Recent Advances in Artificial Intelligence and Machine Learning for Life Cycle-Wide Additive Manufacturing: A Comprehensive Review

Abstract

Additive manufacturing (AM) has emerged as a transformative technology across multiple industries, from aerospace to biomedical applications. The integration of artificial intelligence (AI) and machine learning (ML) into AM processes represents a paradigm shift toward intelligent, autonomous manufacturing systems. This comprehensive review synthesizes recent advances in AI/ML applications across the entire AM life cycle—from design optimization and process planning through in situ monitoring, closed-loop control, and post-process qualification. The analysis is organized by ISO/ASTM AM process families, including powder bed fusion (PBF), directed energy deposition (DED), material extrusion (MEX), vat photopolymerization (VP), binder jetting (BJ), material jetting (MJT), and sheet lamination (SL). For each process family, the review examines the specific AI/ML techniques employed, the data modalities utilized (thermal imaging, acoustic signals, in situ cameras, CT/NDE data), and the current state of deployment from research prototypes to industrial implementation. The analysis reveals that while significant progress has been made in single-stage ML applications such as defect detection and parameter optimization, truly integrated life cycle-wide AI-driven AM workflows remain largely aspirational. Key challenges are identified including data scarcity, model generalization across machines and materials, real-time control constraints, and certification requirements. Finally, future research directions are outlined toward autonomous AM systems enabled by physics-informed ML, digital twins, and hierarchical AI architectures.

1. Introduction

Additive manufacturing (AM), commonly known as 3D printing, has transitioned from rapid prototyping to a viable production technology for end-use components across aerospace, automotive, biomedical, and consumer goods industries [1,2,3,4,5]. Unlike subtractive manufacturing, AM builds parts layer-by-layer directly from digital models, enabling unprecedented design freedom, reduced material waste, and on-demand manufacturing capabilities [6,7,8]. However, the inherent complexity of AM processes—involving multi-physics phenomena spanning thermal, mechanical, and metallurgical domains—poses significant challenges for process control, quality assurance, and qualification [9,10,11,12].

The convergence of AM with artificial intelligence (AI) and machine learning (ML) offers transformative opportunities to address these challenges [13,14,15]. The AM process chain generates vast quantities of multi-modal data from in situ sensors, process parameters, and post-build characterization, making it ideally suited for data-driven approaches [16,17,18,19]. Recent advances in deep learning, physics-informed neural networks, and reinforcement learning have enabled breakthrough capabilities in design optimization, real-time defect detection, process parameter tuning, and property prediction [20,21,22,23].

This review provides a comprehensive synthesis of AI/ML applications across the entire AM life cycle, organized by ISO/ASTM AM process families [24]. The objectives are threefold: (1) mapping the current state-of-the-art for each AM process family; (2) identifying cross-cutting trends, challenges, and research gaps; and (3) outlining future directions toward truly autonomous, intelligent manufacturing systems [25,26].

Scope and Organization

This review is structured around two orthogonal dimensions: the AM life cycle stages and the ISO/ASTM process families (Figure 1).

In response to the second-round review, the manuscript structure is strengthened to keep the primary emphasis on ML algorithm families and their practical AM usage. Section 2 is used as the central analytical axis, while process-family sections are treated as application contexts for evaluating algorithm performance, transferability, and deployment constraints.

The life cycle stages considered span the complete manufacturing workflow. Design and DfAM encompasses ML-assisted topology optimization, generative design, and printability prediction. Process planning and optimization addresses parameter selection, path planning, and build orientation. In situ monitoring and anomaly detection covers real-time sensor fusion and defect classification. Closed-loop control includes adaptive parameter adjustment and RL-based controllers. Post-process qualification addresses property prediction, microstructure analysis, and certification support [2,25].

The ISO/ASTM AM process families provide the organizational framework for the technology-specific analysis [24]. These include powder bed fusion (PBF) with its variants LPBF/SLM and electron beam melting (EBM); directed energy deposition (DED) including LENS and WAAM; material extrusion (MEX) dominated by fused deposition modeling/fused filament fabrication (FDM/FFF); vat photopolymerization (VP) encompassing SLA, DLP, and CLIP; binder jetting (BJ); material jetting (MJT) including PolyJet and inkjet technologies; and sheet lamination (SL) including LOM and UAM [1,27].

Table 1. ISO/ASTM standards referenced in this review and their role in manuscript organization.

Standard	Scope	Use in This Manuscript
ISO/ASTM 52900:2021; Additive manufacturing—General principles—Fundamentals and vocabulary [24]	Terminology and process-family definitions	Defines the seven AM process families used for section-level organization and terminology normalization.
Process-family taxonomy derived from ISO/ASTM 52900:2021 [24]	Classification framework	Supports cross-process comparison of AI/ML maturity, data modalities, and life cycle deployment status.
Journal formatting requirements	Tables, figures, references style	Guides formatting updates for captions, citations, and nomenclature presentation in the revised draft.

outlines the primary standards and their organizational roles in this comprehensive review [24].

Recent studies in quality assurance, real-time anomaly detection, process-scale prediction, and emerging AI paradigms indicate continued expansion of AI/ML research across AM life cycle stages. Representative contributions include ML-based quality-assurance frameworks for AM processes [28], real-time anomaly detection in EBM environments [29], melt-pool characteristic prediction in metal AM via dedicated learning architectures [30], and parameter-to-property mapping in mechanics-focused AM studies [31]. Related work also spans large-scale 3D printing tool ecosystems [32], cross-domain AI trends in AM [33], biomedical metal AM with AI/nanomaterial integration [34], and generative-AI-enabled digital fabrication directions [35].

2. Overview of AI/ML Methods in Additive Manufacturing

Before examining process-specific applications, this section provides an overview of the primary AI/ML methodologies employed in AM research and their typical roles across the life cycle. The selection of appropriate ML algorithms depends on the nature of available data, the specific AM application, and computational constraints [1,13,27]. Figure 2 presents a generalized framework showing how ML integrates across different stages of AM applications.

To improve clarity and consistency, each algorithm-focused subsection below explicitly discusses representative AM use cases, key advantages, and practical limitations/barriers to industrial adoption.

2.1. Classical Machine Learning Approaches

Classical supervised learning methods remain foundational in AM applications, particularly for tasks involving structured, tabular data from process parameters and extracted features [1,27,36]. Support vector machines, random forests, gradient boosting methods, and Gaussian process regression have been extensively employed for modeling process-property relationships in both metal and polymer AM systems. Among these approaches, Gaussian process regression has garnered particular attention due to its inherent uncertainty quantification capabilities, which enable active learning strategies and sequential experimental design that can substantially reduce the experimental burden associated with process development [9,10].

From an algorithm-centric perspective, the principal advantages of classical ML are low data requirements, interpretability, and strong performance on engineered features, whereas common limitations include reduced robustness for raw high-dimensional sensing data, sensitivity to feature quality, and weaker transferability across machines/materials without recalibration.

The development of surrogate models represents another significant application of classical ML in AM. Neural networks and kriging-based models serve as computationally efficient alternatives to high-fidelity finite element simulations, enabling rapid exploration of process parameter spaces that would be prohibitively expensive with physics-based modeling alone [2,37,38]. These surrogate approaches have proven particularly valuable when integrated into optimization loops, where thousands of model evaluations may be required to identify optimal process conditions. The trade-off between model fidelity and computational cost has driven the development of multi-fidelity frameworks that strategically combine expensive experimental data with cheaper simulation results [10,36].

Table 2. Comparison of classical ML algorithms for AM applications.

Algorithm	Data Requirements	Typical AM Applications	Advantages	Limitations
SVM	Small–medium datasets; works with features	Defect classification; process window identification	Effective in high dimensions; robust to overfitting	Sensitive to hyperparameters; limited scalability
Random Forest	Medium–large; handles missing data	Property prediction; feature importance analysis	Interpretable; handles nonlinearity; ensemble robustness	Memory intensive; slow for real-time
Gradient Boosting (XGBoost)	Medium–large; structured data	Process optimization; quality prediction	High accuracy; built-in regularization	Prone to overfitting; sensitive to noise
Gaussian Process	Small datasets preferred	Surrogate modeling; uncertainty quantification; active learning	Provides uncertainty; few hyperparameters	Scales poorly $O\left(n^3\right)$; assumes smooth functions
k-NN	Any size; distance-based	Anomaly detection; similarity search	Simple; no training; adaptable	Slow at inference; sensitive to dimensionality
ANN (Shallow)	Medium–large; needs feature engineering	Process–property mapping; regression	Universal approximator; flexible architecture	Black-box; requires hyperparameter tuning

provides a comprehensive comparison of classical ML algorithms commonly applied in AM, including their characteristics, typical applications, advantages, and limitations.

2.2. Deep Learning Architectures

Deep learning has fundamentally transformed AM monitoring and design optimization by enabling direct learning from high-dimensional sensor data without the need for manual feature engineering [20,21,39]. The ability of deep neural networks to automatically extract hierarchical representations from raw images, time-series signals, and 3D geometric data has opened new possibilities for real-time process monitoring and quality assurance that were previously impractical with traditional ML approaches. Figure 3 illustrates the key application areas of deep learning across the AM workflow.

Deep-learning methods offer clear advantages in representation learning and detection accuracy for image-rich and multi-sensor monitoring tasks; however, major limitations remain in data hunger, computational cost, explainability, and cross-platform generalization under domain shift.

Convolutional neural networks (CNNs) have emerged as the dominant architecture for image-based defect detection in both metal and polymer AM [40,41,42]. Figure 4 summarizes representative generative-design and generative-model architectures (GAN, DCGAN, VAE) that are increasingly used alongside CNN-based monitoring pipelines in AM. Encoder–decoder architectures, including U-Net and various ResNet variants, have demonstrated exceptional performance in pixel-wise segmentation tasks such as melt pool boundary delineation, porosity identification, and surface defect characterization [43,44]. The translation equivariance property of convolutional layers makes these architectures particularly well-suited for processing the spatially structured data generated by layer-wise imaging systems common in PBF processes.

For processing temporal data streams from sensors such as pyrometers, photodiodes, and acoustic emission transducers, recurrent neural network architectures have proven essential [46,47,48]. Long short-term memory (LSTM) networks and their variants can capture the temporal dependencies inherent in AM process signals, enabling the detection of anomalies that manifest as deviations in signal dynamics rather than instantaneous outliers. More recently, temporal convolutional networks and transformer-based architectures have demonstrated competitive or superior performance while offering advantages in training efficiency and parallelization [20,21].

Generative models, including variational autoencoders (VAEs) and generative adversarial networks (GANs), represent an emerging frontier in AM research [49,50]. These architectures have been applied to generative design problems, where they can propose novel lattice structures and metamaterial configurations optimized for both mechanical performance and manufacturability [20,22]. Additionally, generative models offer solutions to the persistent challenge of data scarcity in AM by enabling synthetic data augmentation and the generation of realistic microstructure images for training downstream models.

Graph neural networks (GNNs) and point cloud processing architectures such as PointNet represent the latest generation of deep learning methods finding application in AM [20]. These approaches can process 3D geometric data in its native representation, enabling direct learning from computer-aided design (CAD) models and point cloud measurements without the information loss associated with voxelization or projection to 2D views.

Transformer architectures, which have revolutionized natural language processing and computer vision through their attention mechanisms, are increasingly being explored for AM applications [20,21]. The self-attention mechanism enables transformers to capture long-range dependencies in sequential data without the vanishing gradient problems that can affect recurrent networks. In AM contexts, transformers have been applied to multi-modal sensor fusion, where attention weights can dynamically prioritize information from different sensor streams based on relevance to the detection task. Vision transformers (ViTs) have demonstrated competitive performance with CNNs for image-based defect detection while offering advantages in interpretability through attention visualization. However, the computational requirements of transformer architectures and their data-hungry nature present challenges for deployment in resource-constrained AM monitoring scenarios [21,22]. Figure 5 provides a taxonomy of the major deep learning paradigms and their applications in AM.

Table 3. Deep learning types and algorithms with descriptions for AM applications (adapted from [21]).

DL Type	Algorithm	Description and AM Applications
Discriminative	Multi-layer perceptron (MLP)	Feedforward NN for regression/classification; process-property prediction
	Convolutional NN (CNN)	Image-based defect detection; melt pool monitoring; powder bed analysis
	Recurrent NN (RNN/LSTM)	Time-series process signals; thermal history; sequential anomaly detection
	Transformer	Attention-based models for multi-modal fusion; emerging in AM
Generative	Autoencoders (AE/VAE)	Latent representation learning; anomaly detection; data compression
	GAN	Synthetic data generation; microstructure synthesis; design augmentation
Reinforcement	Model-based (MDP)	Control policy learning with known environment dynamics
	Model-free (Q-learning)	Direct policy learning; process parameter optimization; adaptive control

provides a detailed taxonomy of deep learning types and algorithms commonly applied in AM research, along with their typical applications across the manufacturing life cycle.

2.3. Reinforcement Learning for Control

Reinforcement learning (RL) represents an emerging paradigm for adaptive process control in AM, wherein agents learn optimal control policies through iterative interaction with the manufacturing environment [51,52,53]. Unlike supervised learning approaches that require labeled training data, RL agents discover effective strategies through trial-and-error exploration, receiving reward signals that guide policy improvement. This learning paradigm is particularly appealing for AM applications where optimal control strategies may be difficult to specify a priori and where the process dynamics are sufficiently complex to warrant adaptive approaches.

