A Review of Machine Learning Applications on Direct Energy Deposition Additive Manufacturing—A Trend Study

Abstract

This review explores the evolution and current state of machine learning (ML) and artificial intelligence (AI) applications in direct energy deposition (DED) and wire arc additive manufacturing (WAAM) processes. A Python-based automated search script was developed to systematically retrieve relevant literature using the Crossref API, yielding around 370 papers published between 2010 and July 2025. The study identifies significant growth in ML-related DED research starting in 2020, with increasing adoption of advanced techniques such as deep learning, fuzzy logic, and hybrid physics-informed models. A year-by-year trend analysis is presented, and a comprehensive categorization of the literature is provided to highlight dominant application areas, including process optimization, real-time monitoring, defect detection, and melt pool prediction. Key challenges, such as limited closed-loop control, lack of generalization across systems, and insufficient modeling of deposition-location effects, are discussed. Finally, future research directions are outlined, emphasizing the need for integrated thermo-mechanical models, uncertainty quantification, and adaptive control strategies. This review serves as a resource for researchers aiming to advance intelligent control and predictive modeling in DED-based additive manufacturing.

1. Introduction

Additive manufacturing (AM), commonly known as 3D printing, has revolutionized modern manufacturing by enabling the layer-by-layer fabrication of complex geometries directly from digital models. Originally conceptualized in the 1970s and commercialized in the 1980s with technologies such as stereolithography, AM has evolved to support a wide range of materials and applications, from prototyping to end-use part production [1,2,3]. Its ability to create complex designs, reduce waste, and enable rapid iteration has made AM a key pillar in advanced manufacturing strategies across industries.

Among the various AM techniques, directed energy deposition (DED) stands out as a powerful process for fabricating and repairing large metallic components [4,5,6]. In DED, metallic powder or wire feedstock is delivered to a melt pool generated by a focused energy source, such as a laser, electron beam, or plasma arc. This approach enables localized material addition and is compatible with multi-axis and hybrid machining systems. Depending on the energy source and feedstock form, DED systems are typically classified into four main types, laser-based, electron beam-based, plasma arc-based, and wire arc additive manufacturing (WAAM), each offering trade-offs in deposition rate, resolution, and material flexibility, as illustrated in Figure 1.

Despite its advantages, DED poses significant challenges in terms of process stability and part quality. Issues such as thermal gradients, residual stress accumulation, porosity formation, and inconsistent bead geometry can compromise the mechanical integrity of the final component [7,8]. The process also suffers from high computational costs when modeled using finite element simulations with element activation techniques, particularly for large or complex parts [9]. Moreover, the interdependence of material properties, process parameters, and geometry makes it difficult to identify optimal processing conditions through experimental trials alone.

It should be noted that other additive manufacturing (AM) processes, such as selective laser melting (SLM) and electron beam melting (EBM), also face critical challenges in geometry control, process stability, and thermal management, especially when working with refractory metals (e.g., tungsten, tantalum) or lightweight alloys (e.g., magnesium- and aluminum-based systems). For instance, refractory metals in SLM/EBM are prone to cracking, poor fusion, and thermal stress accumulation due to their high melting points and thermal conductivities [10,11], while low-melting-point materials such as Mg and Al alloys introduce risks of evaporation, porosity, and instability that necessitate advanced real-time monitoring and process control [12,13]. Although these challenges parallel those in DED/WAAM, the present review deliberately focuses on DED and WAAM, whose continuous feedstock delivery, melt-pool dynamics, and toolpath-driven geometry present unique opportunities and demands for ML/AI integration.

To address these challenges, recent research has increasingly turned to advanced data-driven techniques such as artificial intelligence (AI) and machine learning (ML). These methods offer a promising approach to model complex relationships between process parameters and part quality, optimize deposition strategies, and enable adaptive process control. By learning from experimental or sensor-acquired data, ML algorithms can support real-time decision making and reduce reliance on computationally intensive numerical simulations [14,15]. As DED systems become more complex, featuring multi-axis platforms and complex geometries, the demand for intelligent, flexible, and scalable control solutions continues to grow (see Figure 2). As shown in Figure 2a,b, the geometrical complexity, multi-axis deposition, and localized heat effects present significant modeling and control challenges that drive the need for intelligent, adaptive solutions.

