repeated tests). Consistency: not downgraded (results point in the same direction; low errors reported). Indirectness: not downgraded (tasks directly match CSAM geometry; limited use of simulations).
Porosity prediction (RD2/RD3)	Low	Evidence from 3 studies ([16]; [20]; [39]) shows, risk of bias: downgraded (3/3 High; mixed sources from the literature; inconsistent measurements; missing data and estimations). Consistency: downgraded (model results change depending on dataset setup). Imprecision: downgraded (small test samples; little reporting of uncertainty). Indirectness: not downgraded (focus matches porosity; variation fits under bias/inconsistency issues).
Monitoring/anomaly or defect detection (RD3)	Moderate	Evidence from 3 studies ([34]; [38]; [19]) shows, risk of bias: downgraded (0/3 High but 3/3 Unclear; single setup/material; potential split leakage). Imprecision: downgraded (no external validation across sites; small sample size). Consistency: not downgraded (accuracy within each study was consistently high). Indirectness: not downgraded (tasks directly relate to monitoring).
Inverse design/optimisation (RD1/RD4)	Low	Evidence from 4 studies ([36]; [22]; [40]; [17]) shows, Indirectness: downgraded (simulation-only or rule-based planning; relies on surrogate forward models; limited experimental verification). Imprecision: downgraded (few, small demonstrations; limited quantitative validation). Risk of bias: downgraded (1/4 High; 3/4 Unclear; single setup/material; dependent on earlier models). Consistency: downgraded (different methods/metrics make comparisons hard).

.

3. Classification of the Selected Literature

This section summarises the 19 CSAM papers after screening (

Table 6. Research domain classification and data extraction.

	References	Research Domain	ML/AI Models	Data Type	Research Target	Material	Data Volume	Open Challenges	Final Performance /Accuracy
1	[21]	RD2/RD4	Linear Regression (LR), Gaussian Process Regression (GPR), Neural Networks (NN); trained on FEM and mixed (FEM + experimental) datasets	Train: 70% FEM (2-D explicit, 50 impacts per material pair) + 30% experimental; Test: 100% experimental	Forecasting particle flattening ratio and penetration depth to improve coating adhesion from impact velocity, powder Yp, substrate Ys	Powders: Cu, Al, 316 L steel Substrates: PEEK and ABS (unreinforced and long-carbon-fibre-reinforced)	50 simulated impacts (FEM); experimental share not numerically specified (30% of mixed-training set + full test set)	No real-time process control or adaptive ML strategies GPR showed inconsistency in flattening accuracy (possible overfitting). Generalisation to other substrate types not tested FEM-only data lacks some microstructural realism Optimal ratio of FEM-to-experimental data for training is still unresolved.	Flattening (normalised): RMSE 4.7387, MAE 2.0051, R2 0.34—Data: FEM—Model: GPR; Penetration (%): RMSE 1.45, MAE 0.7654, R2 0.73—Data: FEM—Model: GPR; Flattening (normalized): RMSE 1.8327, MAE 1.4462, R2 0.90—Data: Mixed (30% experimental + FEM), Test: experimental-only—Model: LR; Penetration (%): RMSE 0.5842, MAE 0.4130, R2 0.96—Data: Mixed (30% experimental + FEM), Test: experimental-only—Model: NN.
2	[16]	RD3	Linear Regression (LR), Decision Tree Regression (DT), Random Forest Regression (RF), Extreme Gradient Boosting (XGBRegressor), LightGBM Regressor (LGBMRegressor)	227 literature-derived cases; 7 input variables (gas γ, gas T, gas P, standoff, powder D50, powder ρ, substrate ρ)	Prediction of coating porosity from spray/material parameters	Generic metal/alloy cold spray powders; exact materials not detailed	227 labelled cases for training + 10 unseen test sets	Dataset aggregated from the literature, hence measurement inconsistency (porosity via different methods). No in situ data or process feedback. Limited to static prediction (no dynamic control or monitoring). Some algorithms (e.g., linear regression) underperform (R2 = 0.27, MAE = 5.06)	Porosity (%): Best model Decision Tree—R2 0.75, MAE 2.93; Worst Linear Regression—R2 0.27, MAE 5.06; Trends: higher gas pressure and temperature reduce porosity, whereas larger particle size increases porosity.
3	[34]	RD3	Convolutional Autoencoder (CAE) unsupervised anomaly detector + Convolutional Neural Network (CNN) classifier with Integrated Gradients (IG) explainability	Internal stethoscope sensor sound (48 kHz) from powder-feeder vibrations; external microphone baseline sensors under 4 CS conditions (normal + 3 anomalies) for comparison	Real-time detection and classification of powder-flow faults (no-powder flow, feeder clogging, no-gas flow) during LPCS	Tin (Sn) powder, (10–45 µm, d50 ≈ 17 µm), sprayed on aluminium plate in low-pressure CSAM setup	30 s signals × 4 states, segmented into 1 s frames; 80/20 split for train/validation; total ≈ 240 Log-Mel spectra per class	No direct link yet to quantitative flow rate or deposition efficiency. Evaluation limited to 3 specific anomalies under fixed test settings. Interpretation via IG is qualitative and not calibrated to deposition metrics.	Powder-flow fault classification: Anomaly detection (s-CAE) 96.7%, (m-CAE) 90.8%; CNN classification test accuracy 95% (stethoscope input); best class (“no gas”) precision/recall 100%.
4	[22]	RD1/RD4	Hybrid Particle Swarm Optimisation—Multilayer Perceptron (PSO-MLP) (Forward MLP learns SDP → PSO inverse model predicts IPPs)	Experimental SDP data from [23]: 67-point SDP samples with θ (spray angle), v (traverse speed), SOD (stand-off distance); no synthetic data	Inverse prediction of spray angle (α), nozzle traverse speed (v), and standoff distance (d) to reproduce a user-defined SDP (single-track)	Not explicitly specified—generic metallic CSAM powders	DS1: 36 SDPs × 67 pts = 2412 data points (train/validation/test split); DS2: 12 SDPs × 67 pts = 804 data points	Accuracy of inverse model limited by forward-model MLP. Large error for standoff distance d, posed inverse problem: many-to-one mapping of SDP to IPPs. Need larger and more diverse datasets.	Track profile height (forward model): MAE 0.0086–0.0131 mm (DS2); Inverse prediction of (angle, speed, standoff): MAE 0.0591—Optimiser: PSO-MLP; Runtime ≈ 2.7 s.
5	[20]	RD2/RD4	Random Forest (RF), Extreme Gradient Boosting (XGB), Categorical Boosting (CatB)	Literature-derived dataset: 35 papers, 242 porosity labels, 14 parameters; 84% completeness	Predict percentage porosity of cold-sprayed deposits	metallic powders (e.g., CP Ti, Ti-6Al-4V, copper, 304/316 SS) and matching substrates	242 porosity measurements, 14 process parameters (gas temp, pressure, angle, standoff, feeder rate, etc.) from 35 papers (84% complete)	The literature-derived data introduce “a higher level of uncertainty regarding its accuracy and quality” and many missing values. Porosity measurements obtained by two techniques (image analysis ≈ 90%, Archimedes ≈ 10%) prone to operator bias.	Porosity (%): Random Forest test RMSE 0.79% (train 2.77%); CatBoost RMSE 1.34–1.78%; 12 models achieved ≤2% RMSE; Most influential: gas temperature, gas pressure, particle size.
6	[35]	RD2/RD4	Genetic-algorithm (GA)–designed Deep Learning: Wide Neural Network (WNN) for flattening; Tri-layer Neural Network (TNN) for penetration; compared to LR/ANN baselines	Train: 30% experimental + 70% FEM, powders Cu/Al/Steel/Ti on ABS/PEEK/PA66; Test: 100% experimental	Predict particle flattening and penetration to improve coating quality	Powders: Cu, Al, Steel, Ti; Substrates: ABS, PEEK, PA66	Training: 30% experimental/70% FEM; second training set FEM-only; test set experimental-only (absolute size not stated)	GA optimised architecture, not hyperparameters—structure-focused. TNN showed high variance on unseen data (overfitting on penetration prediction)—poor generalisation. Need for multi-output regression for simultaneous flattening + penetration—to avoid separate models DL model interpretation and explainability not addressed—remains a black box. Further generalization to new material combinations untested—trained on known sets only.	Flattening (normalised): RMSE 0.73, MAE 0.51, R2 0.92—Model: GA-designed Wide NN—Test: experimental-only; Penetration (%): RMSE 2.57, MAE 1.46, R2 0.60—Model: Tri-layer NN (variance on unseen data).
7	[36]	RD4	Back-Propagation Neural Network (BPNN)—surrogate/meta-model Optimisers: Particle Swarm Optimisation (PSO), Differential Evolution (DE), Efficient Global Optimisation (EGO + Kriging surrogate)	Synthetic, simulation-generated data: MD snapshots (images) processed to a numerical flattening ratio + the input process variables (velocity, radius, angle)	Maximise deposition quality (flattening ratio µ) by tuning particle velocity, size and impact angle	Copper (Cu) particle on Cu substrate (FCC lattice)	1000 MD samples for BPNN training + 1000 for testing. Optimisation runs: 20 initial EGO samples; 20 particles in each PSO/DE run	MD-based optimisation is time-consuming. Trade-off between efficiency vs. accuracy when choosing optimisation method (EGO is described as the most efficient, but not always the most accurate also, PSO and DE give better solutions, but are more computationally expensive). µ is scalar geometric metric; no stress/damage or adhesion outputs. Shows BPNN works better with PSO/DE, not EGO.	Optimisation of flattening (μ): Best classic PSO objective c = 0.15084 (μ ≈ 6.6) at v = 1200 m/s, r = 16.145 Å, θ = 0.7°; BPNN surrogate cut heuristic optimisation cost by ~50% vs. direct PSO/DE.
8	[12]	RD2/RD4	Univariate Linear Regression (LR); Theory-Guided Machine Learning (TGML)	Mixed dataset: 50% experimental (SEM-based flattening/penetration) + 50% FEM simulations (ABAQUS); impact velocity estimated using analytical physics-based model	Predict single-particle penetration depth and flattening ratio from impact velocity to optimise coating quality	Cu particles on PEEK (fibre-reinforced composite) substrate	56 total samples → 40 simulated + 16 experimental Split ≈ 70%/30% ⇒ near to 50 train/16 test (authors note “Fifty training examples were fed to the model”)	Univariate model unable to capture nonlinear or coupled effects. Low accuracy for flattening predictions (suggests need for cubic/multivariate models). Impact velocity estimation based on theoretical formula, not real sensor data. Lacks generalisation to other materials (only Cu–PEEK studied). Manual feature selection and gradient descent convergence issues limit scalability.	Penetration (normalised): RMSE 0.2713, Correlation 0.95; Flattening (normalized): RMSE 1.0206, Correlation 0.83—Univariate linear regression on mixed (exp + FEM) data.
9	[19]	RD3	ThermoAnoNet—unsupervised deep-learning time-series forecaster built on N-BEATS; baselines: Recurrent Neural Network (RNN) and Temporal Convolutional Network (TCN)	50 fps FLIR A655sc IR image sequences (640 × 480 px) converted to averaged substrate-temperature time series	Detect anomalies (e.g., feed rate changes) in CS by comparing forecasted vs. observed substrate temperature trends	Cu-159-3 copper powder sprayed on Cu plate; N2 carrier gas	37 total sprays (22 for training, 5 for validations, 10 for testing); layer wise mean temp trend extracted and filtered; input: 200 steps; output: 300 step forecasts	Lower sensitivity to small feed rate changes (2 → 4 rpm) Lack of generalisation to complex geometries or obscured views Real-time inference limited to sub-layer detection granularity. Thermal-only modality; requires multi-sensor fusion for robustness. No feedback integration or active process control.	Thermal-trend anomaly monitoring: Accuracy 0.90, F1 0.94 (test); Mean detection delay 0.27–0.54 layers; Spray-class accuracy 1.00 for large feed-rate shifts (drops to 0.60 for small 2 → 4 rpm change).
10	[13]	RD4	Surface-aware Gaussian Process Regression (GPR) with explicit Gaussian-superposing mean; baseline purely data-driven GPR	48 overlapping-track profiles (exp.); 21 single-track profiles for surface reference; inputs: traverse speed, standoff, polar angle, surface type	Predict the overlapping-track profile shape to improve geometric control and toolpath planning in CSAM	Powder and substrate: ASTM grade-2 Ti	48 overlapping-track profiles + 21 single-track profiles → 48 × 151 ≈ 7248 geometry samples (plus 3171 single-track points)	Focused only on 2-track overlaps (no multi-track/layer generalisation). Limited to fixed powder/gas setup (single gas temp/pressure). No feedback loop or adaptive control. GPR surrogate depends on quality of initial mean function. Current results based on 2D cross-sections, not 3D prediction.	Overlapping-track profile: MSE 0.0002183, R2 0.9988, MAPE 0.57%, Max APE 6.91%—Surface-aware GPR outperformed data-only GPR.
11	[37]	RD4	Data-assisted Artificial Neural Network (DANN) (feed-forward ANN with Bayesian-regularised back-prop) + Gaussian function with quadratic regression coefficients (hybrid)	Experimental process parameters (spray-angle, traverse-speed, standoff); 3-D laser-scanned track-profile geometry; virtual Gaussian-generated samples—all numerical	Predict full single-track geometry to improve geometric control	CP-Ti (grade-2) powder (15–45 µm), N2 carrier gas	Training: 36 physical tracks × 67 points = 2412 + 804 virtual points; Test: 12 tracks × 67 = 804	Limited to single-track profiles only; no overlap/layer modelling. Virtual data generated from simplified Gaussian model may bias ANN. Applicability to curved paths, varied spray/nozzle setups untested. Model trained offline; no in situ update or closed-loop feedback.	Single-track profile: MAPE 1.23%, Max APE 5.75%, R2 0.9988, MSE 1.03 × 10−4—Data-efficient ANN + Gaussian hybrid beat pure ANN.
12	[24]	RD4	Feed-forward Artificial Neural Network (ANN) (2 hidden layers, 86 + 54; SCG training)	330 real-time scanned multi-layer profiles (flat + curved substrates) collected from 50 CS runs: 3 process variables + substrate profile as inputs	Predict complete multi-layer deposition profiles from CSAM parameters (SOD, traverse speed, cycles) and previous surface morphology	Cu powder (D50 ≈ 26 µm); Al substrate; 10-layer max, SOD = 20–40 mm, v = 20–100 mm/s	330 profile samples × 206 points (70% train/15% validation/15% test)	Limited to 2D profiles (no 3D volume prediction). ANN performance is sensitive to substrate irregularity and noise. No physical features (e.g., temp/velocity) integrated. No coupling with real-time sensor feedback or path planning yet. ANN may overfit due to large input size (209 features) and limited process space.	Multi-layer profile (flat + curved): RMSE 0.02209, R2 0.9957—2-hidden-layer ANN robust to noisy substrate geometry.
13	[14]	RD1/RD4	Deep Neural Networks (DNNs) for 2D and 3D shape prediction trained on physics-based analytical simulations; compared with Transformers, CNNs, LSTMs, Gaussian Mixture Model (GMM), XGBoost	Simulated single-track shapes generated with a physics-based analytical model, plus a smaller set of experimental cross-sections for validation	Predicting single-track cross-sections and full 3-D deposit geometry; using these predictions inside an adaptive slicing and tool-path framework to inimize waviness/overspray.	Ti-6Al-4V powder (AP&C)	2400 simulated linear tracks (each 30 mm long) (full-factorial DoE on 4 input variables: scanning speed, spray angle, standoff distance).	Lower accuracy at the domain edges; difficulty capturing profile asymmetry at high spray angles; need to handle curved substrates and shadowing in future work.	Shape prediction: 2D DNN—Relative L2 error 0.59%, MSE 8.06 × 10−5, R2 0.999 (simulated); 3D DNN—Relative L2 error 2.62%; Experimental validation: MSE ≈ 9.4 × 10−3, R2 0.981, MAPE 16%.
14	[38]	RD3	2D Convolutional Neural Networks (CNNs) for multimodal defect detection; Grad-CAM used for interpretability; F1 score evaluation under 5-fold CV	IDS optical cam, FLIR A70 thermal cam, N2 gas T/P/flow; ring and side lights; edge-cloud IIoT	Detect oxidation, porosity, and geometric defects in each sprayed layer using multimodal image inputs; enable in situ defect classification and real-time monitoring	CuZn30 (copper-zinc alloy); sprayed via CGS; process gas: N2	DoE-based experiments: 3 × 3 matrix of temp (300–800 °C) × pressure (2.5–4.5 MPa); images + process sensor data per layer; labels from expert + lab tests	Model does not yet handle process drift (e.g., changing material batches or conditions). Data is labelled by experts—still semi-automated and subjective. Grad-CAM explanations are qualitative, not linked to quantitative domain thresholds.	In situ defect classification: F1 0.999 (ring-light oxidation), 0.998 (side-light porosity/shape); Accuracy 1.00; Grad-CAM highlights are consistent with defects.
15	[39]	RD3/RD4	Linear Regression (LR), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GBOOST), XGBoost, Support Vector Regression (linear, polynomial, RBF), Artificial Neural Network (ANN)	35 samples: 22 from literature, 13 from experiment (Cantor alloy CS on steel)	Predict porosity (%) in CS-deposited High Entropy Alloy coatings using material + process parameters	Fe(20)Cr20Mn20Ni20Co20 (Cantor HEA)	35 total samples; 80/20 train-test split; 5 input features; test band analysis + SHAP-based feature ranking	ANN underperformed due to low data volume. Features like traverse speed, spray spacing not included despite influence.	Porosity in HEA coatings (%): Best Support Vector Regression (linear)—RMSE 2.07, PCC 0.85; Linear Regression—RMSE 2.06, PCC 0.83; Several models show 1–3 outliers within 40% prediction band.
16	[23]	RD1/RD4	Multilayer-perceptron Artificial Neural Network (Bayesian-regularised back-prop)	Experimental parameters (spray angle, traverse speed, standoff distance) + measured single-track profiles	Predict full single-track geometry from angle, speed, standoff	Grade-2 (commercial-purity) Ti powder	36 training + 12 test samples (48 tracks); each track represented by 67 polar-length outputs, no simulation data	Model lacks robustness in regions of high geometric variation; ANN requires more data-efficient approaches and improved real-time acquisitions; Model sensitivity limited by small input vector (3 parameters only) vs. large output space (67 neurons); Accuracy degraded at off-normal spray angles (e.g., 48°) diverse data	Single-track profile: MAE 0.05782 mm, MAPE 8.34%, MSE 0.009454, R2 0.9493—MLP (Bayesian regularisation).
17	[17]	RD1/RD3/RD4	Artificial Neural Network (ANN)-based geometry predictors; Convolution-based shape predictor; discusses broader Deep Learning and Reinforcement Learning (RL) for optimisation	Hybrid system combining 3D scan data, reverse engineering, numerical simulation (MATLAB), in situ sensor feedback, and rule-based optimisation (porosity, hardness, roughness) → structured numeric and profile data	Develop an integrated hardware–software CSAM platform that couple’s path-planning, shape prediction, real-time monitoring and subtractive finishing to fabricate complex, precise parts	Cu (pure copper), Al (aluminium) demonstrator parts; system intended for “a wide variety of materials”	Complex digital models from structured light scanning and simulation-based shape evolution (no unified raw dataset shared)	pretreatment and post-treatment modules still not fully integrated; need better inter-module data exchange and quantitative validation	None reported; paper is conceptual (no numerical accuracy metrics).
18	[40]	RD1/RD4	Hierarchical clustering (unsupervised ML) + Optimised Rapidly Exploring Random Tree (RRT) for path-planning	3D CAD models (STL mesh of 122,774 triangles); MATLAB-based simulation; Experimental validation with physical spraying on aluminium substrate; no training set	Generate layer-wise filling (FP), compensation (CP), and linking (LP) trajectories for multi-featured CSAM parts, avoiding over-deposition and ensuring shape fidelity	1050 A Al substrate + commercial Al powder (29 µm); extra demo with Cu powder	Example model: 122,774 triangular facets (slice points generated for every layer)	No ML model trained from data: intelligent planning but not learning-based LP generation is path-optimised but still deterministic. Low resolution at small features (e.g., fingertips) due to nozzle size limits. Spray angle constraints and hardware limitations restrict path variety. Trajectory is fixed after planning; does not adapt to in-process changes. Cannot yet generalise to full curved or non-uniform layer build features.	Dimensional accuracy: Height error 0.8–1.2 mm (5.7–7.4%); Surface flatness ≤ 1.4 mm; Wall straightness 32.8 µm; Uniform layer thickness ≈ 1.25 mm.
19	[41]	RD4	Gaussian curve-fitting regression embedded in Thermal Spray Toolkit	Experimental thickness profiles captured by CMM; simulated Gaussian curves	Optimise spray-distance and scanning-step to maximise coating uniformity and generate robot trajectories	Al 5056 powder (40 µm) on Al substrate	4 spray distances × 5 profiles × ≈30 points ≈ 600 profile points; scanning-step sweep 2–8 mm	Measurement and experimental noise can cause deviations between simulated and actual coating thickness scanning steps exceeding one standard deviation (near to 3 mm) lead to significant loss of surface flatness, reducing coating uniformity. Absence of defined robust parameter ranges makes it harder to ensure consistent quality across different setups. Results are based only on a constant spray angle (90°) and a single material (Al 5056), limiting generalisation to other conditions.	Thickness/profile fit: R2 98.92%; Absolute error ≤ 0.09 mm across cases; Gaussian regression matched CMM profiles closely.