The main strength of RL is policy adaptation under changing process states without explicit labeled supervision; the primary limitations are sample inefficiency, safety risks during exploration, and sim-to-real transfer gaps that currently constrain industrial deployment.

The formulation of RL for AM control typically involves defining the state space (e.g., melt pool temperature, geometry measurements, process parameters), action space (e.g., laser power adjustments, scan speed modifications), and reward function (e.g., minimizing defects, maximizing dimensional accuracy). The choice of reward function is critical and often represents the primary engineering challenge, as poorly designed rewards can lead to unexpected or suboptimal control behaviors. Multi-objective reward formulations that balance quality, productivity, and resource consumption are particularly challenging but essential for practical deployment [52,54]. Figure 6 illustrates the fundamental agent-environment interaction framework underlying RL-based control.

Recent systematic reviews have revealed that reinforcement learning approaches account for over half of AI-based control studies in laser-based AM [52]. The work by Piovarči et al. [55] demonstrated the potential of RL for closed-loop control in direct ink writing, where agents learned to adjust deposition parameters in real-time to compensate for material variability. However, despite these promising laboratory demonstrations, industrial deployment of RL-based control remains limited due to persistent challenges in sample efficiency, safety constraints during exploration, and the difficulty of transferring policies learned in simulation to physical systems [54,55]. The study by Dharmadhikari et al. [56] further illustrates the use of reinforcement learning for process-parameter optimization in AM, reinforcing the relevance of adaptive control strategies for complex manufacturing settings. More recently, deep RL has also been applied to LPBF scan strategy generation for thermal-field management, reporting substantial distortion reduction relative to conventional zigzag/chessboard patterns in experimental validation [57].

Figure 6. The proposed framework adopts a train-application dual-stage architecture to balance offline learning efficiency with online responsiveness. This design facilitates the agent to internalize complex machining logic during learning and deploy it for autonomous planning in real time [58].

Figure 6. The proposed framework adopts a train-application dual-stage architecture to balance offline learning efficiency with online responsiveness. This design facilitates the agent to internalize complex machining logic during learning and deploy it for autonomous planning in real time [58].

2.4. Physics-Informed Machine Learning

The integration of physics-based domain knowledge into machine learning models has emerged as a powerful approach for addressing the fundamental limitations of purely data-driven methods in AM applications [9,21,23,59]. Physics-informed machine learning (PIML) encompasses a spectrum of approaches that leverage physical principles to improve data efficiency, enhance generalization beyond training distributions, and ensure physically consistent predictions—all critical requirements for reliable AM process modeling. As shown in Figure 7, bidirectional prediction frameworks enable both forward modeling of process outcomes and inverse design of optimal parameters.

Physics-informed methods are advantageous for data-efficient training and physically consistent extrapolation in process–structure–property modeling; their practical limitations are higher model complexity, dependence on reliable governing equations/boundary conditions, and increased calibration effort for production-scale deployment.

Physics-informed neural networks (PINNs) represent the most prominent PIML methodology, wherein governing partial differential equations—such as the heat transfer equation, Navier–Stokes equations for melt pool flow, or solidification kinetics—are embedded as soft constraints in the neural network loss function [59,60,61]. This physics regularization enables PINNs to learn physically plausible solutions even in data-sparse regimes and to extrapolate more reliably beyond the training data domain than purely empirical models. Studies applying PINNs to thermal field prediction in LPBF have demonstrated that physics constraints can reduce data requirements by an order of magnitude while maintaining predictive accuracy [62,63]. Recent process-specific studies further reinforce this trend, including architecture-driven PIDL for LPBF thermal-history prediction under limited data conditions [64], and physics-informed recurrent models for real-time long-horizon WAAM temperature-field prediction [65]. In addition, generative deep learning work (AdditiveGDL) demonstrates constrained GAN-based prediction of local thermal distributions from toolpath information for identifying heat-accumulation-critical regions in metal AM layers, further strengthening the role of generative AI in thermal-field-aware process planning [66].

Hybrid surrogate modeling represents another important PIML paradigm, combining physics-based simulations with data-driven corrections to balance computational efficiency and predictive fidelity [2,10,37]. Multi-fidelity frameworks strategically leverage cheap low-fidelity simulations alongside expensive high-fidelity experiments or detailed simulations, with ML models learning the systematic discrepancies between fidelity levels. This approach has proven particularly effective for process optimization, where thousands of model evaluations may be required but high-fidelity simulations are computationally prohibitive. Additionally, the incorporation of physically meaningful features—such as volumetric energy density, cooling rates, and Marangoni numbers—as inputs to ML models improves interpretability and often enhances predictive performance by providing the model with relevant thermophysical context [9,10].

3. Powder Bed Fusion (PBF)

3.1. PBF Overview and Scope

Consistent with the algorithm-centered restructuring, the following process-family discussion maps each life cycle stage to the ML methods used in practice, with explicit reporting of strengths, limitations, and deployment readiness. Powder bed fusion encompasses laser powder bed fusion (LPBF, also known as SLM or DMLS) and electron beam melting (EBM), representing the most extensively studied AM process family for AI/ML applications [11,43,44,67]. The widespread industrial adoption of PBF for manufacturing high-value components in aerospace, biomedical implants, and automotive applications has driven substantial research investment in ML-based quality assurance [68,69]. The maturity of PBF monitoring infrastructure, combined with the criticality of achieving defect-free production for safety-critical applications, has made this process family the primary focus of AI/ML research in metal PBF systems.

The discussion proceeds from general PBF fundamentals (Section 3.2) to LPBF-dominant AI/ML studies and then to EBM-specific opportunities and challenges.

3.2. Process Fundamentals and Data Modalities

PBF processes involve the selective melting of thin powder layers, typically 20–100 μm in thickness, using a focused energy source that scans across the powder bed according to a predetermined trajectory [3,11]. The highly localized energy input creates extreme thermal gradients, with cooling rates reaching ${10}^6$ K/s in some cases, and it induces complex melt pool dynamics that fundamentally govern microstructure formation and defect generation [12,70]. This multi-physics nature—involving rapid heating and cooling cycles, solid–liquid–vapor phase transformations, Marangoni-driven convection within the melt pool, vapor recoil forces, and the development of residual stresses—generates rich, high-dimensional data streams that are particularly amenable to ML approaches [21,71,72].

The sensor modalities employed for process monitoring in PBF have been reviewed extensively [73,74,75,76]. Coaxial melt pool cameras operating at frame rates of 10–100 kHz provide high-speed imaging of melt pool geometry, spatter ejection events, and keyhole formation dynamics. These image sequences are typically processed using CNNs for real-time classification of process states and detection of anomalies [40,43]. Off-axis thermal cameras complement coaxial imaging by providing layer-wise temperature field measurements that enable reconstruction of thermal histories and identification of anomalous thermal signatures. Encoder–decoder neural network architectures have proven effective for semantic segmentation of these thermal images to delineate regions of concern [21,77].

3.3. Electron Beam Melting (EBM): ML Applications and Challenges

Electron beam melting (EBM) is a distinct PBF variant that uses a high-energy electron beam to selectively melt metal powders in a vacuum environment [43,69]. EBM offers advantages for processing reactive materials (e.g., titanium alloys), reduced residual stress due to high build temperatures, and rapid build rates for large components. However, the vacuum environment, high operating temperatures, and unique defect modes (e.g., swelling, lack of fusion, powder charging) present challenges for process monitoring and ML-based quality assurance.

ML applications in EBM are less mature than in LPBF, but recent work has demonstrated the use of CNNs and ANNs for in situ monitoring, microstructure prediction, and defect detection [43]. High-speed imaging and thermal data are used to identify process anomalies, while ML models predict the effects of process parameters on grain structure and mechanical properties. The limited availability of in situ sensors (due to the vacuum environment) and the scarcity of large, labeled datasets remain key barriers.

Future research directions include the development of physics-informed ML models that incorporate EBM-specific process physics, transfer learning from LPBF datasets, and the integration of digital twins for real-time process optimization. Addressing these EBM-specific constraints is critical for broader industrial adoption in aerospace, biomedical, and energy applications [43,69].

Point sensors including pyrometers and photodiodes provide spatially integrated measurements of melt pool emission at sampling rates reaching into the MHz range. Although these sensors lack spatial resolution, their high temporal bandwidth captures rapid process dynamics that may be missed by imaging systems. Time-series analysis methods, including LSTM networks and one-dimensional CNNs, have been successfully applied to classify process conditions and detect anomalies from these signals [43,78]. Acoustic emission sensors offer complementary information by capturing the mechanical signatures of solidification, crack formation, and powder spreading anomalies. Spectral feature extraction combined with classical ML classifiers such as random forests has demonstrated effectiveness for acoustic-based monitoring [43,79]. Finally, powder bed cameras provide layer-wise imaging before melting, enabling detection of recoater streaks, powder contamination, and geometric deviations that may affect subsequent processing [1,80].

3.4. AI/ML Applications Across the Life Cycle

3.4.1. Design and Process Planning

ML-based topology optimization for LPBF has evolved beyond geometric optimization to incorporate AM-specific manufacturability constraints [1,2]. Key developments include:

Printability-aware design: Neural network surrogates predict overhang angles, support volume requirements, and thermal distortion, enabling constraint integration into topology optimization loops. Generative adversarial networks (GANs) and variational autoencoders (VAEs) have been applied to generate lattice structures and metamaterial designs optimized for both mechanical performance and printability [20,22].

Process parameter optimization: This constitutes the most mature ML application in PBF, with extensive work on developing process maps relating laser power (P), scan speed (v), hatch spacing (h), and layer thickness (t) to part density, surface roughness, and mechanical properties [9,44,67]. Gaussian process regression enables uncertainty-aware optimization with active learning, reducing experimental requirements by 50–80% compared to full factorial designs [10].

Build orientation and support optimization: ML models predict build time, support volume, surface quality, and residual stress as functions of part orientation, enabling multi-objective optimization [81].

3.4.2. In Situ Monitoring and Defect Detection

ML-based monitoring in LPBF has achieved remarkable accuracy for detecting process anomalies and predicting defect formation [21,43,44]. The critical review by Herzog et al. [43] systematically analyzed sensor-algorithm combinations across 150+ studies, revealing that CNN-based approaches dominate image analysis while classical ML (SVM, random forests) remains competitive for engineered features from 1D signals. Figure 8 provides an overview of the diverse sensor modalities employed in AM monitoring.

The application of ML to defect detection in PBF has matured considerably over the past decade, with algorithms now capable of identifying the full spectrum of process-induced anomalies with high accuracy [40,42,43]. Lack-of-fusion porosity, which results from insufficient energy density leading to incomplete melting and inter-layer voids, represents one of the most studied defect types. Convolutional neural networks trained on melt pool morphology images have demonstrated detection accuracies exceeding 95% for this defect class [43,82]. Keyhole porosity, arising from excessive energy input that causes deep vapor depression and subsequent gas entrapment upon collapse, presents distinct thermal and optical signatures that have been successfully classified using both supervised and semi-supervised learning approaches [21,83]. The pioneering work on high-speed X-ray imaging combined with ML has enabled real-time keyhole detection and process window identification, providing unprecedented insight into this previously opaque phenomenon [84,85]. Earlier sensor-based monitoring studies also demonstrated the feasibility of build-condition monitoring in LPBF and in situ melt-pool-image analysis for porosity prediction [86,87].

Balling defects, characterized by unstable melt pool dynamics leading to spheroidization of the deposited material, can be detected through analysis of melt pool aspect ratios and thermal signature classification. Spatter ejection and the associated denudation of powder around the melt track affect the quality of subsequent layers and have been successfully tracked using off-axis cameras combined with object detection networks such as YOLO and Faster R-CNN [43,88]. Layer-wise powder bed imaging analyzed with semantic segmentation algorithms enables the identification of recoater streaks and powder spreading anomalies before melting occurs, allowing for potential corrective action [1,89]. Figure 9 illustrates representative examples of the diverse defect morphologies encountered in AM processes.

Table 4. Common defects in metal additive manufacturing, their formation mechanisms, and potential impacts (adapted from [21]).

Scope	Defect Type	Common Causes	Potential Effects
Local	Gas porosity	Raw material contamination (moisture, grease); gas solubility disparity (H2)	Reduction in fatigue properties
Local	Lack of fusion pores	Insufficient input energy density	Reduction in mechanical and fatigue properties
Local	Keyhole pores	Excessive heat input; keyhole instability	Roughness; internal crack risk; fatigue reduction
Local	Microstructural heterogeneities	Improper powder deposition; chemical mobility differences; preferential evaporation	Inhomogeneous mechanical properties
Local	Balling	Low/high input energy density; surface oxidation	Surface roughness; poor layer adhesion
Global	Cracking	Residual stresses; insufficient interlayer bonding	Part failure
Global	Delamination	Residual stresses; weak layer bonding	Part failure
Global	Distortion	Residual stresses exceeding local yield stress	Conformance failure; dimensional inaccuracy

summarizes common defects in metal AM, their formation mechanisms, and impacts on part quality.