ML can be broadly categorized into supervised, unsupervised, and reinforcement learning paradigms. Supervised learning uses labeled datasets to train predictive models, while unsupervised learning identifies hidden patterns within unlabeled data. Reinforcement learning, in contrast, optimizes control policies through trial-and-error interactions with the system [16,17]. These paradigms have enabled various applications in AM, including defect detection, quality prediction, and process parameter optimization. However, despite the growing body of research, current reviews have primarily focused on narrow subsets of DED-related ML research and lack a comprehensive view of the field’s evolution.

This review builds upon prior works, including the conceptual analysis by [18,19], by offering a significantly broader and deeper investigation. Specifically, we (1) use an automated literature retrieval method to collect 370 papers from 2010 to July 2025; (2) conduct a detailed year-by-year trend analysis; (3) categorize studies based on specific ML techniques (e.g., regression, neural networks, fuzzy logic); and (4) emphasize emerging topics such as location-aware modeling, temporal learning, and closed-loop control in DED and WAAM. These contributions distinguish this review and provide a data-driven foundation for guiding future research in intelligent additive manufacturing.

This paper aims to fill that gap by presenting a systematic review and trend analysis of machine learning applications in DED and WAAM processes. By leveraging a Python-based automated literature retrieval tool (Python 3.10), we extracted and analyzed about 370 papers published between 2010 and July 2025. The review categorizes the studies by ML methodology (e.g., regression, fuzzy logic, deep learning, physics-informed ML), application type (e.g., prediction, classification, control), and process scope. Beyond retrospective synthesis, the paper highlights emerging directions such as location-aware modeling, temporal dynamics, and closed-loop control integration. This work offers a structured foundation for researchers aiming to advance intelligent, adaptive additive manufacturing systems.

The remainder of this paper is structured as follows. Section 2 outlines the search methodology and keyword-based filtering approach used to retrieve the dataset. Section 3 presents a year-by-year analysis of ML applications in DED and WAAM, highlighting trends and research growth. Section 4 discusses key technical challenges in integrating ML into AM processes, while Section 5 outlines promising future directions. Finally, Section 6 summarizes the major insights and contributions of this review.

2. Search Strategy

An automated retrieval approach was employed to collect publications relevant to machine learning applications in DED and WAAM processes. Using the Crossref API, papers were queried based on combinations of two keyword sets: one representing DED-related manufacturing processes and the other reflecting machine learning methodologies. Titles and abstracts were screened to ensure the co-occurrence of both keyword categories, and duplicate entries were removed. The search covered publications from 2010 to July 2025, yielding a comprehensive and filtered dataset for subsequent trend analysis.

Two keyword sets were defined to guide the search. The first set (Keyword Set 1) included terms related to DED manufacturing techniques, while the second set (Keyword Set 2) comprised common machine learning methods, as shown in

Table 1. Keyword sets used for the automated literature search.

Keyword Set 1 (DED Processes)	Keyword Set 2 (Machine Learning Methods)
Direct Energy Deposition	Clustering
Beads Clad	Machine Learning
Wire Arc	Machine-Learning
Laser Cladding	Unsupervised
Metal Deposition	Supervised
Directed Energy Deposition	Artificial Intelligence
	Data-Driven
	Reinforcement Learning
	Decision Tree
	Regression
	Neural Network
	Principal Component Analysis
	Random Forest
	K-means
	Support Vector Machine
	K-nearest
	Fuzzy
	Deep Learning

.

For each pairwise combination of a term from Keyword Set 1 and Keyword Set 2, the Crossref database was queried. An initial filter retained only papers whose titles or abstracts included at least one term from each keyword set. Duplicate entries were identified and removed based on DOI, title, and author name matching. The resulting dataset included metadata such as title, journal, publication year, and authors, which were saved in a structured format for analysis.

A Python script was developed to automate the querying process, ensuring consistency and reproducibility. Duplicate results, caused by overlapping keyword matches, were programmatically eliminated. Each entry was assessed for relevance based on the presence of at least one DED-related and one ML-related term in the title or abstract. A final manual check was conducted to confirm that the study explicitly applied ML to DED or WAAM.

This process enabled the construction of a reproducible dataset representing a wide range of studies at the intersection of machine learning and DED/WAAM. The Python code used for dataset construction is provided in the Appendix A. It is important to note that the dataset is constrained by the Crossref API’s indexing scope and does not include publications from proprietary databases (e.g., Scopus, Web of Science) or non-indexed repositories.