ANN = Artificial Neural Network; DNN = Deep Neural Network; BPNN = Back-Propagation Neural Network; DANN = Data-Assisted Artificial Neural Network; CNN = Convolutional Neural Network; CAE = Convolutional Autoencoder; GPR = Gaussian Process Regression; LR = Linear Regression; DT = Decision Tree; RF = Random Forest; XGB/XGBoost = Extreme Gradient Boosting; LGBM = LightGBM; CatBoost = Categorical Boosting; SVR = Support Vector Regression; TCN = Temporal Convolutional Network; RNN = Recurrent Neural Network; N-BEATS = Neural Basis Expansion Analysis for Time Series; PSO = Particle Swarm Optimisation; DE = Differential Evolution; EGO = Efficient Global Optimisation; TGML = Theory-Guided Machine Learning; GA = Genetic Algorithm; WNN = Wide Neural Network; TNN = Tri-layer Neural Network; GMM = Gaussian Mixture Model; RRT = Rapidly Exploring Random Tree; SCG = Scaled Conjugate Gradient; IG = Integrated Gradients; Grad-CAM = Gradient-weighted Class Activation Mapping; FEM = Finite Element Method; MD = Molecular Dynamics; CAD = Computer-Aided Design; STL = Stereolithography; CMM = Coordinate Measuring Machine; IIoT = Industrial Internet of Things; FLIR = Forward-Looking Infrared; CS = Cold Spray; CGS = Cold Gas Spray; LPCS = Low-Pressure Cold Spray; CSAM = Cold Spray Additive Manufacturing; SDP = Single-track Deposition Profile; IPP(s) = Input Process Parameter(s); SOD = Stand-Off Distance; d50 = median particle diameter; ρ = density; γ = specific heat ratio (gas); μ = flattening ratio; Yp/Ys = yield strength of powder/substrate; CP-Ti = Commercially Pure Titanium; Ti-6Al-4V = Titanium alloy (Grade 5); PEEK = Poly-Ether-Ether-Ketone; ABS = Acrylonitrile-Butadiene-Styrene; PA66 = Polyamide-66; HEA = High-Entropy Alloy; N2 = Nitrogen; MAE = Mean Absolute Error; RMSE = Root Mean Squared Error; MSE = Mean Squared Error; MAPE = Mean Absolute Percentage Error; APE = Absolute Percentage Error; R² = Coefficient of Determination; PCC = Pearson Correlation Coefficient; F1 = F1-score (harmonic mean of precision and recall); s-CAE = stethoscope-CAE; m-CAE = microphone-CAE.

) and allocates them to the four research directions (RD) identified earlier: design for CSAM (DfCSAM, RD 1), material analytics (RD 2), real-time Monitoring & Defect Analytics (RD 3), and Deposition & AI Optimisation (RD 4).

Terms associated with RD 4 (e.g., neural network, model, profile, spray angle) appear 28 times and accumulate the highest aggregate link strength (155), while RD 3 keywords (e.g., cold spray, deep learning, anomaly detection) register 22 occurrences and 94 link points. Together, RD 3 and RD 4 include almost two-thirds of all important keywords, showing the field’s focus on parameter optimisation and in-process quality assurance. By contrast, design-oriented RD 1 and materials-focused RD 2 contribute only 11 and 14 keyword mentions, with much lower link strengths (49 and 62), highlighting clear research gaps in part-level design methodology and alloy development.

The paper-level distribution matches this keyword pattern. Sixteen of the nineteen studies (≈79%) target RD 4 objectives such as predicting layer geometry, coating thickness or deposition trajectory, whereas six papers (≈32%) focus on RD 3 tasks centred on defect detection, thermal imaging or anomaly forecasting. Only five studies (≈26%) explicitly address design issues (RD 1) and four (≈21%) concentrate on material analytics (RD 2). Because several papers cover multiple categories, these shares exceed 100%, but the trend is clear: recent CSAM research prioritises real-time monitoring and data-driven process optimisation, while systematic integration of ML into design guidelines and alloy discovery remains at an early stage.

With respect to algorithms, neural-network-based models are the most common, where fifteen papers use basic feed-forward neural networks, convolutional neural networks or related deep-learning variants, reflecting their flexibility for both prediction and classification. Advanced deep models (Convolutional Neural Networks (CNNs), autoencoders, Long Short-Term Memories (LSTMs)) already appear in about a quarter of the studies, while older methods like support vector machines and regressors are used in only one paper, and Gaussian process regression in two. A small but growing number of studies combine data-driven learning with physics-based knowledge using physics-informed ML, offering a clearer option compared to fully data-driven models.

In practical terms, these findings indicate that the strong methods used in RD 3 and RD 4 are deep-learning-based surrogates and anomaly detectors, which could be used earlier in the process. Combining data from different sources, and hybrid physics and ML methods developed for monitoring may help fill the clear gaps in design for CSAM (RD 1) and material analytics (RD 2). The data was extracted from the identified publications and independently verified by the authors for accuracy and consistency. Any discrepancies were resolved through discussion until a consensus was reached. Outcomes of interest were study-reported performance metrics (e.g., coefficient of determination (R²), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), F1 score (F1)). Additional items extracted included reference, research domain, ML/AI model, data type, research target, material, data volume, and reported open challenges (Table 6). Where necessary, metric names were standardised to common definitions (e.g., root mean squared error (RMSE) vs. mean squared error (MSE)); otherwise, results are presented as originally reported. Considering the studies that used different tasks, datasets, and metrics, results were combined narratively; no merging of results into a single combined number or sensitivity analysis was performed.

4. AI/ML Applications for CSAM

In this section, recent applications of artificial intelligence and machine learning in Cold Spray Additive Manufacturing are categorised into four core research directions, building on a systematic literature review and keyword clustering analysis. The domains include: (RD1) Design for CSAM, which focuses on data-driven design frameworks and inverse modelling to achieve desired deposition profiles; (RD2) Material Analytics & Characterization, focusing on porosity prediction, microstructure–property links, and hybrid simulation plus experiment surrogates for adhesion and flattening; (RD3) Real-Time Monitoring & Defect Analytics, which applies AI/ML for anomaly detection, process state classification, and in situ quality control using sensor and image data; and (RD4) Deposition Modelling & AI Optimisation, where physics-guided surrogates and solver-in-the-loop networks accelerate geometry prediction and inverse search, and CFD-assisted optimisation improves nozzle and recipe performance. This work addresses three critical aspects: evaluation using grouped and prospective data splits to avoid context leakage, parameter identifiability (e.g., stand-off distance), and uncertainty-aware methods with cloud-edge deployment for reliable, low-latency control. The four research domains are discussed below.

4.1. Design for CSAM (RD1)

Design in additive manufacturing has progressed from early layer-wise prototyping of simple geometries to a mature engineering discipline that couples digital design, process physics and data-driven control to deliver structural, functional, and multi-material components with predictable performance. International standards such as ISO/ASTM 52900:2021 have advanced AM development by establishing standardised terminology and process categories. These standards define AM as layer-by-layer material joining from 3D model data and explicitly address geometric tolerances, process planning, and verification requirements [1]. From an engineering design standpoint, AM’s unique advantages (design freedom, shorter lead times) are limited by constraints that are process-specific: melt pool dynamics and thermal gradients in fusion-based methods; flow of liquid resin, speed, and process of hardening in photopolymerization; and impact of particles on the surface and how they compact together in solid-state methods. Accordingly, design for AM has shifted from shape-first modelling to physics-based workflows in which deposition strategies, parameter ranges and post-processing are co-designed with geometry to achieve desired microstructures and mechanical properties, minimising residual stresses and ensuring dimensional accuracy at scale [1,2,3,4]. The core insight is that in advanced AM, design cannot be separated from process physics and control, and geometry alone is an insufficient specification for performance. This insight is demonstrated with a design-for-CSAM loop in Figure 3.