The integration of multiple sensor modalities through data fusion techniques has emerged as a promising strategy for improving detection robustness and reducing false positive rates [21,91]. Early fusion approaches concatenate raw sensor signals before processing, intermediate fusion combines features extracted from individual modalities, and late fusion aggregates predictions from modality-specific models. Attention mechanisms and transformer architectures have demonstrated particular promise for learning cross-modal correlations, enabling the model to dynamically weight information from different sensors based on relevance to the detection task [21,22].

The pioneering work by Scime and Beuth [40] demonstrated the effectiveness of deep convolutional neural networks for in situ powder bed anomaly detection. Their multi-scale approach processes images at multiple resolutions to capture both fine-grained defect features and broader contextual patterns, achieving classification accuracies exceeding 93% across five distinct anomaly categories.

3.4.3. Closed-Loop Control

Despite the extensive monitoring capabilities developed for PBF processes, the implementation of true closed-loop control in commercial systems remains remarkably limited [43,52]. The systematic review by Sousa et al. [52] identified only 16 studies implementing AI-based real-time control in laser-based AM, with over half employing reinforcement learning approaches. This gap between monitoring and control capabilities reflects several fundamental challenges. The scan speeds characteristic of LPBF, typically ranging from 0.5 to 2 m/s, impose stringent latency constraints that require sub-millisecond inference times for track-level control interventions. Furthermore, the limited actuator bandwidth available for laser power modulation during scanning restricts the types of corrective actions that can be implemented in real-time. Regulatory requirements for aerospace and medical applications demand predictable, auditable control behavior that current ML-based approaches struggle to guarantee, creating additional barriers to industrial deployment [25,92].

3.4.4. Property Prediction and Qualification

The modeling of process–structure–property (PSP) relationships in PBF through integrated ML pipelines represents a frontier area of research with significant implications for part qualification [9,10,21,93]. Following the philosophy of integrated computational materials engineering (ICME), these pipelines chain multiple predictive models to connect process parameters to final part properties. Thermal history prediction constitutes the first link, wherein surrogate models based on neural networks or Gaussian processes replace computationally expensive finite element simulations for predicting temperature field evolution during processing [48,94]. Microstructure prediction models then map thermal history features—including cooling rates, thermal gradients, and solidification parameters—to grain size distributions, crystallographic texture, and phase fractions [10,95]. Finally, structure–property linkages complete the chain by mapping microstructural descriptors to mechanical properties such as yield strength, fatigue life, and fracture toughness [21,96]. Recent physics-informed deep learning work has extended this workflow to direct prediction of process parameters and molten-pool dimensions, underscoring the potential of hybrid data-driven/physics-informed approaches for AM optimization [97].

Physics-informed neural networks (PINNs) incorporating heat transfer equations and solidification kinetics show promise for improved extrapolation beyond training data domains [9,21].

3.5. Comparison of AI/ML Studies in PBF

Table 5. Representative AI/ML studies in powder bed fusion (PBF).

Study	Stage	ML Method	Data	Key Contribution
Wang et al. 2020 [1]	Multi-stage	Survey	Multi-modal	First comprehensive life cycle taxonomy; identifies gaps in microstructure/alloy design
Sing et al. 2021 [67]	Full chain	Perspective	Multi-modal	L-PBF process chain ML integration framework
Herzog et al. 2023 [43]	Monitoring	CNN, SVM, RF	Optical, thermal, acoustic	Critical sensor-algorithm comparison; identifies limited closed-loop deployment
Zhang et al. 2024 [44]	QC	DL, classical	Multi-sensor	State-of-the-art quality improvement methods
Du et al. 2025 [21]	Multi-stage	Deep learning	Various	Comprehensive DL impact on PSP relationships
Guo et al. 2022 [9]	Multi-stage	Physics-ML	Simulation + exp	Physics-informed paradigm for generalization
Johnson et al. 2020 [10]	Properties	ICME + ML	Multi-fidelity	Sequential learning for materials development
Ladani 2021 [71]	Monitoring	AI/ML	In situ	Data types and variability analysis
Ng et al. 2024 [22]	Multi-stage	Various	Multi-modal	Materials and processes integration
Breitenbach 2022 [81]	Optimization	ML	Various	Systematic review of optimization stages

summarizes representative AI/ML studies in PBF organized by life cycle stage and ML approach.

Table 6. Detailed methodology and results of representative ML studies in PBF.

Study	Material	ML Method and Architecture	Key Parameters	Results	Limitations
Scime & Beuth 2018 [40]	IN625, Ti-6Al-4V	CNN (AlexNet-based); transfer learning	P: 100–400 W; v: 400–1400 mm/s; 500 k images	93.5% accuracy for five defect classes	Limited to powder-bed anomalies; no closed-loop
Gobert et al. 2018 [42]	IN718	CNN; layer-wise imaging	28 k layer images; 195 builds	F1-score: 0.89; porosity correlation: $R^2$ = 0.84	Post-build correlation only; CT ground truth expensive
Okaro et al. 2019 [83]	SS316L	Semi-supervised; autoencoder + RF	Photodiode 100 kHz; 200 k samples	99.2% anomaly detection; reduced labeling 80%	Requires calibration per machine
Mozaffar et al. 2018 [48]	Ti-6Al-4V	RNN-LSTM; surrogate model	Thermal history; FEM training	99.4% prediction accuracy; 1000× speedup vs. FEM	Physics-agnostic; limited extrapolation
Johnson et al. 2020 [10]	IN718, SS316L	GP + Active learning; multi-fidelity	CALPHAD + experiments	60% experimental reduction; UQ enabled	Material-specific; retraining needed
Chen et al. 2021 [84]	SS316L	CNN; X-ray imaging	100 k fps synchrotron	Real-time keyhole detection; 2 ms latency	Synchrotron access limited; lab transfer gap

provides a more detailed analysis of representative experimental studies in PBF, including specific methodologies, key parameters, quantitative results, and identified limitations.

The results presented in Table 6 reveal several important trends in PBF-focused ML research. The work by Scime and Beuth [40] demonstrated that transfer learning from pre-trained image classification networks could substantially accelerate model development for powder bed anomaly detection, achieving over 93% accuracy across multiple defect categories while requiring only modest amounts of AM-specific training data. The semi-supervised approach of Okaro et al. [83] addressed the critical challenge of label scarcity by combining autoencoders for unsupervised feature learning with random forest classifiers, reducing labeling requirements by 80% while maintaining detection accuracy above 99%. The surrogate modeling approach of Mozaffar et al. [48] exemplifies the potential of recurrent networks for thermal history prediction, achieving near-perfect accuracy with computational speedups of three orders of magnitude relative to finite element analysis. However, as noted in the limitations column, persistent challenges include the need for machine-specific calibration, expensive ground truth acquisition through CT scanning, and limited generalization beyond training distributions.

Figure 10 illustrates typical sensor configurations employed for melt pool monitoring in both LPBF and DED processes, highlighting the differences in monitoring approaches between these process families.

Figure 11 presents a detailed flowchart of the ML pipeline for defect detection in SLM using layer-wise imaging, illustrating the data processing stages from image acquisition through classification.

4. Directed Energy Deposition (DED) and Wire Arc Additive Manufacturing (WAAM)

4.1. Large-Scale and Construction Additive Manufacturing: AI/ML Opportunities and Challenges

Large-scale additive manufacturing (LSAM), including construction-scale 3D printing and infrastructure applications, is an emerging frontier where AI/ML can address unique challenges [98,99,100]. Technologies such as concrete extrusion, WAAM for bridges, and hybrid robotic systems enable the fabrication of buildings, bridges, and architectural components at unprecedented scales. These processes introduce new variables—including environmental conditions, material rheology, and site logistics—that complicate process control and quality assurance.

Machine learning is increasingly used for real-time monitoring of concrete printing, predicting layer deformation, and optimizing mix design for printability and sustainability [99,100]. Computer vision systems, often based on deep learning, enable geometric verification and defect detection in large-scale prints, while reinforcement learning and digital twins are being explored for adaptive path planning and autonomous construction robots. In WAAM-based infrastructure, ML models predict bead geometry, residual stress, and structural integrity for large metal components, supporting the construction of bridges and offshore structures [98].

Key challenges include the integration of heterogeneous sensor data in outdoor environments, the need for robust models that generalize across variable site conditions, and the scarcity of large, labeled datasets for civil-scale AM. Sustainability is a major driver, with ML supporting the use of recycled materials, low-carbon binders, and energy-efficient process planning. Future research directions involve federated learning for cross-site model sharing, explainable AI for regulatory compliance, and the development of open benchmarks for construction AM [99,100]. Directed energy deposition processes constitute a diverse family of AM technologies that share the common characteristic of simultaneously delivering feedstock material and energy to a localized melt pool [101,102,103,104]. This family includes laser-based DED (variously termed L-DED, LENS, or LMD), electron beam DED (EB-DED), and wire arc additive manufacturing (WAAM), each offering distinct advantages for specific application domains. The ability to deposit material at high rates, repair existing components, and create functionally graded structures has driven significant industrial interest, particularly for large-scale aerospace and maritime applications [93,104]. The systematic trend study by Pazireh et al. [101] analyzed approximately 370 publications from 2010 to 2025, revealing that ML-related DED research experienced dramatic growth beginning around 2020, coinciding with advances in deep learning and increased sensor availability.

4.2. Process Characteristics and Monitoring Challenges

The fundamental physics of DED processes involves the creation of a melt pool through a focused energy source, into which feedstock material—either powder or wire—is continuously delivered [3,93]. Laser powder DED directs powder streams coaxially or laterally into a laser-generated melt pool, enabling precise control over composition that facilitates the fabrication of multi-material and functionally graded structures [105,106]. Laser wire DED achieves higher deposition rates by feeding wire directly into the melt, offering improved material utilization compared to powder-based approaches. Wire arc additive manufacturing employs electric arc welding processes—including metal inert gas (MIG), tungsten inert gas (TIG), and plasma arc variants—to melt wire feedstock at deposition rates reaching several kilograms per hour, making it particularly suited for large structural components [104,107]. Electron beam DED conducts wire-fed deposition in a vacuum environment, providing advantages for processing reactive materials such as titanium alloys [108].

The open deposition environment characteristic of DED processes and the substantially larger melt pools (typically 1–10 mm diameter) compared to PBF present unique challenges for process monitoring and ML-based quality assurance [73,74,109]. The three-dimensional nature of multi-pass deposition leads to complex thermal histories influenced by heat accumulation in the workpiece, inter-pass temperatures, and geometric variations as the part grows. Process dynamics in WAAM are particularly challenging due to arc instabilities, electromagnetic interference affecting sensor signals, the stochastic nature of metal transfer, and the evolution of thermal conditions as large builds progress [101,107].

The sensor modalities employed for DED monitoring share some commonality with PBF but also include process-specific data sources [102,103]. Coaxial melt pool imaging using high-speed cameras captures melt pool geometry, dimensional stability, and spatter behavior. Pyrometers and thermal cameras monitor temperature fields and inter-pass cooling, with thermal history being particularly critical for microstructure evolution in DED. For WAAM processes, arc electrical signals—including voltage, current, and derived resistance waveforms—encode rich information about process stability and metal transfer characteristics that have been extensively exploited for ML-based monitoring [107]. Additional modalities include wire feed monitoring for detecting feed rate variations and stick-slip phenomena; geometric measurement through structured light scanning and laser profilometry for bead geometry and layer height characterization; and acoustic emission sensing for crack detection and process anomaly identification [101].

Table 7. Comparison of sensory methods for in situ monitoring in AM processes (adapted from [25]).

Method	Advantages	Limitations	Typical Applications
2D imaging	Inexpensive; easy setup; rapid visual data; broad ML compatibility	Limited subsurface detection; affected by lighting and camera placement	Layer deformation; defect segmentation; real-time parameter correction
3D point cloud	High fidelity; comprehensive spatial data; real-time comparison capability	High computational cost; limited sensitivity to minor defects; complex setup	Surface roughness; defect detection; geometric accuracy
Radiography	Non-contact; subsurface defect detection; penetrates dense materials	Time-consuming; expensive; limited resolution for thick components	Pore detection; subsurface structure monitoring (metal AM)
Thermography	Non-contact; effective for surface thermal anomalies; rapid area scanning	Low penetration; ineffective for internal defects; environmental noise	Melt pool temperature; surface-level defect monitoring
Acoustic emission	Internal fault detection; volumetric information; cost-effective hardware	Sensitive to noise; critical sensor placement; complex signal analysis	Crack detection; pore formation; nozzle height monitoring
Spectroscopy	Quick analysis; real-time capable; insights into mechanical/chemical conditions	Primarily metal AM; complex data requiring advanced ML; noisy signals	Porosity monitoring; pore formation; process condition optimization

compares the advantages, limitations, and typical ML applications of different sensing modalities used across AM processes.