Both peer-reviewed and non-peer-reviewed papers were retrieved. While all were included in the trend analysis, some non-peer-reviewed publications were excluded from citation if they lacked sufficient bibliographic detail (e.g., journal name or DOI). However, they were retained in the total publication count reported in the next section.

While the automated search method is reproducible and scalable, it is limited to content indexed by the Crossref database. This constraint may introduce bias by excluding papers published in journals indexed only by proprietary platforms (e.g., Scopus or Web of Science). Despite this, Crossref remains a widely adopted, open-source bibliographic registry, making it suitable for large-scale trend analysis. Future efforts could expand coverage by integrating other databases or APIs to reduce potential selection bias. It is important to note that the novelty of this work lies not in the retrieval mechanism itself but in the trend analysis derived from the dataset. The automated retrieval code functions as a support tool, ensuring mixed keyword coverage, exact keyword naming, and reproducibility across years. While this approach may not capture every single paper in the field, it provides a sufficiently representative dataset that enables a systematic year-based analysis of evolving research trends in ML applications for DED/WAAM.

3. Trends and Insights

The number of publications related to the application of artificial intelligence and machine learning techniques in DED processes has increased substantially over the past decade. Figure 3 and Figure 4 present the trends observed based on the dataset collected through the automated search process described in Section 2. Specifically, Figure 3 illustrates the overall growth in the number of published papers from 2010 to mid-2025 (up to July), while Figure 4 categorizes the publications according to the specific AI and ML subject areas applied within DED and WAAM processes. It is noteworthy that the year 2020 marks a significant inflection point, with a sharp rise in the number of machine learning-related studies applied to DED processes. In addition, it was observed that the keyword “machine learning” appeared more frequently than specific terms such as “regression” and “artificial neural networks (ANN)” across the retrieved publications. Moreover, the topics of “fuzzy logic” and “deep learning” have shown noticeable growth in recent years, indicating emerging research interest in both adaptive control strategies and advanced neural network architectures within DED processes.

Figure 5 presents the yearly trends of the six most common ML methods applied in DED research. Neural networks dominate the field, with a strong rise after 2020, and appear to be stabilizing at a high volume. Regression and deep learning approaches show continued upward momentum. Clustering, fuzzy logic, and data-driven methods also show gradual increases, reflecting growing diversity in ML application areas.

While this section is organized chronologically, its purpose is not merely to report publication counts. Rather, the year-by-year analysis reveals meaningful patterns in how specific ML methods have evolved in response to technological advancements and research demands in DED and WAAM. For instance, the rise of neural networks post-2020 aligns with increased use of GPU availability, while the recent growth in regression and deep learning methods reflects a trend toward data-rich, model-driven process control. Thus, this temporal breakdown provides both a timeline and insight into shifting methodological preferences and applications.

The following subsections provide a detailed year-by-year discussion of these developments.

3.1. Early Developments—Years 2010 to 2015

Research on applying AI and ML to DED/WAAM was relatively limited in the early years, with around 11 publications across 2010–2015. Nonetheless, these studies established important foundations and demonstrated the feasibility of data-driven models in addressing process control and quality issues.

One of the first applications of AI in DED processes was conducted using a fuzzy logic-based inverse dynamic model constructed from input–output data to control the scanning speed for achieving the desired geometry and quality in the laser cladding process [20]. That study introduced a surface response function for the laser scanning speed as a function of the cladding parameters, adaptively adjusting to the process. A similar method was applied for clad height control [21]. In another study, adaptive neuro-fuzzy inference systems (ANFISs) were applied for laser cladding height control, using scanning speed adjustment and CCD feedback as part of an inverse dynamic control approach [22].

Artificial neural networks (ANNs) were also among the first methods investigated. Early work assessed part quality based on scanning speed and laser power [23] and examined the impact of process parameters on the quality of DED-manufactured parts [24]. In a study, optimization of clad geometrical variables was achieved by training a neural network model coupled with particle swarm optimization [25], while logistic regression demonstrated potential for reducing defect areas using experimental data [26]. Additional ANN models, such as those based on Taguchi’s quality loss function, were implemented to predict part quality from scanning speed and feed rate [27]. ANN methods were also applied in industrial settings, including wire arc spraying process control for mold production [28]. Here, spray particle stream data (pressure, current, voltage) were used for ANN-based predictive modeling. Further statistical modeling introduced regression-based hardness predictions from process parameters [29].