4.1.1. Part Design and Deposition Strategy

Cold spray additive manufacturing narrows this landscape to a solid-state deposition route in which powders (roughly 1–100 μm) are accelerated to supersonic velocities (roughly 300–1200 m/s) in pressurised carrier gases and directed through special supersonic nozzles onto a surface. Bonding is achieved via severe plastic deformation, oxide fracture, and mechanical interlocking rather than melting; thereby, preserving feedstock phase constitution and avoiding heat-affected zones [2,8,42]. Successful deposition occurs only when particle speed exceeds a material dependent critical velocity. This threshold increases with decreasing particle size, increasing surface oxide thickness, and increasing hardness, making it a principal design constraint that couples material state to nozzle flow and gas selection. Critical velocity is the key factor in CSAM, and it must be adjusted together with nozzle shape, gas settings, and particle shape to obtain good bonding while avoiding damage to the surface [2,11,43]. Because cold spray is a continuous jetting process, dimensional control depends on path planning and scanning strategy. Toolpaths must integrate jet footprint models, robot kinematics, and access constraints more stringently than in other AM modalities to avoid over or under build, especially near edges and concavities. Recent 3D volume construction and toolpath frameworks explicitly embed cold-spray characteristics to improve geometric accuracy in repair and freeform builds, confirming that algorithmic planning is a first-order design variable in CSAM [44,45].

Material selection for CSAM includes pure metals, alloys, and metal-matrix composites, with feedstock morphology, size distribution, and microstructure shaping deposition efficiency and the mechanical properties of the components [2,46,47]. Spherical and irregular particles exhibit distinct penetration depths and substrate interactions under otherwise similar conditions, and surface condition, especially oxide coverage modulates effective critical velocity and interfacial bonding [2,8,48]. Multi-material CSAM, including simultaneous powder mixing and metal polymer blends, enables tailored property combinations and, in some cases, synergistic responses that exceed linear mixing rules due to interpenetrating microstructures and load-transfer mechanisms [49,50]. Substrate compatibility is unusually broad because the process temperature is below melting. Metals can be deposited on temperature-sensitive substrates, including engineering polymers and CFRPs (Carbon Fibre-Reinforced Polymer), facilitating metal coating and functional modifications without thermal degradation [51,52,53,54,55]. This low-temperature range is deliberately advantageous, as it allows bonding to polymers and composites that fusion-based AM cannot process, while maintaining bioactive or unstable phases in functional coatings.

Process parameters in CSAM are numerous and interdependent. Gas type, temperature, and pressure set the acceleration field. Helium can be necessary for dense deposits of hard or challenging alloys due to its superior sound speed and thermal properties, although it comes at a higher cost [11,56]. Standoff distance, spray angle, and traverse speed control the variation in impact conditions across the surface. Experiments on Al2219–T6 show that spray angle and traverse speed have stronger effects than standoff distance on thickness and properties, highlighting the impact of directional impact and residence time in jet-footprint development [57]. Nozzle exit diameter, divergence angle, and internal contour interacts with gas heating and mass flow to determine particle velocity distributions and gas temperature at the outlet. Recent 3D-CFD and parametric studies map these dependencies to better align design targets with attainable jet conditions [7,9,57,58]. Because these controls interact nonlinearly, multi-parameter optimisation methods such as response surface methodology (RSM), artificial neural networks (ANNs), and hybrid Artificial Neural Network-Genetic Algorithm (ANN–GA) schemes, are becoming increasingly popular to reduce time and costly trial-and-error [15,59,60,61,62,63].

Industrial applications mirror these strengths of CSAM. In aerospace, CSAM is widely used for repair and remanufacture of aluminium and titanium-based components, restoring geometry and function while preserving underlying microstructure; thus, offering a cost-effective solution for high-value parts with minimum heat input [10,64,65]. For lightning-strike protection, aluminium cold spray coatings on CFRP fuselages can provide surface conductivity and avoid laminate damage when applied over a suitably designed plasma-sprayed aluminium interlayer, which improves bonding and prevents erosion of the composite substrate [66]. In biomedical applications, cold spray is employed to produce wear and corrosion resistant coatings, as well as bioactive ceramics and Ti–hydroxyapatite composites that retain the feedstock phase composition and promote bone integration [54,55,67,68]. Emerging areas include thermoelectric device interfacing and even aspects of solid-state battery manufacturing where low-temperature metallisation is essential [46,69]. Across these claims, CSAM’s key distinction is not only its low heat input but its controllable solid-state plastic deformation at impact, which can be leveraged to retain material properties, preserve phases, and enable compatibility with a wide range of substrates [2,42,64].

Despite these advantages, limitations remain. Porosity control and uniform thickness over complex topologies are challenging as jet flow tends to become uneven near edges and separates over sharp features. Many reports note that producing morphologically precise features is particularly problematic in areas with tight curves or deep concavity [2,3,4]. Geometric complexity also increases sensitivity to fixture orientation, robot path discretisation, and overspray, while economics are impacted by high-pressure systems, helium usage, and nozzle wear, which collectively constrain wider adoption [2,10]. At present, the main barrier to CSAM’s industrial scaling is not fundamental feasibility but achieving reliable, uncertainty-aware control of geometry and porosity in complex builds, along with managing the operational costs of advanced gas-nozzle systems.

4.1.2. Design Optimisation

Within this context, design strategies specific to CSAM increasingly integrate adaptive path planning, deposition order, and post-processing to mitigate geometric and residual stress constraints. Automated scan-to-path frameworks that integrate 3D scanning, defect characterisation, and adaptive trajectory generation have been shown to reduce post-processing and improve deposit conformity in repair settings, highlighting workflow integration as a key vector of design maturity [45,70]. Layer-by-layer build up strategies that explicitly address interlayer bonding and stability improve control over wall straightness and topology in freeform components, provided process ranges maintain impact energies above critical thresholds without destroying substrate integrity [71,72]. Shape-aware deposition methods such as “Metal Knitting (MK)” also show how trajectory design can act as a control tool to reduce anisotropy and shape error [73,74]. MK follows continuous circular nozzle paths at fixed angles < 90°, enable straighter walls and thicker, well-bonded deposits with finer geometric control. Although these factors initially elevate porosity and reduce adhesion due to reduced normal impact, these effects can be mitigated by post heat treatments.

Annealing of copper and aluminium deposits relieves residual stresses, reduces microstructural defects, and significantly enhances cohesion and adhesion. Wear resistance approaching bulk material has been reported after appropriate thermal treatments, confirming the utility of post-processing as a design variable [73]. Complementing these strategies, physics-guided parameter estimation frameworks formalise links from operational variables such as spray angle, traverse speed, powder feed rate, standoff distance, nozzle exit diameter and divergence angle, to deposit profiles through analytical gas dynamics, particle ballistics and impact mechanics. Sensitivity analysis emphasises the dominant roles of nozzle geometry, feed rate, and critical velocity, with standoff distance reducing height via velocity loss [75]. Together, these results support a design approach in which deposition strategy, parameter selection, and post-processing are optimised using models that expose sensitivities and identifiability limits.

Turning to the application of artificial intelligence and machine learning in CSAM, the state-of-the-art spans forward surrogates for track and layer shape prediction, surrogate-based inverse design of parameters, and workflow integration for scan-to-path planning and overspray management. Figure 4 shows the minimal AI and ML pipeline used for CSAM design and compensation.

At the design stage, ML supports printability assessment, tolerance analysis, path planning, track-profile prediction, and inverse parameter mapping, as well as nozzle optimisation, often in combination with rapid convolution-based or Gaussian deposition approximations to reduce experimental cost [2,37,76]. Early data-driven forward models predict single-track shapes directly from spray angle, traverse speed, and standoff distance. For example, a compact multilayer perceptron (MLP) with Bayesian regularisation achieved roughly R² = 0.95 and 0.06 mm mean absolute height error on held-out tracks, outperforming a Gaussian baseline particularly at edges, though accuracy degraded for strongly off-normal angles due to sparse training coverage [23]. The critical limitation here is data coverage rather than model class, indicating that active learning and design-of-experiments (DOE) expansion are likely to deliver larger gains than deeper architectures. To mitigate data scarcity, synthetic datasets generated from validated analytical models have been used to train two-dimensional (2D) and three-dimensional (3D) deep surrogates, yielding sub-percent relative errors in simulation and millisecond inference. When embedded into adaptive slicing with zigzag planning and edge-compensation, these surrogates ran more than 250× faster and met flatness targets while limiting overspray on test geometries [14]. The approach is compelling for interactive design, but accuracy declines at high spray angles and domain edges reflect biases and limits of the underlying analytical solver. Integration with real data remains necessary to close the domain gap.

Inverse design efforts frame CSAM parameter selection as optimisation over fast forward surrogates. A representative study couples a small MLP trained on 36 single-track profiles with a particle swarm optimiser to retrieve impact angle, particle velocity, and standoff distance that reproduce a target geometry. The method returned low profile error within seconds and achieved a near perfect correlation on held-out tracks, but the tiny dataset, narrow material or process coverage, and low-capacity network raise external validity concerns, and standoff distance emerged as weakly identifiable, consistent with physical insensitivity and signalling potential non-uniqueness of solutions [22]. Search-on-surrogate is efficient and practical, but it carries the biases of the surrogate model and cannot guarantee global optimisation without uncertainty quantification or feasibility constraints. Addressing non-identifiability explicitly is essential in production. Beyond tracks, end-to-end workflow integrations now combine reverse-engineering scans, mesh-based deposition simulators using cylindrical jet models, and automatic tool–path generators in unified software components. These reduce handoffs and enable pre-spray verification of layer height evolution, detouring around features to avoid overspray and demonstrating that clean multi-feature can build-up to almost 145 × 95 mm in copper [17]. The principal contribution is integration rather than novel physics, and accuracy is currently limited by simplified jet models and uncertainty bounds. Online corrections are needed to assure robustness across geometries and environmental variation. A related design stage toolkit improves uniform-layer slicing, adds edge compensation, and plans transfers via RRT (Rapidly Exploring Random Tree) to avoid existing deposits, eliminating stray material on multi-feature builds, including a 2.3 kg “hand” made from copper. However, uniform-layer assumptions struggle near steep gradients, and concave contours remain challenging without learned deposition fields [40].

Parallel ML threads address parameter optimisation and nozzle design. RSM, ANN, and hybrid ANN–GA have been applied to optimise nozzle parameters, gas heating, and traverse strategies for deposition efficiency and porosity control. Inverse, theory-guided learning has also been used to set process conditions for difficult alloys and fibre-reinforced composites, with helium often required to suppress porosity in high-entropy and hard systems [12,15,56,60]. Nozzle-geometry optimisation via 3D-CFD improves particle-velocity distributions and identifies dominant geometric controls. However, further work is needed on non-adiabatic effects and environmental interactions to strengthen robustness near the substrate and in free-jet regions [9,58,77]. Across these studies, four risks emerge: dataset sparsity, domain shift when training on synthetic data, lack of uncertainty quantification, and parameter non-identifiability in inverse settings.

Synthesising these findings, a coherent trajectory for AI/ML-enabled CSAM design emerges. First, hybrid physics and data surrogates grounding learning in analytical or CFD models but corrected with targeted experiments, can reduce bias while retaining speed [14,77]. Second, active learning and DOE are needed to fill coverage holes at off-normal angles and near boundaries, where both forward and inverse accuracy degrade [14,23]. Third, uncertainty-aware inverse search with feasibility constraints can address non-identifiability, and produce parameter sets with quantified confidence, suitable for shop-floor deployment [22]. Fourth, closed-loop planning, integrating in situ scanning to correct drift and update paths online should be built upon scan-to-path frameworks to harden workflows against environmental and system variation [17,56]. These four pillars (hybrid surrogates, active data, uncertainty-aware inversion, and closed-loop planning) are the differentiators most likely to move CSAM from lab-grade demonstrations to robust industrial practice. In parallel, design must continue to treat critical velocity, nozzle–gas–particle coupling, and path strategy as interdependent variables, while leveraging CSAM’s key advantage (solid-state impact bonding) and to functionalize temperature-sensitive substrates and preserve phases otherwise destroyed by fusion methods. This would allow the design phase to be framed as a multi-scale optimisation challenge (spanning geometric, thermo-fluidic, and microstructural levels), which can then be addressed through an increasingly integrated combination of physics-based models, data-driven surrogates, and autonomous planning.

4.2. Material Analytics and Characterisation for CSAM (RD2)

4.2.1. From Material Analytics in AM to Quality Assurance

Material analytics and characterisation in CSAM focus on linking process conditions with resulting microstructure, porosity, residual stresses, and mechanical performance. These methods establish the evidence base needed to validate design strategies and to guide optimisation of deposition parameters, heat treatments, and post-processing [78,79,80]. In metallic AM, this progression is marked by a shift from monolithic, isotropic design assumptions to designs that explicitly account for anisotropy and the characteristics of inter-track and inter-layer interfaces, whether in fusion routes (e.g., SLM) or solid-state routes (e.g., cold spray), with hybrid process chains leveraging the strengths of each process to meet structural requirements [80,81]. In this framing, DfAM is inseparable from metrology and materials characterisation, since the credibility of any design claim depends on reproducible links between processing parameters, local morphology, phases, and the resulting mechanical response, including failure modes [78,82]. For solid-state deposition, where bonding is driven by high-velocity impacts and severe plastic deformation rather than melting, geometric intent must be balanced with impact mechanics, critical velocity, particle-size thresholds, and post-deposition thermal history, all which shape defect populations (e.g., porosity). Therefore, design should not merely be treated as a geometry creation process, but as a coupled optimisation across feedstock state, trajectory planning, energy delivery, and heat treatment, underpinned by characterisation workflows that can detect anisotropy, residual stress, and bonding quality with sufficient fidelity to inform iterations.

4.2.2. CSAM: Solid-State Process, Properties, and Characterisation Challenges

As defined, CSAM is a solid-state process; in which particles accelerated by a propellant gas impact a substrate and consolidate via mechanical interlocking and metallurgical bonding under severe plastic deformation, enabling dense metallic builds at comparatively low temperatures, with reduced oxidation compared with fusion processes. Bonding requires exceeding a material-dependent critical velocity and size, which are material dependent and interact with particle morphology and oxide state. Atomistic and continuum studies underline trade-offs between critical size and velocity that directly shape process ranges and trajectory design [83]. Post-deposition heat treatment (HT) mediates recovery, recrystallization, and grain growth, often improving inter-particle bonding and ductility while relaxing work-hardening-driven hardness [78,84,85].

Design for CSAM is constrained by anisotropy across multiple scales, from splat and track interfaces to full layer stacks, influenced by impact angle, scan strategy, and local heat buildup. Empirical studies show that cross-hatching can mitigate anisotropy compared with simple bidirectional passes [86]. Porosity and insufficient inter-particle cohesion remain dominant defect modes, with mitigation through parameter optimisation and HT, yet persistent pores frequently originate during deposition from imperfect inter-particle contact and cannot be fully eliminated post-facto, underscoring the importance of precise primary process control [73,85,87]. Residual stresses are typically compressive, reflecting peening-like plasticity, but their quantification at component scale remains challenging, demanding careful selection and calibration of measurement methods.