4.3. AI/ML Applications in DED/WAAM

4.3.1. Bead Geometry and Melt Pool Prediction

Neural networks, Gaussian process models, and ensemble methods predict bead width (w), height (h), penetration depth (d), and dilution as functions of process parameters [101,102]. Melt pool geometry prediction in DED spans multiple scales of complexity [101,102]. Single-track geometry relationships express bead width w, height h, and penetration depth d as functions of process parameters including power P, travel speed v, feed rate $\dot{m}$, and standoff distance $z_{standoff}$. Multi-track overlap predictions address inter-track porosity and surface waviness arising from the interaction between adjacent deposition passes. Multi-layer thermal accumulation models capture how progressive temperature rise affects geometry stability through altered thermal gradients and melt pool dynamics [101].

Physics-informed neural networks incorporating heat transfer and fluid flow equations have shown improved extrapolation for novel parameter combinations [101].

4.3.2. Path Planning and Deposition Strategy

ML-assisted toolpath optimization addresses the complex challenge of thermal history management in multi-layer DED builds [52,101]. Adaptive layer planning approaches predict and compensate for geometry deviation arising from thermal distortion, enabling accurate deposition even for complex geometries where cumulative errors would otherwise compromise dimensional tolerances. Dwell time optimization through ML models predicts optimal inter-layer cooling times that balance productivity considerations against quality requirements, avoiding both insufficient cooling that leads to accumulated heat and excessive cooling that reduces throughput. Deposition sequence optimization, including applications of reinforcement learning, determines scan patterns that minimize residual stress accumulation and distortion through intelligent thermal management [52].

4.3.3. Defect Detection and Quality Monitoring

DED processes are susceptible to several characteristic defect modes that have been addressed through ML-based detection systems [101,102]. Porosity arising from gas entrapment or lack of fusion can be detected via analysis of acoustic emission signals and melt pool thermal signatures. Geometric defects including humping, underfill, and overflow are detected through in situ profilometry combined with CNN-based image analysis. Surface oxidation monitoring employs spectroscopic sensors with ML classification of spectral features. Cracking detection leverages acoustic emission analysis with time-frequency features processed through random forest classifiers. For wire arc processes specifically, arc instabilities manifest in current and voltage signal patterns that LSTM networks can identify with high accuracy [52,110].

4.3.4. Closed-Loop Control in DED

DED offers more favorable conditions for implementing closed-loop control compared to PBF due to the lower scan speeds and larger feature sizes that relax real-time computational requirements [52]. Demonstrated control applications span multiple process variables. Melt pool temperature control employing PID and model predictive control strategies with ML-based surrogate models enables rapid thermal regulation. Bead geometry control through real-time laser power adjustment based on vision feedback maintains consistent deposition dimensions. Reinforcement learning-based controllers using policy gradient methods achieve adaptive parameter optimization that responds to changing conditions throughout multi-layer builds [52,54].

Sousa et al. [52] found that 62.5% of surveyed AI-based control studies in laser AM deployed controllers in real-world experiments, with reinforcement learning prominent among the approaches demonstrating practical implementation.

4.4. Comparison of AI/ML Studies in DED/WAAM

Table 8. Representative AI/ML studies in directed energy deposition.

Study	Stage	ML Method	Data	Key Contribution
Pazireh et al. 2025 [101]	Multi-stage	Trend analysis	Various	Systematic review of 370 papers; identifies growth trends and gaps
Chadha et al. 2022 [102]	Multi-stage	NN, GP, SVM	Process signals	Comprehensive AI in DED review
Sousa et al. 2025 [52]	Control	RL, ML	In situ	Systematic review of AI for real-time control; 16 studies analyzed
Ahmmed et al. 2025 [110]	DT + AI	Various	Multi-modal	Digital twin and cybersecurity integration

summarizes representative AI/ML studies in directed energy deposition.

Table 9. Detailed methodology and results of representative ML studies in DED and WAAM.

Study	Process	ML Method and Architecture	Key Parameters	Results	Limitations
Khanzadeh et al. 2018 [106]	L-DED	Self-organizing maps; clustering	Melt pool images; thermal signature	92% porosity clustering accuracy	Unsupervised; no defect classification
Ren et al. 2021 [60]	L-DED	PINN; physics-constrained	Heat equation; P: 500–2000 W	15% error reduction vs. pure DL	Simplified boundary conditions
Hamrani et al. 2024 [107]	WAAM	LSTM; arc signal analysis	V, I waveforms at 10 kHz; 50 k samples	96.8% anomaly detection; 5 ms latency	Arc noise; material-specific
Williams et al. 2018 [104]	WAAM	ANN; geometry prediction	P, v, wire feed; 200 tracks	$R^2$ = 0.91 for bead geometry	Single-material; no multi-layer
Mukherjee et al. 2021 [105]	L-DED	ML + FEM surrogate	Thermal + mechanical FEM	500× speedup; 5% error	Requires high-fidelity training data
Era et al. 2023 [103]	L-DED	CNN; thermal imaging	IR camera 100 Hz; Ti-6Al-4V	94% defect detection accuracy	Limited to surface defects

presents detailed methodology and results from representative experimental ML studies specific to DED processes, including WAAM applications.

The studies summarized in Table 9 illustrate the diversity of approaches applied to DED process monitoring and optimization. The self-organizing map approach of Khanzadeh et al. [106] demonstrated the potential for unsupervised learning to identify porosity patterns from melt pool thermal signatures without requiring expensive labeled datasets. The physics-informed neural network study by Ren et al. [60] showed that embedding the heat transfer equation as a constraint reduced prediction errors by 15% compared to purely data-driven approaches, particularly for parameter combinations outside the training distribution. For WAAM applications, Hamrani et al. [107] achieved over 96% anomaly detection accuracy from arc voltage and current signals using LSTM networks, with inference latencies of approximately 5 ms that are compatible with real-time control requirements. However, as noted in the limitations, challenges remain including sensitivity to arc noise and the need for material-specific model calibration.

Figure 12 illustrates the distribution of ML algorithms and approaches used in DED research, revealing the dominance of neural network methods and the emerging importance of physics-informed approaches.

5. Material Extrusion (MEX): FDM and FFF

Material extrusion processes, particularly fused deposition modeling (FDM) and fused filament fabrication (FFF), represent the most accessible additive manufacturing technology with the broadest adoption across consumer, educational, and industrial contexts [111,112,113]. The relatively low capital cost, ease of operation, and compatibility with a wide range of thermoplastic materials have made FDM an ideal platform for developing and validating AI/ML approaches that may subsequently be transferred to more complex AM processes [46,47,54,114]. The extensive availability of FDM systems has also enabled the generation of large experimental datasets that are often difficult to obtain for more expensive metal AM processes. Figure 13 illustrates the key components of an FDM system and the diverse sensor modalities used for process monitoring.

5.1. FDM/FFF Process Characteristics

The FDM process involves the extrusion of thermoplastic filament through a heated nozzle, typically maintained between 180–280 °C depending on material composition, with the molten material deposited in a predetermined pattern onto a heated build platform [6]. The process is governed by a multidimensional parameter space encompassing thermal variables (nozzle temperature, bed temperature, chamber temperature), motion parameters (print speed, travel speed, acceleration, jerk settings), geometric parameters (layer height, line width, infill density and pattern), and material flow controls (extrusion multiplier, retraction settings) [111,115].

The comprehensive reviews by Nasrin et al. [111] and Hassan et al. [116] emphasize that the application of ML to polymer AM must contend with material-specific challenges that distinguish it from metal AM. The molecular complexity of polymers introduces dependencies on rheological behavior, crystallization kinetics, and chain orientation that significantly affect final part properties. Environmental sensitivity manifests through humidity absorption affecting filament consistency and ambient temperature variations influencing cooling rates and interlayer bonding. Feedstock variability, including batch-to-batch variations in filament diameter and additive distribution, introduces additional uncertainty that ML models must accommodate. Furthermore, the strong anisotropic properties characteristic of FDM parts—with significant differences in mechanical performance along and perpendicular to the layer direction—complicate property prediction and necessitate consideration of build orientation effects [114,117].

5.2. AI/ML Applications

5.2.1. Process Parameter Optimization

The development of ML models for predicting part quality metrics as functions of process parameters represents the most established application domain in FDM [111,112,118]. Neural networks and ensemble methods such as random forests have demonstrated strong performance in predicting dimensional accuracy, including shrinkage compensation, warpage tendency, and geometric deviation from nominal specifications. Surface roughness prediction models establish relationships between layer height, print speed, temperature settings, and resulting surface finish characterized by parameters such as $R_a$. Mechanical property prediction—encompassing tensile strength, impact resistance, and fatigue performance—has proven more challenging due to the complex dependencies on both process parameters and material state, though hybrid models incorporating physics-based features have shown promise [111,119]. Multi-objective optimization frameworks balance competing objectives such as part quality, build time, and material consumption, with Bayesian optimization using Gaussian processes enabling efficient exploration of high-dimensional parameter spaces while providing uncertainty estimates that guide experimental resource allocation [111,120]. Figure 14 demonstrates the integration of design of experiments with neural network architectures for FDM optimization. Recent primary studies further strengthen this direction: ML-enabled parameter recommendation workflows have been demonstrated for targeting desired mechanical-property outcomes in material extrusion, providing direct process-window guidance rather than only post hoc prediction [121].

5.2.2. Real-Time Fault Monitoring

The monitoring and detection of faults during FDM processing has received substantial research attention, driven by the accessibility of sensor integration on these platforms and the practical industrial need for quality assurance in distributed manufacturing scenarios [46,47,122]. The systematic review by Pereira et al. [47] analyzed ML models for quality control in extrusion-based AM, revealing the diversity of approaches employed and the generally high classification accuracies achieved. New integrated data-driven studies in FDM have also reported autonomous print-process optimization coupled with in situ anomaly detection, indicating a shift from passive quality inspection toward corrective intelligence during the build [123,124].

Vision-based monitoring using cameras represents the most prevalent sensing approach for FDM fault detection [114,125]. Layer-wise imaging enables geometric verification through comparison with expected contours, while real-time nozzle monitoring facilitates detection of extrusion inconsistencies and clogging events. Convolutional neural network architectures, including VGG, ResNet, and custom designs optimized for the FDM domain, have demonstrated classification accuracies exceeding 95% for common defect types including under-extrusion, over-extrusion, stringing, and layer adhesion failures [47,126]. Object detection frameworks such as YOLO have proven effective for identifying and localizing specific defects like stringing and surface blobs in real-time video streams [46].

Beyond vision systems, acoustic sensors provide complementary monitoring capabilities for FDM processes [46]. Characteristic acoustic signatures enable detection of nozzle clogging, filament grinding against drive gears, and slipping events. Spectral features extracted from acoustic signals have been successfully classified using support vector machines and random forest ensembles to distinguish between normal operation and various fault modes [46,47].

Vibration and inertial measurement unit (IMU) sensors enable detection of mechanical anomalies including belt slip, bearing wear, and print head instabilities [47]. Time-series analysis of vibration signals using LSTM networks captures the temporal patterns characteristic of specific mechanical faults. Temperature sensors monitoring hot-end and heated bed conditions enable detection of thermal anomalies and prediction of thermal runaway conditions that could pose safety hazards [46].

Current and power monitoring through observation of motor loads enables detection of mechanical issues within the motion system and provides indirect estimation of extrusion forces.

Common FDM defects that have been successfully addressed through ML-based detection systems include under-extrusion, where insufficient material flow causes gaps in the deposited structure; over-extrusion, where excess material creates surface blobs and dimensional inaccuracy; warping from thermal contraction causing part detachment from the build platform; layer separation or delamination from poor inter-layer adhesion; stringing from unwanted filament trails during travel moves between deposition sites; nozzle clogging through partial or complete blockage of the extrusion path; and z-wobble causing layer misalignment arising from mechanical instabilities [46,47].

5.2.3. Closed-Loop Control and AI-Augmented Manufacturing

FDM serves as a primary platform for closed-loop AI-augmented AM research due to the favorable process dynamics that permit implementation of control loops with achievable latency requirements [54,113]. The AI2AM (AI-augmented additive manufacturing) framework proposed by Sani et al. [54] integrates multiple capabilities into a cohesive intelligent manufacturing system. Real-time monitoring through multi-sensor data acquisition combined with ML-based anomaly detection provides continuous process assessment. Adaptive parameter adjustment implements closed-loop control of temperature, speed, and flow rate based on detected conditions. Predictive intervention capabilities anticipate defects before they occur based on analysis of process trajectory patterns [54]. These developments are consistent with emerging multi-view deep information fusion frameworks in material extrusion AM that combine online quality monitoring with autonomous correction actions, supporting transition toward practically deployable closed-loop polymer AM systems [124].

Demonstrated closed-loop capabilities in FDM include temperature control based on thermal camera feedback that maintains optimal processing conditions despite environmental variations; extrusion rate adjustment based on layer geometry measurement to compensate for dimensional deviations; print pause and parameter modification protocols triggered upon defect detection to prevent error propagation; and reinforcement learning approaches for adaptive optimization of infill patterns and print speeds that respond to evolving build conditions [54,113].

Rajendran et al. [113] articulate a vision for fully autonomous AI-integrated FDM systems incorporating predictive maintenance that anticipates component failures, generative design integration that automatically produces manufacturable geometries, and self-optimizing production lines that continuously improve performance through operational experience.

5.3. Comparison of AI/ML Studies in MEX

Table 10. Representative AI/ML studies in material extrusion (FDM/FFF).