In 2015, control of the wire arc spray process using ANNs was again studied [30], building on earlier work [28] but with added consideration of disturbance control in the process.

Overall, these early efforts emphasized basic applications such as process parameter optimization, defect detection, regression modeling, and initial real-time monitoring approaches. They established a foundation for the methodological expansion seen in later years. A categorization of the reviewed papers for the year 2010 is shown in

Table 2. Categorization of reviewed papers based on main topics from 2010 to 2015.

Topic	Papers (Reference Numbers)
Process parameter optimization	[20,21,24,25,27]
Defect detection and classification	[23]
Real-time monitoring and control	[22,28,30]
Regression modeling	[26,29]

. The majority of the studies cover the process parameter optimization.

3.2. Year 2016

So far, ANN and regression models, which are supervised learning methods, had been explored in the literature. However, one of the first papers in the field of unsupervised learning in direct energy deposition was published in 2016 by Urbanic et al., who utilized both classification (supervised) and clustering (unsupervised) techniques in a laser cladding study to categorize beads based on their similarities [31]. In the same year, ANN was applied from a new perspective to predict powder concentration and catchment efficiency, optimizing the powder nozzle shape [32]. A categorization of the reviewed papers for the year 2016 is shown in

Table 3. Categorization of reviewed papers based on main topics for 2016.

Topic	Papers (Reference Numbers)
Defect detection and classification	[31]
Process parameter optimization	[32]

.

3.3. Year 2017

A study on fuzzy-based PID control demonstrated its robust stability in laser cladding compared to conventional control methods [33]. The AI based research in the DED field in 2017 was concluded with a study on the defect detection using logistic regression and ANN tools to identify defects based on experimentally driven acoustic emission statistical data [34]. A categorization of the reviewed papers for the year 2017 is shown in

Table 4. Categorization of reviewed papers based on main topics for 2017.

Topic	Papers (Reference Numbers)
Real-time monitoring and control	[33]
Defect detection and classification	[34]

.

3.4. Year 2018

The concept of a recurrent neural network (RNN) as an active model for a time-series temperature prediction model in DED was proposed by Mozaffar et al. in 2018 [35]. A passive prediction approach, based on an ANN model trained using process parameters, was studied to optimize the fabrication of required geometries in a laser direct metal deposition process modeled by an ANN for Al alloy [36]. Additional studies examined the regression analysis of process parameters on clad geometry, using experimental data from microscopic measurements [37,38]. A categorization of the reviewed papers for the year 2018 is shown in

Table 5. Categorization of reviewed papers based on main topics for 2018.

Topic	Papers (Reference Numbers)
Deep learning applications	[35]
Melt pool and bead geometry prediction	[36]
Surface property and microstructure prediction	[37,38]
Process parameter optimization	[38]

.

3.5. Year 2019

The application of AI to optimize the microstructure of laser cladding for valve sealing surfaces continued in 2019, with a fuzzy predictor model demonstrating remarkable improvement based on experimental results [39]. A deep study on the ML-trained model for creating a map between the thermal gradient, crystal orientation, and Marangoni effect as inputs and the grain boundary tilt angle was investigated by Li et al. [40]. To control cracks in the laser cladding processes, two more studies focused on leveraging ANN in predicting 2D topography [41] and crack density [42] estimation. Our review demonstrates that one of the earliest studies in the field of convolutional neural network (CNN) based on the images acquired from a camera was introduced by [43]. The application of regression models to laser cladding processes was developed to estimate dilution [44], coating properties [45], crystallized sections [46], and the volume of deposited material [47]. A categorization of the reviewed papers for the year 2019 is shown in

Table 6. Categorization of reviewed papers based on main topics for 2019.

Topic	Papers (Reference Numbers)
Process parameter optimization	[39]
Melt pool and bead geometry prediction	[41]
Real-time monitoring and control	[42]
Deep learning applications	[43]
Surface property and microstructure prediction	[40,44,45,46]
Physics-informed ML models	[47]

. As demonstrated, the majority of the studies in 2019 focused on the microstructure prediction models.