4.2.3. Materials, Techniques, and Property Performance Relationships in CSAM

CSAM research is dominated by aluminium alloys, including Al6061 (gas-atomized vs. solution-treated), Al F357 (effects of powder heat treatment), Al7075 (deposit anisotropy), and advanced Al grades characterised for CSAM-specific applications [84,86,88]. Copper and its alloys are frequent subjects for anisotropy studies and composite development (e.g., Cu–Ti), while Inconel 718 freeform builds demonstrate structural potential and pure Ni has been probed via resonant ultrasound spectroscopy for magneto elastic behaviour. Mg coatings extend the palette and illustrate process versatility [80,89,90,91,92]. Collectively, these studies converge on the central design focus of CSAM for maximising impact-driven bonding and densification without incurring damage, while orchestrating scan strategies and HT to fine-tune anisotropy and residual stress toward target performance.

In CSAM, checking and analysing the material is central to quality control because it shows how process settings affect the microstructure and final properties. Most studies still rely on familiar tools like optical microscopy (OM), scanning electron microscopy (SEM), and microhardness testing. These methods are easy to use and provide quick results, but they depend heavily on the operator and the sample chosen, which makes it harder to automate or compare results between different studies [73,93,94]. Electron backscatter diffraction (EBSD) adds another layer by mapping crystal orientation and phases, which is especially useful for assessing heat treatments. Still, its accuracy depends on fine details such as step size, pattern clarity, and surface preparation. If these are not well controlled, the resolution drops, and predictive models based on the microstructure can fail, exposing a significant weakness for data-driven workflows.

Porosity is the main factor behind strength loss and crack initiation in CSAM. It is most often measured through 2D cross-section image analysis, while X-ray micro-CT is used when full 3D detail and non-destructive inspection are required. The method chosen has a real impact on measurement accuracy and on how process ranges are interpreted [2,73,95]. Heat treatment can reduce porosity, but defects caused by poor splat or inter-layer bonding during deposition often remain. This means HT should be considered as a secondary densification step rather than a primary solution. Hardness is typically assessed by micro or nano indentation, which reveals gradients linked to local work hardening and recovery. In metal-matrix composites (MMCs), however, interactions with the reinforcing phases can affect results, making it difficult to compare hardness across different materials [73,93].

Mechanical performance in CSAM is usually evaluated through tensile testing, often conducted in multiple directions to capture anisotropy, and complemented by fracture and fatigue assessments. In many cases, as-sprayed deposits need post processing to approach the mechanical response of materials behaviour, consistent with the microstructural findings noted earlier [4,73,95]. Standard adhesion tests, such as ASTM C633 (tensile adhesion test, TAT), are suitable for coating-type geometries but do not reflect the bulk nature of CSAM, where stress states and failure mechanisms differ significantly. This mismatch highlights the need for alternative specimen designs that better capture adhesion and bonding behaviours [73]. Residual stresses are typically compressive and are commonly measured using incremental hole drilling in line with ASTM E837-20. Other methods present trade-offs: thin-coating curvature (Stoney-type) approaches fail to scale to bulk builds; X-ray diffraction provides only near-surface stress information; and neutron diffraction, while capable of probing stresses deep within the material, is resource-intensive and limited to crystalline systems.

Alongside the standard techniques, newer analytical methods are being explored. Resonant ultrasound spectroscopy (RUS) has been deployed to extract elastic properties and magneto-elastic softening in cold-sprayed Ni, illustrating non-destructive paths to correlate stress state and bonding, although with calibration needs and material constraints [91]. Fractography remains central for clarifying failure modes and to diagnose interface weaknesses [87,90]. Meanwhile, machine-learning (ML) and computer-vision (CV) tools are beginning to deliver objective microstructure classification, defect detection, and property prediction, including similarity metrics for micrographs (e.g., Structural Similarity Index Measure; SSIM). However, barriers remain, such as limited high-quality labelled datasets (especially for rare defects), computational costs, challenges in integrating with industrial workflows, and a persistent interpretability gap for safety-critical decisions [2,62,87,96,97]. Integrating robust experimental measurement with ML-assisted analysis will be practical to mitigating anisotropy, optimising process monitoring, and improving strength, ductility, and durability in CSAM components, provided data and evaluation protocols are rigorous.

Within this evidence base, the materials investigated, and the associated characterisation methods together form a coherent picture. Aluminium alloys are most common: Al6061 appears repeatedly in atomized and heat-treated powder conditions, Al F357 is used for powder HT effects on microstructure and mechanical properties, and Al7075 is used to examine anisotropy in deposits. Beyond these, advanced Al alloys have been systematically characterised for CSAM [82,84,88,89]. Ti-6Al-4V is the second most-studied CSAM alloy owing to its aerospace relevance. Studies link powder morphology to coating properties, detail heat-treatment-driven microstructural progress and demonstrate hybrid CSAM–SLM routes that broaden the process design space [81,85,87]. Copper and its alloys support anisotropy analyses and composite development. For example, Cu–Ti composites achieving almost 15% elongation after 400 °C HT while maintaining tensile strength above 270 MPa, a combination that illustrates the ductility gains available via HT without catastrophic strength loss [90]. Inconel 718 free-standing components have reached tensile strengths close to 338 MPa after optimisation, demonstrating the potential for structural applications, while pure Ni has served as a platform for resonant ultrasound spectroscopy (RUS)-based elasticity and stress interrogation. Coatings of magnesium and other light alloys demonstrate the extension of CSAM to more reactive systems [80,91,92].

SEM is for particle morphology, cross-sections, and fracture surfaces. Energy dispersive spectroscopy (EDS) augments elemental interpretation, while X-ray diffraction (XRD) supports phase and crystallographic analysis. EBSD offers grain structure, orientation and bonding features, and optical microscopy is still widely used for general microstructural features such as dendrite size [80,84,87,89]. Nanoindentation has emerged as a flexible method applicable to both feedstock and consolidated material. Vickers microhardness is consistently deployed, tensile testing quantifies strength and ductility, fatigue tests explore durability, and fractography remains essential for diagnosing failure mechanisms [80,87,89,90,98]. Key microstructural concepts include severe plastic deformation at particle interfaces and localised solid-state bonding. Heat treatment generally promotes recrystallization (sometimes accompanied by grain growth, depending on material), improving bonding and ductility. Anisotropy remains a recurrent feature, with equiaxed particle footprints in the XY plane, contrasting with lens-like cross-sections in XZ/YZ planes. Cross-hatching scan strategies have been shown to reduce these anisotropic effects compared with simple bidirectional scans [80,84,89,92]. Defects in CSAM are often linked to porosity and weak bonding between particles. Annealing can reduce porosity up to approximately 30% in specific settings, while irregular powder morphology has sometimes produced lower porosity than spherical powders, opposite to trends in fusion-based AM, reflecting the nature of CSAM [73,87]. Process property relations are now sufficiently consistent to inform parameter linkage. For example, higher propellant gas temperatures (almost 800 °C) improve particle deformation and bonding but remain material-specific and sensitive to feedstock and trajectory choices [90].

This study demonstrates strengths such as broad coverage across metallic systems, the use of multiple methods (microscopy, diffraction, indentation, and mechanical testing), and an increasing consistency between experimental observations and mechanistic or atomistic models [83]. Limitations include frequent single-material and single-parameter studies that limit generalisation, challenging quantification of residual stresses, and in some cases small sample sizes in various reports [3,91]. From an application perspective, the evidence indicates that CSAM can be used for structural parts when supported by careful process optimisation and heat treatment. Hybrid approaches offer pathways to geometric complexity with microstructural control and functionally graded properties. Future directions include multi-material and hybrid builds, advanced non-destructive evaluation, and physics-based analytics that can translate microstructure to reliable property predictions.

4.2.4. Data-Driven and Physics-Informed Analytics for CSAM: Porosity Prediction and Surrogate Modelling

On porosity, ref. [20] assembled a “mega-dataset” by mining approximately 35 papers to extract a few hundred porosity labels with around 14 process and material features, emphasising single-powder cases for completeness. Measurement diversity across CSAM studies is useful, but differences in testing protocols, imaging methods, and porosity definitions introduce noise and cross-study variation that weaken model reliability. The study utilised tree aggregate models (RF, XG, and CatBoost), which are robust to missing distributional roughness. These models can handle absent values but cannot correct systematic bias or inconsistent scaling across sources, which must instead be addressed during dataset curation. After feature selection, imputation, and 10-fold cross-validation (CV), several models keep average errors within around 0–2% porosity on a held-out split, and the best RF reaches RMSE around 0.8%, notably better than common image-analysis repeatability. However, random data splits can sometimes inflate performance via information leakage from the same paper (known as same-paper context leakage), and the authors also note that using grouped CV (e.g., leave-one-paper-out) is a more reliable and fairer test of generalisability. Moreover, the reported uncertainties in gas temperature and pressure (about 100 °C and around 1 MPa) should be properly included and propagated (e.g., via Monte-Carlo simulations or interval predictions) rather than treated as exact single values. Despite these limitations, the work delivers a fast, millisecond-scale porosity predictor and a carefully prepared ML workflow that is well-suited for screening and process condition exploration, provided that grouped CV, feature-importance stability analysis, and prospective validation on new laboratory data are performed before deployment. The key insight is that data evaluation design on model choice, and that uncertainty quantification is necessary to map predictions onto process decisions [20]. In this context, Figure 5 shows the minimal AI/ML pipeline, referencing porosity, adhesion, and stress analytics.

For physics-informed surrogates targeting particle flattening and penetration, models for micro-features linked to adhesion have been developed and trained on hybrid datasets with approximately 70% FEM impacts and around 30% SEM measured experiments across metals (Cu, Al, steel, and Ti) and polymers (PEEK, ABS, PA66) [12,21,35]. Hybridization reduces data collection cost, but combination of simulation and experimentation can cause biased mappings unless treated explicitly via multi accuracy or domain-adaptation strategies [21,35]. Their models use minimal input, impact velocity, and yield strengths of powder and substrate, which capture first-order mechanics. However, they leave out other particle characteristics (size, shape, temperature, oxide condition, and impact angle), limiting predictive accuracy. On FEM-only data, GPR achieved the lowest test errors, especially for penetration, reflecting smooth relationships in simulation space. However, when both simulated and experimental data were combined, a small (Moderately complex) model neural network performed better than GPR for predicting penetration, while linear regression slightly outperformed the others for flattening. This outcome suggests that the flattening behaviour may depend on only a few key variables (i.e., has limited effective dimensionality), or that the remaining differences between simulation and experiment (residual simulator biases) can be well captured by a simple linear model. Ref. [35] study evolves compact deep networks via genetic algorithms: a WNN for flattening reaches around 0.73% RMSE with approximately R² = 0.92, while a TNN for penetration performed poorly (RMSE ≈ 2.6%, R² ≈ 0.60) consistent with penetration’s greater nonlinearity and sensitivity to missing descriptors. The authors argue that DL surrogates can replace many FEM runs and enable rapid exploration of new material pairs. As with porosity, safety-critical adoption requires calibrated uncertainty and error budgets defined relative to adhesion thresholds, along with grouped-CV by material and substrate family and validation tests on unseen powders and substrates. The critical insight here is that input sufficiency and evaluation protocols, not just model architecture, determine practical utility, and that multi-output, physics-constrained models with active learning to target high-value experiments are likely to deliver larger gains than simply increasing network complexity.

Synthesis indicates that well-prepared large datasets can power fast porosity prediction and that hybrid simulation plus experiment surrogates can estimate adhesion features, but the main risks are domain shift (simulation-to-experiment), inconsistent labels from mined papers, missing physical descriptors (size, temperature, oxide state, angle), and bias evaluation from random splits. For CSAM designers and process engineers, a pragmatic path forward is clear. First, adopt grouped cross-validation and prospective validation (e.g., leave-one-paper-out, lab, material) to test transfer. Second, use multi-accuracy physics-guided learning to bridge FEM–experiment gaps. Third, expand descriptors with particle morphology, thermal state, oxide condition, and impact angle. Lastly, implement calibrated uncertainty to ensure that predictions map to porosity and adhesion thresholds and associated decision policies [20,21,35]. In parallel, robust, standardised measurement pipelines, covering porosity imaging, EBSD step size, pattern quality, and residual stress calibration are essential to reduce label noise and ensure that ML-based analytics are transferable across labs and material systems [2,3,94,97].

4.3. In Situ Control and Monitoring in CSAM (RD3)

The move toward adaptive and self-aware AM has placed in situ sensing, monitoring, and control at the centre of contemporary research and industrial adoption. Historically, AM quality assurance relied on post-process inspection. However, the demand for higher build rates, certification of critical parts, and reduced material waste has shifted focus to real-time observation and intervention during the build itself. Across different AM processes, the core objectives are consistent, i.e., maintain process stability, detect and prevent defects as they form, and ensure repeatability through feedback loops that connect sensors to models and control [76,96]. The sensors include thermal, optical (vision), and acoustic, while the algorithms range from rule-based diagnostics to ML-based estimation and optimisation. The key insight is that sensing alone is insufficient and the real benefit comes when measurements are combined with models and used to trigger timely and stable control actions. Such a workflow with basic in situ monitoring, decision, and action steps is shown in Figure 6.

This integration remains uneven across AM processes and is often limited by data quality, complexity of models, and the challenges of system integration. However, as CSAM moves from coatings toward bulk builds, in situ control and monitoring become crucial for stabilising a process dominated by high momentum particle jets, impact bonding, and constantly changing surface geometry [2,99]. Compared with melt-based AM, CSAM’s solid-state nature removes melt-pool signatures but introduces process-specific challenges, such as measurement in harsh, particle-laden flows, deposition efficiency that depends on particle trajectories, heat accumulation tied to geometry and path, nozzle wear, and tight requirements for nozzle standoff and traverse speed to achieve uniform layers [2,9,100]. These features limit direct transfer of sensing and control strategies from thermal spray or laser powder-bed fusion. They also highlight the need for robust, high-signal-to-noise sensors, fast data reduction, and low-latency control methods adapted to supersonic flow and impact mechanics.

CSAM monitoring and control remain early-stage, with advances in multi-modal sensing (thermal, optical, acoustic, geometric), CFD-based modelling with empirical surrogates, and emerging AI/ML decision systems. The key challenge and opportunity lie in integrating these three components [2,25]. Thermal sensing is the most established approach. Infrared pyrometers and thermal cameras provide real-time surface temperature data. However, their calibration is complicated by emissivity sensitivity and by changes in surface conditions, especially over complex geometries and under varying oxidation states [100]. Optical diagnostics, notably the HiWatch HR platform, use sizing-particle tracking velocimetry (S-PTV) to measure in-flight particle velocities, trajectories, and size distributions and to produce 2D velocity scan maps that operators can use for process monitoring [26,101]. These systems provide detailed insight into jet quality but face harsh environment constraints such as particle occlusion, limited measurement volumes, and fouling of observation windows.