Study	Stage	ML Method	Data	Key Contribution
Nasrin et al. 2023 [111]	Multi-stage	Various	Process data	Comprehensive polymer AM review; polymer- specific challenges
Nikooharf et al. 2024 [112]	Multi-stage	ML survey	Multi-modal	ML pipeline for polymer AM
Sampedro et al. 2022 [46]	Monitoring	ML classification	Acoustic, vibration	Sensor-modality-driven fault monitoring taxonomy
Pereira et al. 2025 [47]	QC	CNN, LSTM, SVM	Multi-sensor	Systematic QC review; CNN dominance for image data
Sani et al. 2024 [54]	Control	AI2AM	Multi-modal	Closed-loop framework; challenges to deployment
Rajendran et al. 2024 [113]	Multi-stage	AI integration	Various	Vision for autonomous FDM systems

summarizes representative AI/ML studies in material extrusion.

The studies in this subsection span the full material extrusion workflow, from process parameter optimization and real-time fault monitoring to closed-loop control and autonomous system concepts. They were selected to highlight the range of ML methods used in FDM/FFF research, the diversity of sensor and process data involved, and the main deployment limitations that continue to affect practical adoption. The table below organizes these studies by stage and method to support a clearer comparison.

Table 11. Detailed methodology and results of representative ML studies in material extrusion (FDM/FFF).

Study	Material	ML Method and Architecture	Key Parameters	Results	Limitations
Qi et al. 2019 [119]	PLA, ABS	ANN; Taguchi DOE	T: 190–230 °C; v: 30–90 mm/s; 81 experiments	$R^2$ = 0.94 tensile strength; 8% error	Limited to single-objective
Li et al. 2022 [126]	PLA	CNN (ResNet-50); transfer learning	Camera 30 fps; 15 k images	97.2% defect detection; real-time	Lighting sensitive; single material
Gonzalez et al. 2020 [125]	PETG	YOLO v3; object detection	Layer images; eight defect classes	mAP = 0.89; 20 fps inference	Training data intensive
Ghiasian et al. 2018 [120]	Various	GP-BO; multi-objective	Layer height, speed, infill	45% experimental reduction vs. grid	Slow convergence high-dim
Goh et al. 2021 [114]	PLA, CF-PLA	Ensemble (RF + XGB); feature engineering	50 process features; 500 prints	$R^2$ = 0.92 mechanical properties	Feature engineering manual
Piovarči et al. 2022 [55]	Silicone	Deep RL (PPO); closed-loop	DIW; camera feedback	40% defect reduction vs. open-loop	Sim-to-real gap; slow training

provides detailed experimental methodology and quantitative results from representative studies applying ML to FDM/FFF processes.

The results presented in Table 11 highlight both the achievements and persistent challenges in applying ML to FDM processes. The artificial neural network study by Qi et al. [119] demonstrates that carefully designed experiments using Taguchi methods combined with neural network modeling can achieve high predictive accuracy ($R^2$ = 0.94) for tensile strength prediction from process parameters. Transfer learning approaches, as demonstrated by Li et al. [126] using ResNet-50 pre-trained on ImageNet, enable rapid development of accurate defect detection systems (97.2% accuracy) with relatively modest amounts of FDM-specific training data. The Bayesian optimization approach of Ghiasian et al. [120] reduced experimental requirements by 45% compared to grid search while handling multi-objective optimization, though convergence slowed in high-dimensional parameter spaces. Perhaps most significantly, the reinforcement learning study by Piovarči et al. [55] demonstrated the potential for closed-loop control, achieving 40% defect reduction compared to open-loop operation, though challenges remain in bridging the simulation-to-reality gap and reducing the training time required for policy convergence.

The preceding comparison highlights the iterative cycle of experimentation, model training, and optimization that characterizes data-driven process development.

Figure 15 illustrates the general framework for applying ML to process parameter optimization in AM, showing the iterative cycle of experimentation, model training, and optimization that characterizes data-driven process development.

Figure 16 presents a comprehensive taxonomy of ML types and their applications across different areas of 3D printing, highlighting the matching between algorithm characteristics and application requirements.

6. Vat Photopolymerization (VP)

Vat photopolymerization encompasses stereolithography (SLA), digital light processing (DLP), continuous liquid interface production (CLIP), and two-photon polymerization (2PP/TPP), offering high resolution (sub-10 μm) for polymer, ceramic, and biomedical applications [1,22].

6.1. Process Overview

VP processes selectively cure liquid photopolymer resin using controlled light exposure, with several distinct technology variants serving different applications [1,22]. Stereolithography (SLA) employs a UV laser that scans and cures resin point-by-point, building parts through sequential solidification of individual voxels. Digital light processing (DLP) uses a digital projector to cure entire layers simultaneously, offering faster build rates for parts that fit within the projection field. Continuous liquid interface production (CLIP) introduces an oxygen-inhibition window that enables continuous printing without discrete layers, eliminating layer interfaces and improving surface quality. Two-photon polymerization (2PP) exploits nonlinear two-photon absorption to achieve sub-micron resolution for applications requiring the highest precision [22,27].

6.2. Current State of AI/ML Integration

VP processes are relatively under-represented in the AI/ML-for-AM literature compared to PBF and MEX [1,22,27]. Several factors contribute to this gap. The inherently more controlled environment of photopolymerization results in fewer defect modes than metal AM processes, reducing the urgency for intelligent monitoring systems. Lower industrial adoption rates compared to metal and polymer extrusion processes mean fewer high-value applications driving quality-focused research investments. Additionally, the narrower material palette available for VP constrains application diversity and the corresponding research attention [22].

Despite this relative under-representation, meaningful ML applications have emerged in the VP domain. Exposure optimization models predict optimal curing parameters including exposure time, intensity, and wavelength combinations required to achieve desired geometric accuracy and mechanical properties [27]. Resin formulation research employs ML for property prediction in novel photopolymer and ceramic slurry compositions, accelerating materials development. Support structure optimization through automated generation of self-supporting geometries and minimal support designs reduces post-processing requirements and material waste. Print orientation optimization algorithms minimize support requirements and build time through intelligent part placement. Dimensional compensation approaches predict and compensate for the shrinkage and post-cure deformation characteristic of photopolymer materials [1,22].

Emerging opportunities suggest growing research attention for VP in the coming years. Biomedical applications including patient-specific implants, tissue engineering scaffolds, and drug delivery systems impose stringent quality requirements that will drive ML-based quality assurance development. Precision optics manufacturing for micro-optics and photonic devices requires tight tolerances that may benefit from ML-based process optimization. Ceramic AM via VP for high-performance ceramic components requires defect prediction capabilities to ensure structural integrity. Multi-material VP systems with functional gradient capabilities represent an emerging technology class requiring sophisticated process control [22].

Table 12. AI/ML application areas in vat photopolymerization (VP) processes.

Application	ML Methods	Current Status	Key Challenges
Exposure optimization	NN, GP regression	Early research	Material-dependent curing kinetics
Resin formulation	ANN, random forest	Limited studies	Complex rheology– property relationships
Support optimization	Topology optimization + ML	Emerging	Multi-objective trade-offs
Dimensional compensation	Regression models	Moderate activity	Anisotropic shrinkage prediction
Bioprinting	CNN, RL	Growing interest	Cell viability constraints; bioink complexity
Ceramic VP	Neural networks	Very limited	Debinding/sintering prediction

summarizes the key AI/ML application areas for vat photopolymerization processes, their current development status, and future opportunities.

The relatively sparse VP-specific AI/ML literature represents a significant research gap, particularly as applications in dental aligners, hearing aids, and biomedical devices continue to expand [22,27].

Figure 17 provides an overview of AM process categorization and the corresponding ML application areas, highlighting how different process families present unique opportunities and challenges for AI/ML integration.

7. Binder Jetting (BJ)

Binder jetting deposits liquid binder onto powder beds comprising metals, ceramics, sand, or polymers, followed by thermal processing including curing, debinding, and sintering to achieve final part density [1,27].

7.1. Process Characteristics

The BJ process chain involves multiple sequential stages, each introducing parameters that influence final part quality [27]. The printing stage selectively deposits binder to create green parts with sufficient handling strength for subsequent operations. Curing through thermal treatment strengthens the binder to enable safe handling during depowdering, which removes unbound powder from the green part. Finally, sintering or infiltration densification treatments achieve the mechanical properties required for functional applications. This multi-step chain introduces unique challenges for ML integration, as final quality depends on parameter interactions across all stages, with early-stage decisions constraining later optimization possibilities [1].

7.2. AI/ML Opportunities and Challenges

Several ML application areas have been explored or identified for BJ processes. Binder saturation optimization addresses the critical balance between green strength, dimensional accuracy, and sintering behavior through predictive models relating binder levels to multiple quality outcomes [27]. Powder spreading analysis employs machine vision to detect layer defects, voids, and density variations before binder deposition commits the layer. Sintering shrinkage prediction enables compensation for the anisotropic dimensional changes, often 15–25%, that occur during densification. Density and porosity prediction models relate process parameters across the full chain to final part density distributions. Microstructure evolution predictions address grain growth and phase transformations occurring during high-temperature sintering [1].

However, significant challenges limit ML adoption in BJ. The multi-stage process chain requires end-to-end optimization approaches capable of modeling complex interdependencies across printing, curing, and sintering stages. Long feedback cycles inherent to sintering processes, which may require hours to days, delay quality feedback and impede rapid iteration for model development. Material diversity spanning wide ranges of powder-binder systems with different behaviors complicates the development of generalizable models. Limited commercial deployment compared to PBF and FDM processes results in smaller datasets and fewer research groups focusing on BJ-specific challenges [27].

Research in BJ-specific AI/ML remains limited compared to other process families, representing a significant gap that warrants increased attention as BJ adoption grows for metal and ceramic part production [1,27].

Table 13. AI/ML challenges and opportunities specific to binder jetting (BJ) processes.

Process Stage	ML Opportunities	Data Sources	Challenges
Printing	Binder saturation optimization; droplet placement	Printhead sensors; cameras	Binder–powder interaction modeling
Curing	Thermal profile optimization	Temperature sensors; time-series	Limited available data
Depowdering	Automated inspection; fragile part handling	Vision systems; force sensors	Green part fragility
Sintering	Shrinkage prediction; density optimization	Thermal profiles; dimensional scans	15–25% anisotropic shrinkage; hours-long cycles
End-to-End	Multi-stage optimization; PSP relationships	Combined data streams	Complex interdependencies; delayed feedback

summarizes the unique challenges and opportunities for AI/ML integration in binder jetting processes, highlighting the multi-stage nature that distinguishes this process family.

8. Material Jetting (MJT)

Material jetting technologies, including PolyJet, MultiJet, and inkjet printing approaches, deposit droplets of photopolymer, wax, or metal nanoparticle inks that are subsequently cured or sintered to form parts [22,27].

8.1. Process Characteristics

MJT systems offer unique capabilities that distinguish them from other AM technologies [22]. High resolution achieved through layer thicknesses of 16–32 μm enables fine feature reproduction suitable for detailed prototypes and precision components. Multi-material capability through simultaneous deposition of multiple materials enables complex assemblies in single builds. Full CMYK color capability supports realistic visual prototypes for design verification. Most distinctively, voxel-level material control enables functionally graded parts with spatially varying properties through precise compositional gradients [27].

8.2. AI/ML Applications

Current research areas for MJT encompass several aspects of the droplet-based deposition process [22,27]. Droplet dynamics modeling predicts spreading, coalescence, and surface formation behaviors that determine dimensional accuracy and surface quality. Print head management systems detect and compensate for nozzle failures that can introduce systematic defects. Color accuracy through ML-based color calibration and compensation ensures faithful reproduction of intended appearances. Multi-material interface prediction and optimization addresses the challenging transitions between different materials that affect both aesthetics and mechanical performance. Texture and surface finish control at micro-scales enables functional surfaces with tailored properties [22].

Emerging opportunities in specialized MJT applications include bioprinting, where cell-laden bioink deposition requires precise control to maintain cell viability and achieve desired tissue architectures. Electronics printing employing conductive ink deposition for printed electronics and sensors demands ML-based optimization of sintering and conductivity. Pharmaceutical applications exploiting drug-loaded polymer printing for personalized medicine require precise dosage control and dissolution behavior prediction [22,27].

Table 14. Emerging AI/ML applications in material jetting (MJT) processes.

Application	ML Approach	Potential Impact	Current Status
Droplet dynamics	CFD + ML surrogate	Improved deposition accuracy	Research stage
Nozzle management	CNN for jet monitoring	Reduced print failures	Limited industrial use
Multi-material interfaces	Neural network optimization	Enhanced mechanical properties	Emerging research
Bioprinting	RL for viability optimization	Cell survival >90%	Active research area
Printed electronics	Conductivity prediction	Reliable electrical performance	Early development
Pharmaceutical	Dissolution modeling	Personalized drug delivery	Regulatory limited

outlines the emerging AI/ML application areas in material jetting processes.

Dedicated AI/ML research for MJT remains sparse in the published literature, with most work embedded in broader AM reviews rather than process-specific investigations [22,27].

9. Sheet Lamination (SL)

Sheet lamination processes, including laminated object manufacturing (LOM) and ultrasonic additive manufacturing (UAM), bond sheets of material through adhesive bonding, ultrasonic welding, or laser welding to construct parts layer by layer [1,27].