3.6. Year 2020

The turning point for ML applications in DED processes was in 2020, during which the number of published papers spiked significantly. CNN exhibited the capability to determine the laser cladding dilution rate based on laser power [48], predicting porosity in laser metal deposition based on thermal data using a physics-driven approach [49], and online monitoring of DED process parameters using time-series image-trained CNNs [50].

The application of the fuzzy comprehension method to experimental data across various process parameters has demonstrated its effectiveness in identifying the optimal conditions for producing a cladding layer with the most favorable surface shape [51], as well as the optimum heat-affected zone and depth of penetration for metal deposition [52].

Machine learning was further leveraged for one-bead curvature predictions using ANN [53], laser quality analysis based on coaxial images using CNN [54], real-time data extraction and machine learning training for online control [55], and classification techniques to predict clad aspect ratios [56]. ANN has been investigated as a robust model in the prediction of dilution on a single track [57]. Nature-based optimization techniques have been implemented on neural networks in training for modeling of hardness in laser cladding processes [58] and the optimization of an ANN-based hardness model to provide process parameters for optimum hardness [59].

The application of porosity recognition for Al alloy in laser metal deposition, based on image-based classification, was introduced by Garcia-Moreno et al. in 2020 [60]. Further regression models using Gaussian process regression models for DEDs have been explored [61,62,63]. Regression models have proven effective in constructing strain rate models for DED deformation processes [64]. In an innovative method, the regression method was studied for dwell-time (interpass time) computation to control the DED processes [65]. The optimization of wire arc and DED processes, both single-objective and multi-objective, using regression models as surface functions has been analyzed in the literature [66,67,68].

The final ML-related study identified for the year 2020 in this review implements reinforcement learning into a laser arm movement framework in multi-track deposition [69].

A categorization of the reviewed papers for the year 2020 is shown in

Table 7. Categorization of reviewed papers based on main topics for 2020.

Topic	Papers (Reference Numbers)
Process parameter optimization	[48,58,59,62]
Deep learning applications	[48,49,50]
Defect detection and classification	[54,60],
Fuzzy logic-based applications	[51,52]
Melt pool and bead geometry prediction	[53,56]
Real-time monitoring and control	[55,65]
Regression modeling	[58,59,61,62,63,64,66,67,68]
Reinforcement learning applications	[69]

. As demonstrated, the majority of the studies in 2020 focused on the regression models and process parameter optimization.

3.7. Year 2021

In this year, fuzzy logic was applied in predicting corrosion [70], quantitative quality [71], and geometrical parameters [72], using fuzzy logic to optimize the laser cladding processes based on the process parameters.

The clustering and classification approaches proposed for detecting porosity in wire-based DED using microscopic images found that random forest is highly effective in segmenting porosity and identifying affected areas [73,74]. The clustering approach based on bead profiles showed potential in determining deposition speed according to the bead similarities to avoid material over-deposition [75].

A study by Wang et al. reviewed the applicability of a deep learning CNN to actively control the layer height of wire arc additive manufacturing (WAAM) [76]. Mohajernia et al. [77] proposed a bead-geometry-based model for 1D, 2D, and 3D bead data, using an ANN to predict stress and hardness across various process parameters in the depth, length, and width of the beads. Barrionuevo et al. [78] applied artificial neural networks (ANNs), support vector machines (SVMs), and boosting regressor models to multi-layer beads to predict geometry height, offering a promising approach for assessing layer penetration in wire arc additive manufacturing (WAAM). Additionally, the application of ML in DED spans predicting and optimizing the surface roughness of beads in WAAM processes [79], intelligent multi-disciplinary bead modeling and optimizing material usage [80], identifying real-time geometric defects (e.g., voids) [81], energy input uncertainty characterization using Monte Carlo [82], data-driven time-series LSTM and extreme gradient boosting (XGBoost) in melt-pool temperature prediction [83], and reducing metal deposition distortion using physics-guided ML using digital twin technology [84]. The advancement of ML in DED has led to the development of novel hybrid modeling approaches for generating data to be used in future ML training and modeling [85].