In practice, airborne acoustic sounds in CSAM are generated within a supersonic jet environment where turbulent mixing noise, shock-associated noise, and screech can mask weaker process cues. Nozzle wear and geometry shift further alter the sound pattern and direction, making exact monitoring hard and pushing the need for more advanced metrology and feature engineering. Recent CSAM work shows that low-frequency (0–20 kHz) microphones can still notice changes in nozzle gas pressure/temperature, powder feed rate, and nozzle wear via power-spectrum analysis, and that sliding-window derivatives of power in the time domain show slow (e.g., throat wear) versus sudden (e.g., delamination) problems, while being fairly unaffected by microphone position once power is balanced. To handle noise and overlapping sources more widely in AM, reviews highlight background-noise removal and strong feature picking to separate process-important signals from time-changing environmental effects [102].

To strengthen reliability, other machining processes (especially milling and turning) also fuse acoustic emission (AE) with vibration/force signals (and process settings) for quality prediction. In precision grinding, mixed deep models that use both sound emission and vibration signals (plus process settings) reach very high accuracy for surface-roughness (SR) prediction. For example, a DBO-optimised 1D-CNN-LSTM showed R² ≈ 0.991, with MAE ≈ 0.005, RMSE ≈ 0.0067, and MAPE ≈ 4.9%, while an SVR model using multi-domain features from CEEMDAN and PCA merging reached R² ≈ 0.995 with MAE ≈ 0.002 and RMSE ≈ 0.005; both clearly gain from adding grinding settings alongside sensor signals [103,104]. Multi-sensor combination also improves tool/state classification. For example, combining AE, force, and acceleration for grinding-wheel wear recognition resulted in over 92% accuracy with a better SVM [105]. Beyond prediction, explainable ML helps show what the combined models learn. In machining, a sequence-to-sequence fusion network linked acoustics and vibrations to machining-power signals with only 2.5% error (vs. 5.6% from acoustics alone and 8.2% from vibration alone), and Integrated Gradients showed which time regions/signals most shaped the predictions [106].

Similarly in AM, XAI (Grad-CAM on a 1D-CNN) highlighted frequency bands (≈5–6.5 kHz; ≈10–12.5 kHz) that correlate with laser motion and printing position in LPBF while yielding strong classification rates, demonstrating how physically meaningful spectral regions can be surfaced from acoustic data [105]. Taken together, these results indicate a clear opportunity for CSAM to combine aeroacoustics with complementary sensors (e.g., vibration, force, thermal) and adopt interpretable AI to ensure that anomaly flags and quality predictions can be linked back to specific, physically plausible frequency-time patterns, improving both reliability and insight.

Moreover, geometric sensing has advanced through laser profilometry. A custom nozzle-tracking system used in situ track-shape evolution across the spray duration, providing real-time insight into the deposition efficiency, layer stability, and demonstrating that skewed Gaussian functions represent track profiles better than symmetric alternatives [107]. Other systems, such as a Keyence LJ-X8000 (Keyence, Japan) mounted on the robot for pre-spray scans and per-layer captures, detecting under-build or heat accumulation and triggering path adjustments or pauses as needed, demonstrate a practical closed-loop system even without ML, with strong potential for future uncertainty handling and predictive control [17].

At the system level, integrated low-cost Industrial Internet of Things (IIoT) setups that stream optical and thermal images along with N₂ temperature, pressure, flow, and robot triggers, have been demonstrated for layer-wise quality assurance. Lighting setups are tailored (side lighting for porosity and geometry and ring lighting for oxidation), and downstream CNNs deliver very high F1 scores near 0.998–0.999 on internal splits, including 100% accuracy on a CuZn30 test set [38]. While cost-efficiency and open-source tooling are impressive, such high metrics invite caution about data leakage or overly easy test conditions. Long-term stability, lighting stability, and retraining schedules must be validated for production deployment. To reduce fragmentation across sensing modalities, Figure 7 presents a simple multi-sensor fusion action pipeline for CSAM.

4.3.1. Modelling Approaches: CFD, Surrogates, and AI/ML Integration

Computational fluid dynamics (CFD) remains the dominant mechanistic route for understanding gas dynamics and particle motion in CSAM. Two-dimensional axisymmetric models, often based on one-dimensional isentropic gas-flow relations with Lagrangian particle tracking, can predict particle acceleration, temperature histories, and exit velocity distributions and show within 5–10% agreement with experimental measurements [6,7,15,61,63,108,109,110].

Empirical approaches, especially Response Surface Methodology (RSM) with Central Composite Designs remain prevalent and often deliver quadratic regressions with R² > 0.96 for parameter screens and trade-off studies [61,63]. Fast, one-dimensional analytical models can describe deposition ranges as functions of pressure, temperature, particle size, and standoff. Recent work introduced clear metrics, maximum ideal particle velocity (γ), acceleration capacity (β), and free-jet deceleration (α), to compare process regimes and nozzle designs efficiently [6]. For geometry, coating-thickness simulators exhibit reliable accuracy on curved surfaces and skewed-Gaussian track models outperform symmetric fits [5,107].

AI/ML usage is expanding in both modelling and monitoring. Gaussian Process Regression with physics-informed mean functions that include prior knowledge about surface geometry (geometric priors) improve overlapping-track profile prediction by blending these geometry-based assumptions and flexible residual learning [13]. Combination of Latin Hypercube Sampling, neural networks, and genetic algorithms achieve particle velocity optimisation with almost 4% error, underscoring the efficiency of hybrid design-of-experiments plus machine learning [111]. Edge-deployed computer vision advances, combining visual transformer-based video super-resolution with fully convolutional networks, report 96.34% accuracy for multi-object extraction at near 118.83 ms inference times, showing that feasibility of real-time, high-resolution perception at the toolhead is feasible [112]. In broader AM, deep mixed-effects frameworks that integrate neural networks with random-effects terms have improved real-time thermal monitoring and anomaly detection, a transferable concept for CSAM where nozzle, powder, and path introduce layered variability [113].

Beyond modelling, pre-run recipe predictors for cold spray porosity and quality have matured. On a 35-sample Cantor-alloy dataset (22 literature + 13 new records, with five preset descriptors: gas type, gas pressure/temperature, powder type, standoff), linear-kernel SVM and ordinary squares outperform deeper learners (PCC ≈ 0.85 and 0.83; RMSE ≈ 2.1% porosity), and SHAP analyses prioritise gas and powder type, then temperature and standoff. However, the small, mixed dataset invites heterogeneity and overfitting risks, making leave-one-paper-out validation more appropriate than random cross-validation for deployment claims [39]. On a larger 227-record literature dataset covering 35 studies and seven descriptors, decision trees achieved R² ≈ 0.75 (MAE ≈ 2.9%) on ten unseen cases, whereas linear regression lagged (R² ≈ 0.27). It should be noted that testing on only ten cases is statistically weak, and label noise and inconsistent protocols limit achievable accuracy [16]. These models are millisecond-fast and can help filter out risky parameter settings before a run, but they are not substitutes for in situ sensing and should also provide calibrated uncertainty to avoid false “go and no-go” outcomes.

4.3.2. In Situ Monitoring and First Steps Toward Closed-Loop Control

Signal and vision-based irregularity detection illustrates the feasibility of layer-wise fault detection and response. By attaching a stethoscope on the powder hopper to capture clean internal vibrations, one study converted audio to log-mel spectrograms and used an autoencoder for anomaly detection, followed by a CNN to classify “no powder,” “clogging,” and “no gas.” On short blind tests, recall reached 100% with near 97% accuracy for the stethoscope signal, while an external microphone recalls only near 34%, showing the clear SNR advantage of mechanically coupled sensing [34]. Integrated-gradient attributions highlighted important frequency bands at 600–1600 Hz, offering interpretability, but portability across guns, powders, and rigs and lifecycle issues (mounting, calibration, clog-detection delay) remain to be proven.

Thermal forecasting via ThermoAnoNet uses a single FLIR A655sc (50 fps) and educes each frame to average substrate temperature. An N-BEATS forecaster trained with Bayesian-tuned thresholds, detects mid-build feed-rate jumps with near 96% accuracy (F1 ≈ 0.97) on validation and near 90% (F1 ≈ 0.94) on eleven blind tests, typically within almost 0.3 layer; performance degrades for subtle 2–4 rpm changes (recall ≈ 0.60) [19]. Temporal-forecast thresholds performed better than RNN, Temporal Convolutional Networks (TCN) and statistical baselines in F1 and detection delay without early alarms, but emissivity calibration, lens contamination, and camera placement remain critical for robustness.

Most importantly, closing the loop has been demonstrated in practice, with robot-mounted laser profilometry enabling pre-layer scans and per-layer checks. When heat build-up or under-build is detected, trajectories are adjusted or paused in real time, and data feeds back for process and path optimisation [17]. Even without ML, this architecture delivers layer-wise sensing and immediate corrections. Moreover, adding uncertainty thresholds and learned predictors could reduce false stops and give more accurate corrections.

Decades of work in thermal spray show that nonlinear links between parameters, in-flight particle states, and deposit responses are learnable and controllable. Early neural networks mapped these relationships [114], later embedded in process-control equipment for adaptive responses [115], and eventually produced repeatable builds (e.g., thin-walled moulds) in wire-arc spraying [116]. More recently, big-data frameworks have supported quality monitoring and laid pathways to adaptive control, contingent on standardised, high-quality data pipelines [117]. In CSAM, automated path and trajectory planning for repairs show the link between sensing and decision-making for in-process optimisation, though long-term robustness under drift, wear, and environmental variation remain underreported [70].

Against this background, persistent issues remain. Data availability and labelling quality are major limitations. Large, diverse, well-labelled datasets covering defect types, materials, and printers are expensive to build [76,96,118,119]. The “black-box” nature of powerful solvers makes root-cause analysis and operator trust difficult, even though attribution tools such as SHAP and Integrated Gradients exist [34,39,96]. Real-time AI also requires significant computing at the edge and careful software and hardware integration into different spray systems [96,112] In CSAM specifically, maturity is uneven. Intelligent monitoring is uncommon in commercial systems, traditional controllers in thermal and cold spray often cannot detect rapid quality variations induced by hardware fluctuations, and anisotropy and porosity persist as quality risks, needing both process and ML mitigation [2,76,86,94,95,120,121]. Process efficiency and control also depend on gas selection, blending, nozzle design, and substrate response, showing the close links between hardware, flow physics, and quality metrics [9,27,48].

Taken together, the studies show that temperature, optical S-PTV, aeroacoustics, and laser-profilometry sensing can each provide useful signals for CSAM. CFD, analytical surrogates, and empirical RSM are still effective for design-space exploration and nozzle and path optimisation. Moreover, AI/ML can pre-screen recipes, detect faults, and, in some cases, close the control loop with low latency. The most significant limitation is fragmentation, where most systems are demonstrated separately, a thermal camera in one case, an audio classifier in another, without being connected to stable, latency-managed controllers that can guarantee corrective actions under complex geometries [25].

A clearer path for progress emerges from the evidence base. First, validation must match deployment, with grouped and prospective evaluations (leave-one-paper-out, lab) required to estimate real generalisation beyond random splits, especially for the literature-mined datasets. Second, multi-modal sensing, for example, combination of audio and thermal and geometry, with calibrated uncertainty can reduce single-sensor weaknesses and cut nuisance alarms. Third, active learning and online adaptation should be standard to handle drift from nozzle wear, powder lot changes, and environmental variation. Fourth, true closed-loop control audits must report allowed delay, stability margins, and recovery effectiveness for specified fault classes and geometries, enabling certification level assurance. In short, the opportunity in CSAM is to connect robust sensing into physics-based, data-efficient models that act quickly and in a stable manner, turning today’s monitoring prototypes into production-ready control systems.

4.4. Process Modelling and Optimisation (RD4)

4.4.1. From Model-Driven Understanding to Optimisation in AM

Process modelling and optimisation have long supported progress in AM, providing a structured way to connect process parameters to microstructure and, ultimately, properties and performance. As AM expanded beyond single-physics and single-material systems, predictive models moved from empirical relations to physics-based and hybrid approaches that can capture combined thermomechanical effects and enable proactive control and design-space exploration [113]. Within this paradigm, modelling works as the main analytical tool for establishing process–structure–property (PSP) relationships, while optimisation strategies turn this understanding into parameter settings, build strategies, and cost–quality balances. They shift from trial-and-error to model-based exploration, driven by the rising cost of extensive experiments and the need for reliable results across materials and geometries, motivating the use of simulation, surrogate modelling, and data-driven optimisation throughout the AM lifecycle. Figure 8 summarises the modelling, optimisation, and validation loop used across this section.

At the same time, AI/ML has become an important tool that links with physics-based modelling at many stages for design of experiments, in situ monitoring and control, fault detection, and post-process quality prediction. Neural networks, support vector machines, and sophisticated solvers have been used to predict porosity, adhesion, and hardness, often performing better than simple regressions, while newer methods such as deep mixed-effects modelling and edge AI aim to deal with batch variations and provide on-device, real-time decision-making. However, dataset scarcity, mixed data sources, and limited external validation continue to limit generalisation. This highlights the importance of physics-guided features, constraints, and strong validation methods.

4.4.2. Modelling and Optimisation for CSAM

Cold Spray Additive Manufacturing creates unique modelling needs that are very different from fusion-based AM. The process couples compressible gas dynamics, nozzle flow development, particle acceleration, particle–gas and particle–wall interactions, high-rate plasticity, and adiabatic heating upon impact, followed by path-dependent deposit build-up shaped by local motion and shadowing effects [6,7,9,108,109,122]. CFD tools model nozzle–gas–particle interactions and velocity and temperature fields; explicit finite-element analysis (FEA) simulates impact, deformation, and bonding; analytical models define deposition conditions and critical velocity conditions; and geometric and kinematic models describe track-and-layer-scale build-up [5,6,7,15]. In addition, surrogate and data-driven models expedite design-space exploration and support inverse design, especially when combined with robot programming and toolpath planning [13,14,24,123].

Process-specific optimisation is challenging because of the strong interdependence among key parameters, such as gas temperature and pressure, nozzle geometry, standoff distance, traverse speed, and spray angle. The process must achieve material-specific critical particle velocities while maintaining geometric accuracy, yet the build quality is highly sensitive to local tool orientation and shadowing effects. Furthermore, the coupling between the nozzle, gas, and particles introduces multi-objective trade-offs among particle velocity, temperature, heat flux, and surface pressure, all under economic constraints related to gas selection and blending strategies [9,15,27]. Moreover, integration of in situ monitoring with model-based control is still at an early stage, with many diagnostic systems working as standalone tools rather than being part of closed-loop control frameworks [25,26].