9.1. Process Variants

Several distinct process variants fall under the sheet lamination category. LOM bonds paper, plastic, or metal sheets with adhesive and shapes them using laser or knife cutting, offering rapid and cost-effective prototyping for large parts. UAM employs ultrasonic welding to bond metal foils of aluminum, copper, titanium, and steel alloys at relatively low temperatures, enabling multi-material structures and the embedding of sensors and fibers within solid metal parts without thermal damage. Selective lamination approaches represent hybrid techniques combining sheet bonding with other AM or subtractive processes for specialized applications [1,27].

9.2. AI/ML Opportunities

UAM presents particularly interesting opportunities for ML applications given its complex process physics and unique capabilities [27]. Bond quality prediction models relate ultrasonic parameters including amplitude, force, and speed to interface strength and void content at the bond interface. Embedded component integration optimization determines parameters for fiber and sensor embedding that achieve mechanical integration without damaging sensitive components. Multi-material bonding prediction addresses the compatibility and optimal parameters for joining dissimilar material combinations, a key UAM advantage. Thermal management modeling predicts and controls heat generation during welding to prevent thermal damage to previously deposited layers or embedded components [1].

Published AI/ML research specifically targeting sheet lamination processes remains very limited, representing perhaps the largest gap among AM process families [27]. The relatively niche application space and smaller research community contribute to this under-representation. As UAM capabilities for multi-material and smart structures attract growing interest, research attention may increase [1,27].

Table 15. AI/ML opportunities in sheet lamination (SL) and ultrasonic additive manufacturing (UAM).

Process Aspect	ML Opportunity	Sensor Inputs	Research Status
Bond quality (UAM)	Interface strength prediction	Ultrasonic power; force; thermal	Virtually unexplored
Multi-material joining	Compatibility prediction	Material properties; process parameters	No published studies
Embedded components	Damage prevention optimization	Force sensors; thermal imaging	Very limited
Thermal management	Heat generation modeling	Thermocouples; IR cameras	Early conceptual
Cutting optimization (LOM)	Path planning; kerf prediction	Vision systems	Basic implementations

highlights the key AI/ML opportunities specific to sheet lamination and ultrasonic additive manufacturing.

10. 4D Printing: Stimuli-Responsive Additive Manufacturing

4D printing extends the capabilities of traditional 3D printing by enabling the fabrication of objects that can change their shape, properties, or function over time in response to external stimuli such as temperature, moisture, light, magnetic fields, or pH [127,128]. This paradigm leverages smart materials—including shape memory polymers, hydrogels, and composites—to create structures with programmable, time-dependent behavior, opening new frontiers in biomedical devices, soft robotics, aerospace, and consumer products.

10.1. Fundamentals and Materials

4D printing relies on the integration of stimuli-responsive materials with advanced AM techniques. Commonly used materials include shape memory alloys, shape memory polymers, hydrogels, and multi-material composites. The design of 4D-printed objects requires careful consideration of material properties, printing parameters, and the intended stimulus-response mechanism [127].

10.2. Applications

Key applications of 4D printing include:

• Biomedical devices: Self-deploying stents, tissue scaffolds, drug delivery systems, and patient-specific implants that adapt to physiological conditions.
• Soft robotics: Actuators and grippers that change shape or function in response to environmental cues.
• Aerospace: Morphing structures for adaptive aerodynamics and deployable components.
• Consumer products: Smart textiles, responsive packaging, and toys.

10.3. AI/ML Integration in 4D Printing

Artificial intelligence and machine learning play a crucial role in 4D printing by enabling:

• Design optimization: ML algorithms can predict and optimize the stimulus-response behavior of complex geometries and multi-material systems, reducing the need for costly experimental iterations [128].
• Process parameter tuning: Data-driven models help identify optimal printing parameters for achieving desired actuation profiles and response times.
• Inverse design: Generative models and reinforcement learning can be used to design structures that achieve specific time-dependent transformations under given stimuli.
• Real-time monitoring and control: ML-based sensor fusion enables in situ monitoring of shape change and adaptive feedback control during and after printing.

10.4. Challenges and Research Outlook

Despite rapid progress, several challenges remain for widespread adoption of 4D printing:

• Limited availability of robust, biocompatible, and multi-responsive materials.
• Complex multi-physics modeling of coupled stimulus-response behaviors.
• Need for standardized testing and qualification protocols.
• Integration of digital twins and AI for predictive life cycle management.

Future research will likely focus on the development of new smart materials, advanced multi-material printing strategies, and the integration of AI-driven design and control frameworks for fully autonomous, adaptive manufacturing systems [127,128].

11. Cross-Cutting Themes and Challenges

11.1. Federated Learning, Blockchain, and Distributed AI in Additive Manufacturing

The increasing digitization and connectivity of AM systems create new opportunities and challenges for distributed AI, privacy, and secure data sharing. Federated learning enables collaborative model training across distributed manufacturing sites or organizations without requiring raw data to leave local premises, preserving intellectual property and sensitive information [20,110]. This approach is particularly valuable for rare defect detection, sustainability metrics, and biomedical applications where data privacy is paramount. Edge AI and continual learning further support real-time adaptation and model improvement across machine fleets [129].

Blockchain and distributed ledger technologies are being explored for secure, transparent tracking of part provenance, process parameters, and material flows throughout the AM life cycle [130]. Blockchain can provide immutable audit trails for quality assurance, regulatory compliance, and supply chain traceability, while supporting decentralized marketplaces for design and manufacturing services. Integration with AI/ML models enables automated verification of process compliance and anomaly detection in distributed production networks.

Key challenges include the scalability and latency of blockchain solutions for high-throughput AM environments, the need for standardized data schemas and interoperability, and the development of privacy-preserving ML algorithms that balance utility and confidentiality. Future research will focus on hybrid architectures combining federated learning, blockchain, and digital twins to enable secure, adaptive, and trustworthy AM ecosystems [20,130].

11.2. AI/ML for Bioprinting and Nanomaterial-Enabled Additive Manufacturing

Bioprinting and nanomaterial-enabled additive manufacturing represent rapidly advancing frontiers with unique challenges for AI/ML integration [22,27]. Bioprinting involves the precise deposition of cell-laden bioinks to fabricate tissue scaffolds, organoids, and functional biological constructs. Machine learning is used to optimize print parameters for cell viability, predict tissue maturation, and enable real-time monitoring of construct geometry and cell distribution. Reinforcement learning and computer vision approaches are being developed for adaptive control of extrusion and droplet-based bioprinting, supporting the fabrication of complex tissue architectures.

Nanomaterial integration in AM, including the use of nanoparticles, nanofibers, and functionalized hydrogels, enables the creation of parts with enhanced mechanical, electrical, or biological properties. ML models assist in predicting the effects of nanomaterial loading on printability, mechanical performance, and functional response. Data-driven approaches accelerate the discovery of optimal nanocomposite formulations and guide the design of hierarchical structures for targeted applications in electronics, energy, and biomedical devices.

Key challenges include the limited availability of high-quality, multi-modal datasets (e.g., imaging, omics, mechanical testing), the need for interpretable models to support regulatory approval in biomedical applications, and the integration of multi-scale simulation with experimental data. Future research directions involve the use of digital twins for tissue engineering, federated learning for collaborative model development across institutions, and explainable AI for safe and effective clinical translation [22].

11.3. Sustainability-Focused Machine Learning in Additive Manufacturing

The integration of sustainability principles into additive manufacturing is a rapidly growing area, with AI/ML playing a pivotal role in enabling green manufacturing, resource efficiency, and circular economy models [128,131]. Machine learning can optimize process parameters to minimize energy consumption, material waste, and emissions, while predictive maintenance and quality control reduce scrap rates and extend equipment life. Life cycle assessment (LCA) models, enhanced by ML, allow for real-time evaluation of environmental impacts across the product life cycle.

Recent research has demonstrated the use of neural networks, genetic algorithms, and response surface methodology (RSM) to identify optimal trade-offs between part quality, throughput, and environmental footprint [131]. AI-driven design for sustainability (DfS) tools can recommend lightweighting strategies, material substitutions, and end-of-life recycling pathways. In distributed manufacturing scenarios, federated learning enables collaborative model development for sustainability metrics without sharing proprietary data.

Key challenges include the lack of standardized sustainability metrics for AM, limited availability of high-quality environmental data, and the need for interpretable models that can support regulatory compliance and eco-labeling. Future directions involve the integration of blockchain for transparent tracking of material flows and carbon footprints, and the development of digital twins that incorporate environmental performance as a core optimization objective [130,131].

The preceding process-specific sections reveal several cross-cutting challenges and research themes that emerge consistently across all AM process families, transcending the particulars of individual technologies [2,9,17,25]. Addressing these challenges will be essential for advancing AI/ML applications in AM from laboratory demonstrations to industrial deployment. Figure 18 provides an overview of how ML integrates across diverse AM process stages and applications.

11.4. Comparative Evolution of ML Models in AM

Table 16. Comparative evolution of ML model families in AM, including advantages, limitations, and practical applicability.

Model Family	Typical Strengths	Typical Limitations	Representative Performance	Representative AM Use Case
Classical ML (SVM/RF/GP)	Works with small/structured datasets; interpretable baselines; UQ (GP)	Feature engineering burden; weaker raw-image learning	89–99% anomaly/classification metrics in curated settings	Fast deployment for tabular process-property mapping and acoustic anomaly screening in PBF/DED
CNN-based DL	Strong image-feature extraction; high defect-detection sensitivity	Data-hungry; transfer across machines may degrade	93–97% reported detection accuracy in LPBF/FDM studies	In situ vision monitoring for powder-bed anomalies and extrusion faults
Sequence models (LSTM/TCN)	Captures temporal process dynamics; suitable for sensor streams	Drift-sensitive; sequence labeling cost	Up to 96.8% anomaly detection for WAAM arc signals	Online monitoring of arc/thermal/current signals in DED/WAAM
RL controllers	Adaptive policies and autonomous recovery potential	Sample inefficiency; safety and sim-to-real gaps	40% defect reduction demonstrated in DIW/FDM-like settings	Closed-loop parameter correction under changing build conditions
Physics-informed ML (PINN/Hybrid)	Better extrapolation; lower data requirements; physically consistent outputs	Model complexity; PDE and boundary-condition sensitivity	15% error reduction or order-of- magnitude data savings in reported studies	Thermal-field and process-structure-property prediction for metal AM qualification workflows

compares model families in terms of strengths, limitations, deployment readiness, and representative quantitative performance ranges reported in this review. This synthesis is further supported by recent comprehensive metal-AM-focused reviews that critically analyze life cycle-level AI integration barriers and deployment gaps [132].

The literature trend indicates a progression from feature-engineered classical models (early adoption phase), to vision-dominant deep learning (monitoring scale-up phase), and now toward physics-informed and control-oriented methods (industrialization phase). This transition is driven by three factors: increased sensor bandwidth, demand for robust cross-machine generalization, and the need for certifiable model behavior in safety-critical sectors.

Practical applicability differs by objective: classical ML and CNNs are currently strongest for quality monitoring and defect detection, while RL and hierarchical AI remain pre-industrial for most high-consequence metal AM production environments.

Application-level evidence further clarifies this differentiation. In LPBF and EBM monitoring pipelines, CNN-based image models consistently provide strong defect/anomaly screening performance (typically in the low-90%-to-high-90% range in controlled studies), making them useful for near-term in situ quality surveillance; however, their cross-machine and cross-material robustness often degrades without adaptation. In DED/WAAM, temporal models (e.g., LSTM-class approaches) are particularly useful for arc/thermal signal interpretation and online anomaly tracking, but sensitivity to drift and process non-stationarity remains a practical limitation. Physics-informed and hybrid models improve extrapolation and reduce experimental burden in thermal/process–structure–property prediction workflows, which is valuable for qualification-oriented studies. By contrast, RL-based control demonstrates promising autonomous correction behavior in pilot-scale settings, yet still faces sample-efficiency and safety-validation barriers that limit routine industrial deployment.

Table 17. Summary of key challenges in deploying AI/ML for additive manufacturing.

Challenge Category	Specific Issues	Impact	Mitigation Strategies
Data scarcity	Limited labeled datasets; rare defect examples; expensive ground truth acquisition	Underfitting; poor generalization; biased models	Transfer learning; synthetic data via GANs/simulations; active learning
Data heterogeneity	Non-standardized formats; varying sensor configurations; inconsistent metadata	Impedes data pooling; limits reproducibility	Standardized schemas (ASTM); federated learning; data ontologies
Model generalization	Machine-to-machine transfer; material changes; parameter window limits	Narrow validity; production brittleness	Domain adaptation; physics-informed constraints; multi-fidelity learning
Real-time implementation	Sub-millisecond latency for PBF; computational constraints; hardware integration	Prevents closed-loop control; limited industrial deployment	Edge AI; model compression; FPGA/GPU acceleration
Interpretability	Black-box models; trust deficits; certification barriers	Regulatory rejection; operator skepticism	Explainable AI (SHAP, Grad-CAM); physics-informed models; attention mechanisms
Digital twin integration	Data synchronization; model orchestration; cybersecurity	System complexity; vulnerability propagation	Standardized APIs; robust networking; holistic security architectures
Life cycle integration	Siloed tools; missing feedback loops; workflow barriers	Sub-optimal overall performance; manual hand-offs	End-to-end pipelines; digital threads; organizational change

summarizes the key challenges facing AI/ML deployment in AM, categorized by domain, along with their impact and potential mitigation strategies.