4.4.3. Synthesis of Findings on Modelling and Optimisation Approaches for CSAM

A clear picture emerges across analytical, numerical, surrogate, and AI/ML approaches, each offering different strengths but also showing limitations that hold back industrial use. One-dimensional analytical formulations and deposition monitoring analyses remain valuable for quick screening and for defining the limits of feasible operation [6]. CFD has advanced to give detailed understanding of gas–particle interactions and their sensitivity to nozzle design, supporting systematic optimisation of de Laval and CD nozzles [7,15,63,110]. Explicit dynamic FEA, often with Johnson–Cook laws, penalty contacts, and adiabatic heating, explains impact physics and bonding, but at very high computational cost, which limits wide parameter studies [12,21]. Notably, CFD-based design-of-experiments combined with ML can achieve multi-objective improvements (nearly 9.1% in particle velocity, 4.4% in particle temperature, 17.9% in heat flux, and 27.6% in surface pressure), showing the benefits of hybrid RSM, ANN, and GA approaches for nozzle optimisation [15]. However, validation is often tied to narrow operating ranges or specific materials, raising questions about cross-material generality.

Surrogates built FEM and experiments greatly reduce evaluation time, allowing fast “what-if” analysis and reverse searches. A representative study used 2D axisymmetric ABAQUS impact simulations across velocities of 50–350 ms⁻¹ for Cu, Al, and 316 L on PEEK and ABS with ∼70% FEM and ∼30% SEM data to train linear models, Gaussian process regression (GPR) and a small neural network. GPR performed best for penetration when trained only on FEM data, while the neural network outperformed GPR when experimental data were included, suggesting penetration behaviour is more nonlinear than flattening, which is lower-dimensional [21]. A follow-up used GA-designed deep surrogates (wavelet neural networks (flattening) and triple neural networks (penetration)) across broader material/substrate sets, where flattening accuracy was strong, but penetration generalised poorly without richer inputs or physics-guided constraints. It was highlighted that claims of predicting novel combinations would need tests outside the training set with calibrated uncertainty [35]. Analytical plus linear surrogates work well for quick trend analysis but reach accuracy limits, especially for flattening where complex multivariate effects dominate [12].

Purely experimental surrogates show impressive speed for reverse parameter prediction. A small MLP trained on around 2.4 k pointwise heights from 36 single-track profiles enabled particle-swarm optimised inversion for angle, speed, and standoff in around 3 s with median errors close to 0.06 mm; yet parameter identifiability problems, especially low sensitivity to standoff, and training on a single setup limit generalisation [22]. The literature-based porosity models trained on 242 measurements with 14 descriptors provided best random forest results of around RMSE = 0.8%, showing potential as planning aids, but mixed data sources and random splits may inflate accuracy, calling for grouped validation and forward-looking tests [20]. Physics-guided surrogates (e.g., Gaussian-profile (corrected by Gaussian-process residuals), or ANNs augmented with analytical features) improve data efficiency and stability and are fast enough for real-time planning but require proof of closed loop use and robustness across surfaces and materials [13,37]. At the extreme of speed, one-shot ANNs map process inputs directly to multi-layer track geometries with R ≈ 0.996 and RMSE ≈ 0.022, an excellent fit but at risk of failing outside the training range [24]. Hybrid solvers plus network frameworks keep physical consistency while cutting run-times from hours to seconds, reaching ∼99.5% flatness with <0.5 mm overspray on a concave part, though results near edge cases and across materials are still scarce [14,123].

Deposition efficiency gains of about 18% have been achieved through step-by-step monitoring and parameter tuning, and real-time spray monitoring allows verification before coating production, improving consistency and reliability [26]. Process parameter ranges provide practical guidelines. For example, helium at around 700 °C and 25 mm spray distance for Fe, Co, Ni, Cr, and Mn (high-entropy alloys), and 6 bar/400 °C for coating polymers [48,124]. Despite these advances, most monitoring systems still operate as independent tools rather than as parts of model-based control, with multi-modal data fusion slowed by different sampling rates, data formats, and processing needs, and systematic model updates from monitoring feedback are still rare [25,26,101]. Feedback control is basic (temperature feedback and laser-assisted heating in the literature are the most developed). While feedforward control based on predictive models shows promise but depends on sophisticated models and dependable sensors [113]. Lessons from other spray processes suggest that embedded neural controllers and fuzzy-logic rules can improve reproducibility, but their transfer to CSAM’s solid-state physics is still under-tested [114,115,116,125,126,127]. Recent CSAM-specific trajectories show automated geometry-based path planning and orientation control for complex surfaces but closed-loop corrections during spraying are not consistently demonstrated [44,70,72,128]. The main bottleneck is the connection between robotic trajectory planning and sensor-driven, ML-based controllers that can keep deposition stable and geometrically accurate under disturbances.

In-line nitrogen–helium blending is a practical way to optimise coating cost while maintaining required deposition quality thresholds. A cost model can select the gas ratio that minimises overall coating cost for a given job [27]. Post-processing and hybrid routes, including laser-assisted cold spray (LACS), heat treatment, and hot isostatic pressing (HIP), reduce porosity, relieve residual stresses, and improve ductility/microstructural integrity; however, they each introduce additional process parameters and trade-offs that must be balanced for performance and cost-effectiveness [3,81,95,99,100,120,129].

CFD studies are extremely beneficial and can provide solutions to ‘‘what-if’’ scenarios. ML studies also show promise but lack comprehensive validation across diverse regimes and materials [6,15,61,109,112,113]. Monitoring demonstrations frequently stop at proof-of-concept rather than industrial-scale assessments, and system-level integration studies are sparse. Standard validation procedures for both monitoring systems and models, and the use of grouped and prospective validation, are essential to build trust and support technology transfer.

4.4.4. Numerical Analysis: Physics-Based Simulations and Reduced-Order Modelling for CSAM

Numerical analysis in CSAM has two main roles. First, to explain the coupled, multi-scale physics that control acceleration, impact, and deposit build-up. Second, to provide simpler models (surrogates) that enable optimisation and, eventually, control. A core area, therefore, includes high accuracy solvers (CFD for nozzle-gas-particle dynamics; explicit FEA for impact and bonding; analytical models describing deposition conditions models, and reduced-order or data-assisted models that simplify these solvers without losing essential physics [5,6,7,12,13,21].

From a physics perspective, compressible flow through de Laval and constant-diameter nozzles sets the particle velocity and temperature, with gas type, inlet temperature/pressure, and nozzle geometry defining outlet conditions and, in turn, whether critical velocities for different materials can be reached [6,9,109]. Lagrangian tracking of particles in the jet resolves drag, heat transfer, and travel time, while particle-wall impact must capture fast plastic deformation with adiabatic heating and friction to represent bonding and flattening [12,21]. At the mesoscale, deposit evolution emerges from local kinematics (spray angle, stand-off, traverse speed) and geometric overlap between passes. Thickness and shape prediction can be obtained via analytical superposition (Gaussian profiles) or explicit height-evolution solvers [5,14,41].

CFD studies typically rely on RANS-based turbulence closures to render nozzle internal/external flows computationally feasible. When coupled with particle tracking, they expose sensitivity to expansion ratio, throat and exit diameters, and stand-off, enabling nozzle-level optimisation [7,15,110]. For impact mechanics, explicit dynamics with Johnson–Cook plasticity, penalty contact, and adiabatic heating is the common approach, resolving deceleration, jetting, and plastic flow at microsecond timescales [12,21]. Simplifications such as 2D axisymmetric domains and normal-impact assumptions reduce runtime but miss angled sprays, varied particle shapes/sizes, and curved substrates, limiting practical validity for process planning [21,109]. Molecular dynamics provides mechanistic clarity at nanoscales but does not transfer quantitatively to micron-scale CSAM without careful bridging, making it valuable for testing search algorithms and informing qualitative trends [36].

Verification, validation, and credibility checks are often incomplete across the literature. Mesh and time-step independence, contact-parameter sensitivity, and calibration to high-strain-rate data are not always reported, while validation is often limited to narrow parameter ranges or single materials [6,109]. Data-driven property models built from mixed sources can show low apparent error (e.g., RF RMSE near 0.8% for porosity), but random splits give overly optimistic results and hide domain shifts, requiring grouped and prospective validation and uncertainty calibration before use [20]. As a result, simulation trustworthiness, not just accuracy, becomes the limiting factor for applying numerical outputs in optimisation and control loops. Reduced-order modelling (ROM) and surrogates act as the computational link.

Physics-guided surrogates that include analytical structure (Gaussian mean functions with GP residuals or ANN models shaped by analytical features) regularly improve data efficiency and stability in small datasets for single and overlapping tracks [13,37]. One-shot ANNs can copy multi-layer geometries at real-time speed, but their lack of built-in physics makes them fail when spray conditions, materials or nozzles differ from the training set [24]. At the particle-impact scale, hybrid methods that train surrogates on explicit FEA plus some experiments allow millisecond-level prediction for penetration and flattening. The observed pattern (GPR better on FEM-only data, small NNs better when experiments are included) shows penetration has higher complexity than flattening and supports multi-accuracy learning instead of single-accuracy fits [21]. Expanding capacity with GA-designed deep surrogates improves flattening accuracy but stays weak for penetration unless more input data or physics-based rules are added [35]. Hybrid solver-in-the-loop frameworks maintain physical consistency, while replacing slow height-evolution solvers with deep neural network (DNNs) reach near-real-time design-space search. For example, flatness of 99.5% was achieved for a part with inward curves (concave) with only 0.45 mm excess height (overspray), showing that a physics-based model secures feasibility while a DNN surrogate provides the speed for optimisation [14]. In this context, Figure 9 shows the hybrid physics and ML pipeline (calibrate physics, train surrogate, optimise with constraints and UQ, plan and run, monitor and update).

Numerical optimisation is effective when it is tightly coupled to credible models. CFD-assisted RSM, ANN, and GA pipelines deliver multi-objective improvements by exploring nozzle design spaces that are impractical to test by trial-and-error experimentation [15]. Inverse design from measured geometry is possible with small MLP surrogates combined with PSO, reaching close to 0.06 mm median error within seconds. Still, limits in identifiability (e.g., low sensitivity to stand-off) show when inverse problems are not well-defined and when extra physics or sensors are required [22]. Optimisers like PSO, DE, and EGO are more sample-efficient than GAs and fit non-smooth objectives from contact and plasticity, but methods with uncertainty handling and constraint control are still uncommon in CSAM, even though they clearly improve safety and reliability [36,63].

Taken together, the numerical analysis subdomain yields three critical insights. First, accuracy must be matched to decision level: use high-accuracy CFD/FEA to set feasibility and calibrate physics-guided features; use ROM and surrogates for planning, inverse design, and control. Second, trustworthiness is more important than reported accuracy: mesh and time-step checks, grouped and prospective validation, and multi-fidelity training are essential before optimisation, not optional additions [20,21,109]. Third, optimisation must include uncertainty and constraints, linking cost models (e.g., N₂–He blends) and safety margins while recognising identifiability limits. This supports hybrid methods that use online monitoring to update models and correct drift during production [26,27].

5. Challenges and Opportunities

Although AI and ML are increasingly embedded in smart manufacturing system, Cold Spray Additive Manufacturing presents distinctive barriers alongside clear opportunities for progress. Based on the preceding literature review, this section synthesises the principal challenges and the corresponding opportunities for research and industrial deployment of AI and ML in CSAM, framing where current methods fall short and where targeted developments can deliver the greatest impact.

5.1. Data

5.1.1. Limited Data vs. Big Data

ML for CSAM still works mainly in a limited-data setting. Most studies train on small, single-machine datasets or collections from papers whose labels and features are inconsistent across powders, nozzles, imaging setups, and porosity definitions, giving overly low error rates under random splits and weak general validity [16,20,39]. Inverse models trained on only a few tracks can solve for parameters in seconds, but the reported accuracy often comes from local smoothness and limited data coverage. Standoff distance shows poor identifiability, highlighting how narrow design spaces hide uncertainty [22,23]. The literature-scale “mega-datasets” are quite useful. Mixed measurement protocols, uncertain inputs (e.g., gas temperature and pressure), and paper-specific context can leak into splits, making results look better unless grouped or forward-looking testing is enforced [16,20]. Therefore, the main need is not a new model type, but a new way of collecting and testing data like leave-one-paper-out, lab and material-out validation, and clear handling of uncertainty from reported ranges [20]. Hybrid simulations plus experiment surrogates partly balance small samples by starting to learn with physically reasonable trends, but they also bring in domain mismatch. Models that work well on FEM-only data lose accuracy when experimental variability appears unless training is made to handle different noise levels in simulation and real data [21,35]. Similarly, synthetic tracks from analytical or CFD solvers speed up testing but copy solver biases at non-standard angles, sharp edges, and complex shapes, limiting use for real geometries unless corrected by targeted experiments [6,14]. Active learning gives a clear solution, with the use of forward models to find where gradients change quickly or uncertainty is high, then collect detailed measurements, that match the performance drops seen at angled sprays and near edges [14,23]. On the other hand, CSAM is generating big data through multi-sensor, high-speed monitoring, thermal cameras, optical S-PTV, aeroacoustics, robot-mounted laser scans, and low-cost IIoT vision, producing frame or scan-level data fast enough for layer-by-layer quality control [17,26,38,101,112]. However, more data does not mean better data, as without consistent meta-data (sensor position, optics, heat emission handling, lighting), synchronisation with robot triggers, and labels tied to physical outcomes (e.g., porosity types, adhesion limits), large data streams mostly raise storage needs while keeping uncertainty [73,93,100]. Reports of near-perfect accuracy on restricted datasets show this issue, where very high metrics can just reflect easy test conditions, data leakage, or narrow data spread, and should be checked with forward-looking tests under change [38]. In practice, a balanced strategy emerges with the use of offline computation to calibrate CFD and FEA and build physics-guided surrogates, deploy compact predictors at the edge for millisecond inference, and drive data growth intentionally via active learning, not passively via uncontrolled logging [14,15,17,112]. In short, CSAM needs “smart data” rather than just more data, with grouped and prospective validation, physics-based feature design, and targeted data collection in hard cases. These are the key steps that can transform both small, carefully built datasets and big sensor streams into reliable ML-based analysis for design, monitoring, and optimisation.

5.1.2. Bias, Measurement Inconsistencies, and Data-Leakage Risks

The risk-of-bias check (Table 5, PROBAST) showed 7 of 19 studies at ‘High’ overall risk, with the rest ‘Unclear’; none reached ‘Low’ risk in all areas. The main causes were small or mixed datasets (especially taken from the literature), unclear predictor and outcome meanings, and validation setups that allowed data leakage (e.g., random splits that mixed samples from the same study, machine, or project across train/test). Measurement differences were especially strong for porosity, where studies combined image-analysis and density-by-displacement method, creating operator, and procedure-related effects. These factors explain why porosity models were less stable and more sensitive to dataset setup, and why results often worsened under distribution changes.