11.5. Data Challenges

The availability of high-quality, labeled datasets remains the primary bottleneck constraining ML advancement in AM applications [21,43,44,133]. Several inter-related issues contribute to this challenge. Rare defect modes, by their very nature, generate insufficient training examples for supervised learning approaches, yet these uncommon failures are often the most consequential for part qualification. Ground truth acquisition for defect characterization typically requires expensive post-build analysis such as computed tomography scanning or destructive mechanical testing, creating economic barriers to large-scale dataset generation. The resulting datasets are often severely imbalanced, with the “normal” or “acceptable” class dominating, which can bias model training and inflate apparent accuracy metrics. Furthermore, the AM research community lacks publicly available benchmark datasets comparable to those that have accelerated progress in fields such as computer vision and natural language processing [17,134].

Figure 19 illustrates the application of thermal imaging and machine learning for melt pool analysis, demonstrating how infrared data streams can be processed through ML pipelines for real-time quality assessment.

Synthetic data generation through physics-based simulations and generative models, including GANs and more recently diffusion models, offers promising approaches to address data scarcity [20,22]. Simulated datasets can provide arbitrary quantities of labeled examples, including rare defect scenarios that may be dangerous or impossible to generate experimentally. However, the fidelity of synthetic data to real-world conditions requires careful validation, and the “sim-to-real” gap remains a significant concern for models trained primarily on simulated data [54,55]. For polymer extrusion specifically, recent systematic synthesis of multiscale thermal prediction models highlights the need to couple physics-informed learning with digital-twin-oriented thermal history estimation, reinforcing the importance of process-aware models rather than purely black-box approximators [135].

Data heterogeneity across the AM ecosystem poses additional challenges for large-scale model development [2,15,17]. Sensor data formats, sampling rates, coordinate system conventions, and metadata structures vary significantly across machine manufacturers and models. Different research groups employ diverse experimental protocols, data collection methodologies, and annotation practices. This heterogeneity severely impedes data pooling efforts that would enable the training of larger, more generalizable models. The lack of standardized data schemas for AM—encompassing process parameter definitions and units, sensor metadata and calibration information, spatial coordinate systems and registration protocols, quality metrics and defect taxonomies, and model documentation requirements—represents a critical infrastructure gap [1,25]. While efforts toward AM data standards are emerging through organizations such as ASTM International and various consortia developing AM-specific ontologies, widespread adoption remains limited.

11.6. Model Generalization and Transfer

A persistent limitation of published ML models for AM is their narrow validity, typically restricted to specific combinations of machine, material, and process parameter range under which training data were collected [27,43,101,129]. This brittleness manifests across multiple dimensions. Machine-to-machine transfer is complicated by differences in beam profiles, scanner dynamics, chamber atmospheres, and powder handling systems that alter process signatures even for nominally identical parameter settings. Material-to-material transfer faces challenges from divergent thermal properties, melt pool behavior, and defect formation mechanisms across different alloy systems or polymer formulations. Models trained on limited parameter windows may fail catastrophically when applied at the boundaries of or beyond their training domain. Additionally, temporal drift arising from machine wear, powder recycling degradation, and environmental variations over time can erode model performance in longitudinal deployment scenarios.

The research community has proposed several approaches to address these generalization challenges [9,21,136]. Transfer learning with domain adaptation layers enables models pre-trained on data-rich source domains to be fine-tuned for data-sparse target domains with minimal additional training examples [136]. Physics-informed constraints that embed fundamental governing equations improve extrapolation behavior by preventing predictions that violate physical principles [9,23,59]. Multi-fidelity learning frameworks strategically combine expensive experimental data with cheaper simulation results to expand the effective training distribution [10,36]. Federated learning architectures enable collaborative model training across distributed manufacturing facilities while preserving data privacy, though practical implementations in AM remain nascent [20,110]. Most ambitiously, the development of foundation models pre-trained on large, diverse AM datasets could provide adaptable representations for downstream tasks, analogous to language models in natural language processing [133,137].

11.7. Real-Time Implementation

The translation of ML models from offline analysis to real-time closed-loop control demands inference latencies compatible with process dynamics, which vary substantially across AM technologies [43,52,54,109]. In LPBF, where scan speeds typically range from 0.5 to 2 m/s, track-level control interventions require sub-millisecond decision making to modify laser parameters before the beam has traversed a significant distance. DED processes operate at comparatively lower speeds of 5–20 mm/s, relaxing latency requirements to the 10–100 ms range. FDM systems, with typical print speeds of 30–100 mm/s, can accommodate response times on the order of 100 ms to 1 s for meaningful corrective action.

Achieving these latencies requires addressing multiple implementation challenges. Model compression and quantization techniques enable deployment of neural network models on edge devices with constrained computational resources [91]. Hardware acceleration through GPUs, FPGAs, and dedicated AI chips can achieve the throughput necessary for real-time inference. Deterministic inference timing becomes critical for safety-critical control applications where bounded response times must be guaranteed. Integration with machine control systems and adherence to industrial communication protocols present additional engineering challenges for practical deployment [138].

Table 18. Computational latency requirements for real-time ML deployment across AM processes.

Process	Typical Speed	Control Level	Required Latency	Hardware Options
LPBF	0.5–2 m/s	Track-level	<1 ms	FPGA; dedicated AI chips
LPBF	–	Layer-level	10–100 ms	GPU; high-end CPU
DED	5–20 mm/s	Track-level	10–100 ms	GPU; embedded systems
WAAM	5–15 mm/s	Layer-level	100–500 ms	Standard GPU
FDM	30–100 mm/s	Real-time	100 ms–1 s	Edge devices; Raspberry Pi
VP	Varies by layer	Layer-level	1–10 s	Standard computing

summarizes the computational latency requirements across different AM process families and application scenarios.

11.8. Interpretability, Trust, and Certification

Regulatory acceptance of ML-based quality assurance requires interpretable, auditable models whose decision-making processes can be understood and validated [21,25,43,120]. Explainability addresses the fundamental need to understand why a model predicts a particular outcome, whether defect or acceptable quality, enabling human operators to assess the reasonableness of model predictions. Uncertainty quantification provides confidence bounds on predictions essential for risk assessment in safety-critical applications [63]. Failure mode analysis characterizes when and how models fail, identifying the boundaries of reliable operation and informing appropriate application contexts. Comprehensive audit trails documenting model versions, training data provenance, and prediction histories support regulatory compliance and quality system requirements [139].

Physics-informed approaches improve interpretability by embedding known physical relationships, making model behavior more transparent, though this structural constraint may sacrifice some flexibility compared to purely data-driven approaches [23,37]. Attention mechanisms and gradient-based explanation techniques such as Grad-CAM and SHAP values provide post hoc interpretability for deep learning models, highlighting the features most influential in model predictions.

For safety-critical aerospace and medical applications, certification frameworks for ML-based manufacturing quality systems remain under active development, representing a crucial enabler for broader industrial adoption [140].

11.9. Digital Twin Integration

Digital twins (DTs) provide comprehensive architectural frameworks for integrating AI/ML capabilities across the complete AM life cycle [25,92,105,110,141]. Real-time process mirroring maintains a virtual replica continuously updated with live sensor data, enabling monitoring and visualization of process states [142]. Predictive capabilities powered by ML enable forecasting of quality outcomes, maintenance requirements, and long-term performance trajectories. What-if analysis through virtual experimentation supports process optimization without consuming physical resources [18]. Historical tracking through complete digital threads spanning design through qualification enables traceability and continuous improvement [25].

However, significant challenges impede DT–ML integration. Data infrastructure requirements for real-time synchronization between physical processes and digital replicas demand robust networking and storage systems. Model management and version control across potentially hundreds of ML components operating within a digital twin framework require sophisticated software engineering practices. Computational resource requirements for maintaining high-fidelity simulations alongside real-time ML inference can be substantial. Critically, digitalization introduces new cyber-physical attack surfaces, requiring robust cybersecurity measures alongside AI capabilities [110].

The comprehensive review by Ahmmed et al. [110] emphasizes that the integration of digital twins, AI/ML, and AM creates interconnected systems where vulnerabilities in one domain can propagate to others, necessitating holistic security architectures.

11.10. Life Cycle Integration Gap

Despite significant progress in developing ML solutions for individual life cycle stages, truly integrated design-to-qualification AI pipelines remain largely conceptual [2,25]. The current state of the field reveals several persistent gaps. Most implementations continue to address single life cycle stages in isolation, with separate models for design optimization, process parameter selection, in situ monitoring, and post-build qualification operating independently without systematic information exchange. Feedback loops from post-process qualification back to design optimization, which would enable iterative improvement based on as-built performance, remain rare in both research and industrial practice. End-to-end data flow and model chaining across the life cycle present significant research challenges in terms of data format compatibility, uncertainty propagation, and computational orchestration. Beyond technical barriers, organizational structures and workflow conventions within manufacturing enterprises often impede the kind of cross-functional integration that holistic AI systems would require [2,9].

The vision of fully integrated workflows—encompassing design, process planning, closed-loop build control, post-processing optimization, property prediction, and feedback to design—as articulated in panoramic reviews of the field [2,25], remains a compelling but unrealized goal. Achieving this vision will require not only advances in individual ML components but also fundamental progress in systems integration, data infrastructure, and manufacturing informatics [129,143].

12. Future Directions

Based on the analysis of the current literature and identified gaps, this section outlines key research directions for advancing AI/ML in additive manufacturing. These directions span from near-term improvements to existing approaches through longer-term paradigm shifts in how AM systems are designed, operated, and certified.

Accordingly, the remainder of this section is organized as forward-looking research priorities (including PINN/PIML, RL, and foundation-model directions) aimed at advancing industrially deployable AI-enabled AM systems.

Figure 20 provides an integrated roadmap view of the challenge-opportunity landscape that motivates the research directions discussed in this section.

Table 19. Summary of future research directions in AI/ML for additive manufacturing.

Research Direction	Key Approaches	Current Status	Timeline	Expected Impact
Autonomous AM	Hierarchical AI agents; self-optimization; cognitive systems	Early research	5–10 years	Lights-out manufacturing; in-space production
Physics-informed ML	PINNs; hybrid surrogates; neural operators	Active research	2–5 years	Improved extrapolation; reduced data needs
Foundation models	Pre-trained multi-modal models; few-shot learning	Emerging	3–7 years	Cross-process transfer; rapid deployment
Edge AI	On-machine inference; federated learning; model compression	Advancing	2–4 years	Real-time control; distributed manufacturing
Advanced RL	Safe exploration; sim-to-real transfer; hierarchical control	Limited deployment	3–6 years	Adaptive closed-loop control
Standardization	Benchmark datasets; evaluation metrics; data schemas	Early efforts	2–5 years	Reproducibility; fair comparison
Human–AI Collaboration	Explainable AI; mixed-initiative systems; operator interfaces	Growing interest	1–3 years	Enhanced productivity; trust building

summarizes the key future research directions, their current status, expected timeline, and potential impact on the field.

12.1. Short-, Medium-, and Long-Term Research Roadmap

Table 20. Roadmap classification for future AI/ML directions in AM.

Horizon	Timeline	Priority Themes	Rationale
Short-term	1–3 years	Benchmark datasets, explainable monitoring, edge deployment pilots, standards alignment	Leverages mature sensing pipelines and existing model classes; primarily integration and validation work.
Medium-term	3–6 years	Safe RL, multi-fidelity pipelines, federated learning across plants, digital-twin orchestration	Requires nontrivial software/infrastructure upgrades and robust cross-domain transfer validation.
Long-term	6–10 years	Autonomous AM cells, foundation-model workflows, certification-ready closed-loop autonomy	Depends on trustworthy autonomy, regulatory pathways, and stable life cycle-scale data ecosystems.

maps key directions into short-term (1–3 years), medium-term (3–6 years), and long-term (6–10 years) horizons based on current TRL, data infrastructure maturity, and certification readiness.

12.2. Towards Autonomous Additive Manufacturing

The vision of autonomous AM systems represents a paradigm shift from current human-supervised operations toward manufacturing systems capable of independent decision-making and self-optimization [25]. Fan et al. have proposed a hierarchical framework comprising four integrated layers that collectively enable autonomous operation. The knowledge layer provides structured databases of materials, processes, design rules, and historical experience captured in knowledge graphs and ontologies that serve as the foundation for intelligent reasoning. Building on this foundation, the generative solution layer employs AI-driven design exploration, process planning, and parameter optimization using generative models and optimization algorithms to propose candidate solutions. The operational layer handles real-time monitoring, anomaly detection, and process control through ML-based decision support systems that maintain process quality during execution. At the highest level, the cognitive layer deploys high-level AI agents capable of observation, analysis, planning, and execution of tasks that traditionally require human intervention [25,143].

Several key capabilities must be developed to realize autonomous AM. Self-optimization through continuous learning from production data can progressively improve process parameters without explicit human tuning. Automated anomaly diagnosis and root cause analysis can identify the underlying causes of detected defects, moving beyond simple anomaly detection to actionable insights. Adaptive recovery mechanisms can enable dynamic responses to process deviations without requiring human intervention to restart or adjust builds. Lifelong learning capabilities can allow systems to accumulate knowledge across builds and materials over time, developing expertise analogous to that of experienced operators [25,43].

Applications in challenging environments such as in-space manufacturing, where human intervention is limited or impossible and communication delays preclude remote supervision, provide compelling use cases that may drive autonomous AM development [25,26]. The 5–10 year estimate for broad autonomous deployment reflects current gaps in certifiable closed-loop control, cross-machine robustness, and process-wide digital-thread integration rather than limitations of individual model accuracy alone.

12.3. Physics-Informed and Multi-Fidelity Learning

Hybrid approaches that systematically combine physics-based models with data-driven components address key limitations of purely data-driven ML, including poor extrapolation, high data requirements, and limited interpretability [9,10,21,23,37,59]. Figure 21 illustrates the architecture of physics-informed neural networks for AM applications.

Physics-informed neural networks (PINNs) embed governing equations—heat transfer, fluid flow, solidification kinetics—as soft constraints during training, guiding models toward physically plausible predictions [59,60,61]. This approach yields several benefits: improved extrapolation beyond training data domains because physical laws remain valid where data are absent; reduced data requirements through physics-based regularization that constrains the hypothesis space; enhanced interpretability through predictions that conform to known physical principles; and natural handling of noisy or sparse experimental data by leveraging physical constraints as additional information [23,62].

As shown in Figure 22, physics-informed neural network (PINN) framework integrates data-driven learning with physics-based inputs and constraints to enable accurate thermal field prediction in metal additive manufacturing.

Multi-fidelity frameworks offer another avenue for improving data efficiency by strategically combining expensive high-fidelity experiments with cheaper computational simulations [10,10]. Transfer learning techniques enable models pre-trained on abundant simulation data to be adapted to experimental conditions with limited fine-tuning [136]. Active learning strategies optimize the selection of additional experiments or simulations to maximize information gain per data acquisition cost [36]. Rigorous uncertainty propagation across fidelity levels ensures that model predictions appropriately reflect the contributions of different data sources [63].

Emerging approaches extend these concepts further. Neural operators such as DeepONet and Fourier Neural Operators learn solution operators that map input functions to output functions, enabling rapid predictions for parametric families of physics problems [145]. Graph neural networks provide natural representations for mesh-based simulations and irregular geometries common in AM [20]. Differentiable physics frameworks enable end-to-end gradient-based optimization through physics simulations, opening new possibilities for integrated design-process optimization [9].

12.4. Foundation Models and Large-Scale Learning

Recent advances in foundation models for natural language processing and computer vision, where models pre-trained on massive datasets exhibit remarkable adaptability to diverse downstream tasks, suggest analogous opportunities for AM [20,25,133,137]. AM foundation models could be pre-trained on large, diverse datasets spanning multiple processes, materials, and machine platforms, learning transferable representations that capture fundamental aspects of AM physics and defect formation [133,137]. Multi-modal learning approaches could jointly learn from images, time-series sensor data, process parameters, material specifications, and even text documentation describing procedures and outcomes [91]. Few-shot adaptation capabilities would enable rapid fine-tuning to new materials or machines with minimal additional data, dramatically reducing the cost of deploying ML for novel applications [21,136]. Natural language interfaces powered by large language models could provide intuitive access to process planning recommendations and troubleshooting guidance for operators without ML expertise [25].

Significant challenges remain in realizing this vision, including assembling sufficiently large and diverse training datasets given the fragmented nature of AM data, ensuring model robustness and safety for manufacturing-critical applications, and managing the substantial computational requirements for training and deploying such models [17,129,143].

12.5. Advanced Sensor Integration and Edge AI

Next-generation monitoring capabilities require coordinated advances in both sensing technologies and computational infrastructure [22,43,73,75]. Emerging sensor technologies promise unprecedented visibility into AM processes. High-speed X-ray imaging enables real-time subsurface defect detection by visualizing melt pool dynamics, keyhole formation, and pore nucleation as they occur [84,85]. Hyperspectral imaging provides composition monitoring capabilities valuable for multi-material and functionally graded structures. Quantum sensors offer extreme sensitivity for magnetic field and stress measurement that could enable new non-destructive evaluation modalities. Distributed fiber-optic sensors can map temperature and strain fields across entire build volumes with high spatial resolution [22].

Effectively utilizing these rich data streams requires advances in edge AI deployment. On-machine inference using specialized AI accelerators such as GPUs, FPGAs, and dedicated AI chips can achieve the latency requirements for real-time control [146,147]. Federated learning across machine fleets enables collaborative model training while keeping sensitive manufacturing data local to individual facilities [20]. Privacy-preserving computation techniques protect intellectual property while enabling model improvement from distributed data sources [110]. Real-time model updating through continual learning allows deployed models to adapt to changing conditions without full retraining [129].

12.6. Reinforcement Learning for Adaptive Control

Reinforcement learning offers a natural framework for learning control policies through direct interaction with the process, discovering optimal strategies without requiring explicit supervision [52,54,55]. The current state of RL in AM shows promise, with over half of AI-based control studies employing RL approaches according to recent surveys [52]. However, practical deployment remains limited by several challenges that define the key research priorities in this area.

Sample efficiency presents a critical concern, as RL algorithms typically require thousands to millions of interaction episodes to learn effective policies—a requirement incompatible with expensive physical AM experiments. Model-based RL approaches that learn dynamics models and plan using those models, combined with sim-to-real transfer techniques that enable policies trained in simulation to transfer to physical systems, offer paths to reducing experimental requirements [54,55]. Safety constraints must be rigorously enforced, ensuring that exploration during learning remains within safe parameter bounds that do not damage equipment or produce hazardous conditions. Constrained RL formulations that guarantee constraint satisfaction during learning are essential for practical deployment [54]. Multi-objective control that simultaneously balances quality, productivity, and resource consumption requires RL formulations capable of handling competing objectives with appropriate trade-off mechanisms. Hierarchical control architectures that combine track-level, layer-level, and part-level decision making can decompose the control problem into tractable sub-problems while maintaining global coherence [52].

12.7. Standardization and Benchmarking

Sustained community-wide efforts are essential for establishing the infrastructure required for systematic progress in AI/ML for AM [2,17,25]. Publicly available benchmark datasets with thorough documentation, covering diverse processes, materials, and defect types, would enable fair algorithm comparison and accelerate research by reducing duplicative data collection efforts. Standardized evaluation metrics for different task types—detection, classification, regression, control—would ensure consistent reporting and meaningful cross-study comparisons. Reproducibility guidelines establishing documentation standards for ML studies in AM, including requirements for reporting data preprocessing, model architecture details, training procedures, and statistical validation, would improve the reliability of published results. Interoperable data exchange formats defining schemas for process and quality data would enable data pooling and model sharing across institutions [1,17].

12.8. Human–AI Collaboration

Rather than pursuing full automation as an immediate goal, effective human–AI teaming may represent a more realistic and valuable paradigm for the near term [25,133]. AI-assisted decision support systems can augment the capabilities of process engineers by synthesizing complex data streams and providing actionable recommendations while leaving final decisions to human judgment. Explainable recommendations that clearly communicate the reasoning behind suggestions, along with associated uncertainty estimates, enable operators to appropriately calibrate their trust in AI guidance. Human-in-the-loop learning frameworks allow continuous model improvement through operator feedback, leveraging human expertise to correct model errors and handle novel situations. Mixed-initiative systems that dynamically balance automation and human control based on situation complexity and model confidence can combine the consistency of algorithmic processing with the adaptability of human oversight [25].

13. Conclusions

This review has synthesized the rapidly evolving landscape of AI/ML applications in additive manufacturing, organized systematically by ISO/ASTM process families and AM life cycle stages to provide a comprehensive panoramic perspective on the field.

Several key findings emerge from this analysis. First, a clear maturity gradient exists across application domains: ML-based process parameter optimization and defect detection represent the most mature applications, with CNN-based monitoring systems achieving high accuracy across PBF and MEX processes and demonstrating readiness for industrial deployment. In contrast, closed-loop control systems and integrated life cycle approaches remain largely at the research prototype stage, with substantial development required before production use.

Second, significant process coverage imbalance characterizes the literature. Metal AM, particularly LPBF, dominates published research, with extensive work addressing monitoring, defect detection, and process–structure–property relationships for metallic materials. Polymer AM through FDM/FFF shows strong recent growth, particularly in practical fault monitoring applications accessible to a broad user base. However, vat photopolymerization, binder jetting, material jetting, and sheet lamination remain significantly underexplored, presenting opportunities for future research to address these gaps.

Third, methodological evolution continues to reshape the field. Deep learning has become essential for processing high-dimensional sensor data, while physics-informed approaches show particular promise for improving generalization and interpretability. Reinforcement learning for adaptive control is emerging as a promising paradigm but has not yet reached production readiness. The integration of multiple methodological approaches—combining data-driven learning with physics-based constraints and uncertainty quantification—increasingly characterizes leading research.

Fourth, persistent challenges continue to limit the translation of research advances to industrial deployment. Data scarcity, model generalization across machines and materials, real-time computational constraints, and certification requirements for safety-critical applications represent inter-related barriers that require coordinated advances in standardization, physics–ML integration, and system-level architectures to overcome.

Fifth, the future trajectory of the field points toward increasing integration and autonomy. The field is evolving from isolated intelligent tools addressing individual tasks toward integrated, autonomous AM ecosystems enabled by digital twins, hierarchical AI architectures, and physics-informed learning frameworks that span the complete manufacturing life cycle.

Recent evidence reinforces these findings across both metal and polymer AM: RL-based LPBF scan-strategy optimization with experimentally validated thermal/distortion gains [57], process-aware thermal prediction advances through PIDL/PIML formulations for LPBF and WAAM [64,65,148], and emerging autonomous material-extrusion workflows that couple in situ monitoring with corrective control and target-property parameter recommendation [121,123,124]. At the synthesis level, recent review evidence also supports the life cycle-scale integration priorities and deployment barriers highlighted in this paper [132,135].

Across the representative studies synthesized in this review, defect/anomaly detection in monitoring-heavy workflows frequently reports performance in the low-90%-to-high-90% range, while closed-loop AI control evidence remains limited to a small number of studies and mostly pilot-scale demonstrations. In addition, data-efficient methods such as active learning and multi-fidelity modeling have reported substantial experimental burden reductions (typically in the order of 45–60%), indicating clear near-term value for industrial process development pipelines.

The vision of truly autonomous, self-optimizing AM systems capable of design, process optimization, real-time quality assurance, and continuous improvement without human intervention remains a compelling research frontier. Realizing this vision holds the potential to transform manufacturing across industries, enabling distributed production of customized components with unprecedented quality and efficiency.

Table 21. Summary comparison of AI/ML maturity across AM process families.

Process	ML Maturity	Dominant Methods	Primary Applications	Key Research Gaps
PBF (LPBF)	High	CNN, LSTM, GP, PINN	Defect detection; parameter optimization; property prediction	Closed-loop control; cross-machine transfer; certification
PBF (EBM)	Medium	CNN, ANN	In situ monitoring; microstructure prediction	Smaller research base; vacuum environment challenges
DED	Medium–High	NN, GP, RL	Bead geometry; thermal history; adaptive control	Multi-layer thermal accumulation; large-scale builds
WAAM	Medium	LSTM, RF	Arc signal analysis; geometry prediction	Arc instability; material diversity
MEX (FDM)	Medium–High	CNN, RF, XGBoost	Fault monitoring; parameter optimization	Closed-loop control; multi-material
VP (SLA/DLP)	Low	ANN, regression	Exposure optimization; dimensional compensation	Limited research attention; biomedical applications
BJ	Low	NN, GP	Sintering prediction; binder optimization	Multi-stage chain; long feedback cycles
MJT	Very Low	Limited	Droplet dynamics; color calibration	Minimal dedicated research; niche applications
SL (UAM)	Very Low	Limited	Bond quality prediction	Virtually unexplored; unique physics

provides a comparative summary of AI/ML maturity and key research needs across AM process families, synthesizing the findings from this comprehensive review.

Table 22. Application-priority suggestions for AI/ML adoption by sector and life cycle stage.

Sector	Near-Term AI Focus	Medium-Term AI Focus	Implementation Note
Aerospace	In situ defect detection and traceable qualification analytics	Physics-informed PSP prediction and certifiable adaptive control	Prioritize auditability and uncertainty quantification for certification pathways.
Biomedical	Geometry/porosity quality prediction and lot-level consistency	Patient-specific optimization and digital-twin-informed qualification	Emphasize explainability, biocompatibility constraints, and validation traceability.
Automotive	Throughput-quality optimization and predictive maintenance	Plant-scale federated learning and closed-loop productivity tuning	Focus on scalable deployment and model portability across machine fleets.
Construction/Large-scale AM	Geometric verification and path correction under field variability	Autonomous robotic planning with robust environmental adaptation	Require robust sensing under outdoor disturbances and site-level data standards.

provides priority suggestions for AI/ML adoption by sector and life cycle stage.