5.1.3. A Unified Protocol for Data Collection and Database Construction

Building reliable CSAM–ML datasets require a reporting protocol that captures both measurement methods and process details, to ensure that the cross-study fusion does not collapse under hidden variations. At a minimum, porosity must be tied to explicit image-analysis settings (magnification, thresholding, field-of-view, segmentation method), specimen location, and sectioning orientation, because 2D quantification and micro-CT give different sensitivities and operator-based bias that otherwise spread into labels [2,73,93,95]. BSD should record step size, pattern quality metrics, and surface preparation, alongside heat-treatment history, given its central role in assessing recrystallisation and inter-particle bonding. Moreover, tensile data should declare specimen geometry, extraction direction and gauge position relative to build axes to make anisotropy comparisons clear [84,91,93]. Residual-stress reports must state the method, calibration, and depth sensitivity, because near-surface XRD and bulk neutron diffraction imply different ranges of interpretation, and hole-drilling requires standard calibration.

Equally, the process side must be standardised as well. Nozzle geometry needs a controlled vocabulary and numeric descriptors type (de Laval or constant diameter), throat and exit diameters, divergence angle, and expansion ratio, paired with inlet gas type, temperature and pressure, since these jointly control particle acceleration [6,7,9]. Trajectory parameters should include spray angle, standoff distance, traverse speed, path topology (e.g., crosshatch, MK), layer count and pause time between passes because geometry and porosity depend on jet footprint overlap and local motion [5,73]. For in situ sensing, the protocol should log device model, optics, frame rate/sampling, calibration, and region-of-interest for thermal cameras; measurement volume and optical alignment for S-PTV; sensor type and placement for aeroacoustics; and scan spacing and filtering for laser profilometry, enabling repeatable monitoring and control studies [26,100,101].

Because literature-mined datasets mix laboratories and protocols, meta-data must also carry accuracy tags (simulation versus experiment; solver type; mesh/time-step checks), reported input uncertainty, and grouping keys (paper, rig, material) to ensure that grouped and forward-looking validation can be enforced and uncertainty can be handled rather than treated as exact [20,21,35]. Finally, hybrid CSAM datasets should apply multi-accuracy links between FEM, analytical surrogates, and experiment to support transfer learning without bias transfer, reflecting the gaps between simulation and experimental work, highlighted across penetration, flattening, and geometry prediction.

5.1.4. Data Pre-Processing and Feature Extraction

Robust CSAM pipelines begin with careful synchronisation and cleaning of time-series data. Thermal, aeroacoustics, and profilometry streams should be time-aligned using robot triggers and controller events, with sliding-window splitting to separate slow changes (e.g., nozzle wear) from sudden problems (e.g., delamination) [38]. Thermal channels require emissivity management and vigilance for lens contamination; both factors otherwise misrepresent temperature trends used for forecasting [19,100]. For optical feeds, illumination is part of preprocessing, as side and ring lighting choices used for different defects imply corresponding normalisation steps to stabilise downstream inference.

Geometry should be encoded with physically useful summaries rather than raw height maps. Fitted-Gaussian track values, overlap ratios and per-pass height increments provide compact, robust descriptors of layer build-up [5]. Analytical gas-jet metrics (maximum ideal particle velocity (γ), acceleration capacity (β), and free-jet deceleration (α)) are effective features for mapping recipes to outcomes and should be computed alongside nozzle and flow descriptors (exit diameters, divergence, expansion ratio, pressure, and temperature) [6,7,9]. In-flight particle statistics from S-PTV, including velocity and size distributions, add important variables for deposition efficiency [26,101].

For microstructures, computer-vision descriptors and similarity metrics such as Structural Similarity Index Measure (SSIM) can stabilise porosity and defect labelling but remain sensitive to operator-based preparation and class imbalance for rare defects, reinforcing the need for careful data selection [96,97]. Physics-informed representations (analytical means with Gaussian-process residuals or ANN models combined with analytical features) regularly improve data efficiency and extrapolation within small, targeted datasets [13,37]. Finally, the literature-mined datasets require checks for missing data filling, tests of feature stability under grouped (per-paper and lab) validation, and explicit tracking of reported input uncertainty to avoid errors from mixed protocols [20]. Where deployment at the tool-centre point is intended, feature sets should also be checked for real-time limits shown in recent vision pipelines [112].

5.1.5. Cloud-Edge Synergy

Edge inference is feasible for CSAM with efficient CNNs, while low-cost IIoT stacks already stream aligned thermal and optical data and process signals for layer-wise quality checks [103,108]. Robot-mounted profilometry closes the loop by enabling pre-layer scans and real-time path changes, whereas embedding fast surrogates at the edge can turn these checks into predictive adjustments under strict delay limits [13,14,17]. Heavy CFD/FEA can be stored in the cloud or offline for calibration and controller updates [7,15].

5.1.6. Transfer Learning

Transfer learning in CSAM remains brittle as models trained on mixed datasets often lose accuracy under leave-one-paper-out, leave-one-material or leave-one-nozzle evaluation because protocol, sensor, and material changes dominate the main signal [16,20,39]. A practical route to better transferability is multi-accuracy learning that treats FEM and experimentation as different fidelities and embeds hierarchical and random-effects terms to absorb lot-, batch- and rig-level variation; thereby, separating systematic changes from learnable patterns. Explicit conditioning on oxide state, particle size, morphology, impact angle, and material pair should accompany any cross-material claims to reduce mismatch at deployment [21,35]. Evaluation must reflect transfer of grouped and forward-looking splits, plus propagation of reported input uncertainties are needed to avoid data leakage and overstated generalisation [20]. In practice, pre-training on physics-based simulations with selective fine-tuning on targeted experiments is effective when enough input features are available. Notably, penetration requires richer descriptors than flattening, consistent with observed error patterns across fidelities. Operationally, studies should report transport tests across powders, substrates, and nozzles with calibrated uncertainty. When transfer fails, active data collection should target under-sampled cases identified by physics and model error maps.

5.2. ML Models

5.2.1. Interpretability

Interpretability in ML for CSAM is a key requirement for root-cause analysis and operator trust, given the documented “black-box” problem in AM monitoring [96]. For porosity prediction in high entropy alloy Cantor coatings, SHAP analysis showed that gas type and powder type were consistently the top contributors across the best models (support vector machine regression with a linear kernel and linear regression). Performance dropped when those categorical factors were removed, showing that the attribution matches process physics around particle velocity and feed characteristics [39]. For real-time flow/defect monitoring, a stethoscope-guided pipeline (CAE to CNN) and Integrated Gradients were used to visualise which parts of the sound spectrogram influenced decisions. The maps highlighted a 600–1600 Hz range for the “no-powder-flow” problem, and the classifier reached ~95% test accuracy with the stethoscope input, showing that explanations can be linked to physical sound sources (feeder/nozzle states) instead of being just generic heatmaps [34]. Crucially, attribution reliability must be established under grouped and prospective validation to avoid paper-specific errors and data leakage common to literature-mined datasets [16,20]. A porosity dataset was collected from earlier studies and, after careful cleaning (fixing missing values, converting categories, and selecting features), showed only 0–2% average error on a test set. However, it also showed problems with variety and uncertainty because the data came from very different experiments and measurement methods, with missing information and even conflicting parameter trends across sources [20]. In this case, testing by leaving out one study, machine, or powder type, and running future trials on new nozzles or powders, is important so that SHAP/IG explanations work on new CSAM systems instead of just remembering lab-specific details.

Physics-based checks should clearly link explanations to factors like critical velocity ranges, oxide-film behaviour, and overlapping jet footprints, instead of assuming highlight maps are proof. Reviews of CSAM show that successful deposition depends on keeping particle velocity within a certain range, i.e., if the velocity is too low, particles rebound, and if it is too high, then they cause erosion. Oxide films also crack, break, or peel during impact, which affects bonding and porosity. These effects are strongly influenced by the choice of gas (He vs. N₂), gas temperature/pressure, and the oxygen level and sphericity of the powder [2,44]. Model forms that show internal structure can further enhance interpretability. For track-profile/footprint modelling, a surface-aware Gaussian Process Regression was used with an explicit Gaussian mean function (domain knowledge) plus learned residuals. This physics-informed approach means improved predictions over simple Gaussian mixing and fully data-driven models, directly linking model components to geometric overlap mechanisms important for toolpath tuning [13]. Finally, explanations should be paired with calibrated uncertainty, to ensure that operators can weigh predicted benefits against measurement noise and protocol differences before triggering corrective actions. GP-based models naturally provide posterior variance over predicted profiles, offering a transparent way to threshold alarms or recommend parameter nudges when attribution points to gas/powder-driven velocity deficits or overlaps are detected.

5.2.2. Overfitting vs. Underfitting

Random splits artificially raise accuracy, with grouped and prospective tests better estimating generalisation and usually reporting lower R² and increased RMSE [16,20]. Underfitting appears in linear baselines for mixed datasets, whereas overfitting emerges in complex models trained on small, single-rig data or when evaluation leaks context [16,23,24]. Model choice should be justified against input adequacy. Flattening behaves lower-dimensional than penetration, explaining why simpler models can be enough for the former while deeper networks struggle on the latter without richer features [21,35].

5.2.3. Physics-Based vs. Data-Driven ML Modelling

CFD and analytical models provide feasibility checks, sensitivity, and nozzle-design guidance; explicit FEA captures impact physics; and hybrid surrogates combine stability from physics with millisecond prediction speed [6,7,13,14,15,21]. In CSAM geometry prediction, models trained only on small, single-setup datasets often underperform. Adding domain-informed surrogates (e.g., a Gaussian profile used as an extra input) and virtual data augmentation improves data efficiency and accuracy compared with a purely data-driven ANN, within the tested process range [24,37]. The opportunity is multi-fidelity training with feasibility constraints and uncertainty, aligning inverse searches with parameter identifiability and safety limits [22,35].

5.3. Multiscale Spatiotemporal Modelling and Control

Thermal, S-PTV, aeroacoustics and laser-profilometry signals evolve on different timescales and must be combined with models that link particle-scale impact to track/layer build-up and component-level geometry and porosity [5,18,100,107]. Time-series forecasting on thermal streams detects feed-rate changes within fractions of a layer, showing useful horizons for corrective control, while aeroacoustics can flag nozzle wear and delamination as global signals [18,19]. Closed-loop demonstrations with robot-mounted profilometry confirm that pre-layer scans and online path edits reduce overspray and under-build, but formal delay and stability margins are rarely reported [17,25].

5.3.1. Sensing System

In CSAM, sensing must handle supersonic particle jets, changing surface states, and the need for fast-response control methods while still providing signals that map onto usable controls. Thermal cameras and pyrometers remain the most practical tools for tracking heat build-up and detecting process drift, but their accuracy depends on emissivity correction as surface roughness and oxidation change during a build, with calibration challenges increasing over complex geometries [100]. As a result, thermal data are most reliable when paired with regular calibration routines and used together with other methods that are less sensitive to radiative properties [25,100].

Optical sizing–particle tracking velocimetry (S-PTV), shown by the HiWatch HR platform, directly measures in-flight particle velocities, trajectories and size distributions, enabling 2D velocity scan maps for jet tuning and recipe checking. However, the usable measurement volume is small, and window contamination, particle blocking, and line-of-sight limits coverage near complex geometries [26,101]. These data are nevertheless very useful for aligning nozzle design and operating conditions with target particle-state distributions [6,9].

Aeroacoustics has emerged as a strong, non-contact overall diagnostic. where frequency-time signatures respond to gas-supply changes, powder feed variations, standoff deviations, nozzle wear and even delamination, with low sensitivity to precise microphone placement reducing setup effort [18,102]. Performance drops at very low powder loadings and when wear is highly uneven, requiring extra sensing or sensor fusion to separate causes [18]. Mechanically coupled audio sensing at the powder feeder, captured via a stethoscope-style mount, greatly improves signal-to-noise ratio and enables reliable anomaly classification compared with free-field microphones, showing the benefit of high-SNR channels [34].

Geometric sensing through laser profilometry closes the loop on what ultimately matters, i.e., deposited shape. Robot-mounted scanners measure track-shape changes at the tool-centre point, with skewed-Gaussian fits giving accurate cross-sections and enabling pre-layer scans and real-time trajectory updates to reduce under-build and heat-induced distortion [17,107]. At system level, low-cost IIoT platforms now combine optical/thermal imagery with gas temperature, pressure, and robot triggers, delivering layer-wise quality assurance in near real time. Downstream CNNs can achieve very high F1 scores internally, though forward-looking validation under drift and lighting changes are essential [25,38]. The immediate opportunity is systematic multi-sensor fusion with calibrated uncertainty, time-stamped synchronisation, and lifecycle-aware indicators that signal nozzle erosion or powder-lot changes early to ensure that the models and paths can be adapted before defects build up.

5.3.2. Process–Structure–Properties–Performance (PSPP) Relationships

Credible PSPP chains for CSAM must begin from parameters that set gas-particle states and jet footprints, continue through inter-particle bonding and porosity formation under severe plastic deformation, and end in anisotropic mechanical responses shaped by heat treatment [2,5,73,82,83,84,89,90,95]. ML predictors of porosity and geometry can fit into this chain if labels are standardised and uncertainty-based, and if grouped validation confirms transfer across materials and rigs [13,14,20].

5.4. ML/AI for CSAM Sustainability

Sustainability in CSAM is governed by three tightly connected factors: process energy (mainly gas heating and delivery), consumables and wear (especially nozzles), and yield losses from geometric error and porosity. ML and AI contribute to all three by embedding physics-based predictions into planning, monitoring, and control. On gas usage, pre-run “recipe screeners” for porosity and quality reduce the space for risky parameter sets, cutting aborted builds and unnecessary helium use. Decision trees (SVMs and ensembles trained on the literature and lab data are already millisecond-fast) require grouped and prospective validation to avoid overly positive errors from context leakage [16,20,39]. Cost-aware N₂–He mixture models further separate performance from helium price changes and allow economical settings without lowering deposition quality [27]. At the same time, process ranges that reach target particle velocities at moderate gas temperatures can be identified via CFD-assisted optimisation, reducing heating demand at the source [5,7,15].

Yield improvement depends on preventing rework and scrap through predictive geometry and porosity control. Hybrid physics-guided surrogates (analytical and CFD means corrected by learned residuals) allow fast track and layer prediction and path adjustment, limiting overspray and edge swelling, and thereby reducing material waste and extra machining [5,13,14]. Robot-mounted laser profilometry has shown practical, layer-by-layer checks and on-the-fly trajectory edits to stop under-build or heat build-up, directly lowering the chance of part rejection [17]. Pre-spray verification with S-PTV further ensures jet quality before committing material, helping avoid low-efficiency runs [26].

Consumable life and unplanned downtime are addressed by intelligent monitoring. Aeroacoustics diagnostics can detect gas-supply drift, powder-feed irregularities and nozzle wear as overall signals, enabling maintenance before major failure; mechanically coupled audio sensing at the feeder further improves signal-to-noise for fault detection [18,34,102]. Low-cost, open IIoT stacks integrate thermal/optical streams with process signals and achieve high internal accuracy for layer-wise QA, lowering adoption barriers for SMEs, provided forward-looking validation is performed to manage drift [38]. Edge deployment is feasible, with transformer-based video super-resolution and efficient CNNs meeting sub-150 ms inference, aligning with robot motion and allowing in-cycle decisions that cut waste [112].

While the original identified literature does not directly use the term “sustainability”, it discusses the challenges and opportunities of CSAM in terms of cost efficiency, waste reduction, and industrial viability. The three-lever framework (energy, consumables, yield) is introduced here as a summary to join these scattered themes into a clear sustainability perspective, aligning with the practical concerns highlighted across RD1–RD4. Overall, sustainability gains in CSAM arise from AI/ML-enabled improvements such as physics-guided surrogates and recipe screeners that lower helium and heating demand (energy), aeroacoustics-driven diagnostics and predictive maintenance that mitigate nozzle wear (consumables), and real-time sensing-control (e.g., robot-mounted profilometry) plus pre-run porosity prediction that reduce rework and scrap (yield). These benefits are delivered through physics-based predictions (ML) combined with optimisation and closed-loop decision-making (AI) in cloud–edge architectures.

6. Answers to the Research Questions

This investigation was structured around three key research questions addressing critical implications of AI/ML on CSAM. In terms of CSAM problems currently being addressed with AI/ML (RQ1), four main research domains have been identified. In design for CSAM, predictive ML models estimate single-track and multi-layer geometries, help with edge-compensation, and work backwards from a desired shape to process parameters, while path planning (algorithmic or ML-assisted) accounts for the CSAM spray footprint and access constraints. In material analytics, models relate process and feedstock descriptors to porosity, adhesion, and mechanical properties. Image-based ML aids microstructure and defect identification, and hybrid models trained on both simulations and experiments estimate particle-scale outcomes such as flattening and penetration. In in situ monitoring and defect analytics, thermal data, aeroacoustics, robot-mounted laser profilometry, and camera-based vision are used to spot drift and defects, predict process states, and, in first demonstrations, apply layer-wise trajectory corrections. Finally, in deposition modelling and optimisation, ML supports CFD and analytical frameworks to speed up nozzle and flow parameter optimisation, guide cost-effective gas blending, and connect modelling to planning within solver-in-the-loop architectures.

In terms of the AI/ML model types being used to solve CSAM problems (RQ2), they span traditional algorithms and modern neural approaches, often combined with physics guidance. Random forests, gradient-boosted trees, decision trees, linear models, and SVMs remain competitive when datasets are small or mixed. Gaussian process regression is widely used where uncertainty estimates are needed and where physics-informed mean functions can build in analytical relations, particularly for geometry prediction with overlapping tracks. Neural models include compact feed-forward networks for forward and inverse tasks, autoencoders for anomaly detection, CNN pipelines for layer quality, time-series forecasters for thermal streams, and specialised architectures (e.g., wavelet or multi-branch networks) for particle-impact surrogates. These learning components are frequently used together with RSM and Central Composite Design (CCD) or Latin Hypercube Sampling LHS-driven designs of experiments, evolutionary or swarm optimisers, and Bayesian or evolutionary search for balancing multiple objectives. A notable trend is hybrid methods, where analytical or CFD features provide stability and interpretability while learned residuals or surrogates deliver speed, and multi-fidelity training explicitly blends FEM outputs with experimental data to reduce simulation-to-reality gaps.

Across the literature, three themes summarise benefits, limitations, and open challenges when applying AI/ML to CSAM (RQ3). On the benefits side, very fast predictors for porosity and geometry shorten design cycles and enable pre-run recipe screening; hybrid CFD-plus-ML frameworks deliver simultaneous improvements in particle velocity, heat flux, and surface pressure for redesigned nozzles; and practical monitoring pipelines based on thermal, audio, vision, and profilometry data show high accuracy internally and, in some cases, achieve layer-wise corrections that reduce overspray and under-build. On the limitations side, most studies still operate in a data-scarce setting with single-rig or literature-mined datasets that mix protocols and labels; random-split validation often overstates accuracy through context leakage; domain shift between simulation and experiment, as well as across materials, powders, nozzles, and geometries, leads to poor transfer; uncertainty estimates are often missing; and critical variables such as standoff may be hard to identify from geometry alone. Calibration and integration issues persist, especially changing emissivity in thermal sensing, window fouling and small measurement volumes in optics, and stand-alone single-sensor deployments that are not integrated into latency-aware, stability-audited control loops. As a result, open challenges focus on data standards and grouped and prospective validation, multi-fidelity and domain-adaptive training with calibrated uncertainty, identifiability-aware inverse design, reliable multi-sensor fusion at the edge, and proven closed-loop control that reports allowable delays and recovery margins for representative geometries. This description addresses the research questions mentioned in Section 2.1.

It is to be noted that although this systematic review focuses on CSAM, the main machine learning challenges identified here, such as small and mixed datasets, simulation-to-reality gaps, fragmented sensing, and the need for robust, deployment-ready validation, also appear across other AM categories as well. According to ISO/ASTM 52900:2021, AM encompasses seven categories: binder jetting, directed energy deposition, material extrusion, material jetting, powder bed fusion, sheet lamination, and vat photopolymerization [1]. Each category faces similar data scarcity issues due to the high cost of experimentation involving trial-and-error, proprietary restrictions on sharing industrial data, and the complexity of multi-physics numerical simulations governing process behaviour as well as part quality. For instance, powder bed fusion processes like selective laser melting struggle with thermal monitoring and porosity prediction [130], while material extrusion (e.g., fused filament fabrication) encounters challenges in predicting mechanical anisotropy and surface quality [131]. The commonality of these AI/ML challenges across AM categories have been widely reported [132,133], and suggests that methodological advances developed for CSAM (specifically in handling limited data, physics-informed modelling, and uncertainty quantification) could transfer to other processes, accelerating the broader industrial adoption of data-driven AM process control.

Based on the content of the systematic literature review, a quantitative meta-analysis was not performed for this work. It is due to the fact that the outcomes, scales, and validation setups of the identified articles were widely different, with many missing uncertainty ranges. Instead, a structured, grouped summary is provided with certainty marked accordingly (‘Moderate’ for geometry/monitoring; ‘Low’ for porosity and inverse design). These judgments arise from the bias sources (Section 2.8) above and should be kept in mind when interpreting main results.

7. Future Work

A priority is to standardise CSAM–ML data and reporting to ensure that models from different labs can generalise without mistakes caused by unseen problems in the training data. Use a simple but complete framework that records process details (gas type, pressure, temperature; nozzle type and geometry; path settings such as angle, standoff, traverse speed, path pattern, and pauses), sensor setup, calibration (how emissivity is handled for thermal cameras, optics used, where microphones are placed, profilometer scan spacing and filters, S-PTV measurement volumes), material information (particle shape and size distribution, oxide state, substrate pairing), and characterisation settings (porosity method with image settings and where the section was taken, EBSD step size and pattern quality, residual-stress method and depth, tensile specimen geometry and extraction direction). Always include grouping keys (paper, experimental setup, material, nozzle) and report input uncertainties to ensure that testing on unseen groups, checking performance on future data, and carrying measurement uncertainties through the model become the normal practice.

With a clean and consistent dataset in place, modelling should move toward physics-based, multi-fidelity surrogates with well-calibrated uncertainty. Analytical and CFD-based features (such as ideal particle velocity and acceleration capacity, free-jet deceleration metrics, and Gaussian footprint parameters) can be defined as the base trend, while Gaussian processes capture remaining patterns and yield credible intervals, which can map to geometry and porosity thresholds. At the particle-impact scale, training should explicitly treat FEM and experimental data as different noise levels and expand inputs beyond velocity and yield strength to include impact angle, morphology, oxide condition, and thermal state. Model performance must be reported alongside calibration quality because coverage of prediction intervals at decision-relevant thresholds is more informative for deployment than just RMSE.

Considering current models struggle at angled sprays, near edges and in concavities, an active learning targeted at these “hard zones” is essential. Uncertainty-driven sampling and gradient-change criteria can identify high-value points, and fast characterisation loops combining profilometry with targeted micro-CT and minimal sample preparation can then supply better labels. This strategy converts data collection from passive logging to purposeful sampling of the conditions that affect accuracy, inverse identifiability, and closed-loop stability.

Inverse design should move from single “best” recipes to sets of parameters with confidence ranges, with explicit feasibility and safety constraints. Multi-objective formulations can balance flatness and edge accuracy against porosity risk and helium cost, while physics-based priors from critical-velocity limits and nozzle-gas-particle coupling narrow the search. Modern optimisers should operate on possible surrogate outputs rather than point predictions, returning parameter ranges with confidence. When a target geometry is not enough to determine weakly sensitive variables such as standoff, the algorithm should flag that non-identifiability and recommend extra sensing before committing to a build.

To translate monitoring into practical manufacturing value, a fused, latency-bounded closed-loop controller should be built and evaluated. A minimal, robust stack would combine one thermal camera for heat-buildup trends, one aeroacoustics channel for global events such as gas-supply drift, powder-feed anomalies, or nozzle wear, and a robot-mounted laser profilometer to measure track cross-sections and layer height increase at the tool centre point. Lightweight classifiers and forecasters can run at the edge, fitting skewed-Gaussian tracks online and forecasting thermal drift. A state estimator can merge the streams, and a model-predictive controller can adjust traverse speed, dwell, local angle, or pauses. The study should report total loop delay, stability margins, false-stop rates, and recovery effectiveness on test samples that include flat walls, edge-rich patterns, and concavities, providing evidence ready for certification and transfer.

These elements can be integrated within a CSAM digital twin that links offline physics calibration with online surrogate updates and drift detection. Before a run, the twin checks process parameters with uncertainty, simulates path plans with edge compensation, and selects economical gas blends subject to performance constraints and helium price. This framework enables clear sustainability metrics. For example, energy and gas per usable volume, helium saved relative to a baseline, scrap and rework avoided by online corrections, and predicted nozzle life from audio signals can all be tracked. A before-and-after study on matched parts would supply the ROI evidence (cost per kilogram saved and, where possible, CO₂-equivalent reductions) that industry stakeholders expect.

Finally, to generalise beyond one lab or material, use deliberate transfer-learning and strong benchmarks. Mixed-effects neural models can distinguish experimental setup, powder, and nozzle variability; domain-adaptation layers can bridge simulation to experiment gaps; and success should be measured under testing models on materials or nozzles that were not included in training, to ascertain if they can generalise to new cases while maintaining uncertainty calibration. Releasing standardised datasets with grouped splits, reference fusion pipelines, sensor and geometry benchmarks, and clearly documented tests that show the effect of removing or changing features or model parts will make results reproducible and comparable. Together, the strategy above advances the four research directions identified in this review (hybrid surrogates, targeted data collection, uncertainty-based inversion, and closed-loop control) from early-stage prototypes to reliable, verifiable solutions for industrial CSAM.

8. Conclusions

Integrating ML into Cold Spray Additive Manufacturing has demonstrated strong potential to accelerate the transition from trial-and-error methods toward data-driven, physics-based production. This review was motivated by the rapid growth of CSAM research and the increasing use of AI/ML tools to address its long-standing challenges. A systematic organisation of recent literature highlighted four core domains of progress, i.e., design and inverse planning (RD1), materials analytics and characterisation (RD2), in situ monitoring and defect analytics (RD3), and deposition modelling with AI-enabled optimisation (RD4). Together, these domains outline how ML can address CSAM’s process–structure–property linkages and provide a foundation for scalable industrial adoption.

Within these domains, physics-guided optimisation reliably improves jet behaviour and deposit quality. Studies that combine CFD of nozzle–gas–particle flow with data-driven optimisers (RSM, ANN, and GA) report simultaneous gains in particle velocity, particle temperature, heat flux, and surface pressure for redesigned nozzles, performing better than trial-and-error methods and giving practical guidance on enter and exit sizing and expansion control. At the geometry level, physics-informed surrogates, Gaussian and analytical models with GP and ANN solvers, solver-in-the-loop networks, enable very fast (millisecond-scale) track and layer prediction and support inverse design and path correction while keeping results feasible. One-shot ANNs show excellent accuracy within trained ranges and analytical screening metrics (γ, β, α) allow quick comparison of operating ranges. The clear trend is that adding physics constraints to guide search and using surrogates for speed gives more accurate, lower-porosity, better-quality builds than unguided tuning.

Despite these advances, CSAM still lacks a fully integrated, fast closed-loop control system that combines thermal, vision, aeroacoustics, and profilometry signals as well as directly controls robot motion in real time to maintain geometry and porosity under changing conditions. The literature shows strong individual methods, thermal cameras for temperature trends, S-PTV for particle mapping, aeroacoustics for nozzle wear and delamination and supply drift, and laser profilometry for layer-wise detection and correction (skewed-Gaussian track models; robot-mounted scans that pause and adjust paths). There are also low-cost IIoT systems with high offline accuracy and anomaly detectors using audio and thermal data. However, these systems are working separate at different sampling rates, sensitive to calibration (e.g., emissivity, optics fouling), and without a single controller that ensures stable delay from measurement to decision to action, or reports on margins and recovery near edges, concave areas, or under drift (nozzle wear, gas changes). This gap is most serious where CSAM physics is extremely complex (near edges, concave regions, and during orientation changes), where open-loop planning and layer-by-layer checks are too slow to prevent build errors and porosity problems.

As a result of these gaps, model-based design and optimisation (RD1 and RD4) are now advanced enough to set practical working ranges, critical velocity ranges, nozzle control, and toolpath strategies such as Metal Knitting and edge compensation and even to allow near real-time planning. However, without fast signal fusion and feedback from different types of sensing (RD3) and robust materials analytics and characterisation (RD2; to track powder and oxide-state changes, and standardise porosity labels), small overlaps between process settings and minor changes in the environment or material can cause an evidently correct recipe to be off target during production. In practice, this shows up as weak identifiability of standoff distance from shape or geometry data; bias in prediction models caused by using overly simple jet-flow formulas such as cylindrical or Gaussian models; and consistent errors at high spray angles, near edges and in concavities where data are limited and jet separation occurs.

Accordingly, by grounding future research in these practices and by directly linking design, materials characterisation, monitoring, control, and modelling optimisation in physics-based ML systems, CSAM can progress from small-scale demonstrations to repeatable, uncertainty-bounded, industrial-level production. In short, integrating physics-based models with fused sensing and timely control is the decisive step from early-stage prototypes to dependable CSAM production. This timely systematic literature review provides a critical foundation and a clear pathway for researchers and engineers to advance CSAM from the laboratory to the manufacturing industry.