Further capabilities of ANN models include enhancing structural homogeneity and property consistency across components by adjusting DED parameters [86]; integrating FEM and ANN to predict interpass temperature, thereby reducing thermal defects in WAAM [87]; dynamically adjusting process parameters [88]; applying a multi-feature data fusion model to detect defects in the process [89]; leveraging acoustic emission signal patterns processed by a neural network to distinguish between normal cladding and defect-prone conditions [90]; predicting dilution rates in high-entropy alloy coatings [91]; optimizing dilution rates in Ni coatings [92]; optimizing nozzle structure parameters like angle, outlet width, and radius to control powder flow and quality [93]; predicting residual stress based on FEM-generated thermal history data applied to ANN [94,95]; real-time estimation of melt pool depth [96]; assessing single-bead shape by energy density and deposition rate [97]; and using CNN to track key melt pool aspects such as spatter distribution and area [98].

The application of regression-based models, such as the relationship between specific energy and powder density in FEM simulations [99], a deposition-width model using multiple linear regression [100], an orthogonal polynomial regression model for predicting bead geometry based on process parameters [101], and a model for predicting the shape of the cladding layer [102], are among the regression studies in DED processes.

A categorization of the reviewed papers for the year 2021 is shown in

Table 8. Categorization of reviewed papers based on main topics for 2021.

Topic	Papers (Reference Numbers)
Fuzzy logic-based applications	[70,71,72]
Unsupervised learning and clustering	[75]
Defect detection and classification	[73,74,81,90]
Deep learning applications	[76,85,87,94,95,97,102]
Real-time monitoring and control	[76,80,82,83,88,96,97]
Process parameter optimization	[92,93]
Surface property and microstructure prediction	[73,86,91,92,101]
Regression modeling	[101,102]
Melt pool and bead geometry prediction	[78,83,88,96]

. The studies show that the trend of applications moved towards deep learning models.

3.8. Year 2022

The predictive models in the year 2022 show an extension of approaches to wider applications. The classification approach in 2022 was advanced by applying experimental adaptive data capture for monitoring the melt pool state and the impact on the material quality [103]. The scholastically modeled behavior of powders in the metal laser deposition (MLD) nozzle showed that the deposition wall roughness requires the random collision angle to be taken into account in the MLD process [104]. The studies show the extension of predictive models by generating correlations between the contact tip-to-work piece distance in WAAM processing using feature extraction from acoustic signals and the random forest technique [105]. The novelty of decision making on the quality of laser cladding part quality based on users’ requirements was studied in [106] that applied a fuzzy logic data-driven model by focusing on carbon emission in the process. The application of fuzzy logic in DEDs in the year 2022 includes criteria-defined decision making in WAAM [107], optimizing the process parameters based on an adaptive neural fuzzy interface system (ANFIS) surrogate model [108], application of ANFIS to surface roughness prediction [109], and optimizing part quality based on process parameters utilizing the ANFIS model [110].

Further regression approaches were explored by optimizing thin-plate fabrication, using ANOVA to assess input effects and applying multiple linear regression models for predictions [111]; optimizing bead geometry with a hybrid genetic algorithm and linear regression model [112]; reducing costs and optimizing geometry through process parameters using linear regression models [113]; finding multi-objective optimal solutions from the Pareto front based on application criteria with an SVR predictor model [114]; applying response surface functions extracted from FEM simulations for the multi-objective optimization of laser cladding [115]; conducting sensitivity analysis of geometry with respect to process parameters using correlation methods [116]; performing statistical analysis of thermal data using online monitoring [117]; examining the influence of process parameters on bead width and height [118]; predicting single-track geometry with a hybrid multi-physics and regression model [119]; and applying SVR to predict fatigue life based on pore size and microstructure data during crack initiation [120].

Machine learning methods were used in a wider range of applications in DED processes by applying ANN to predict the deposition quantitatively using ANN [121] or CNN for layer quality [122] or process-parameter optimization [123]. A support vector machine (SVM) was trained by Surovi et al. [124] to predict early defects based on acoustic features derived from printed segments. The decision tree and random forest techniques showed potential in determining defects based on process parameters [125,126]. Defect detection using several ML tools was the focus of studies [127]. The SVM classifier was investigated for bead geometry prediction [128]. A study by Barrionuevo et al. [129] explored three ML methods, Gaussian process regressor, extreme gradient boosting regressor, and ANN, to provide analytical models for predicting melting efficiency. The potential of Gaussian process regression to estimate surface roughness in WAAM processes was also shown [130]. ML methods showed capabilities in the optimization of toolpath planning based on bead geometries [131,132]. Effective heat control was identified as a crucial factor for optimizing WAAM processes [133,134,135]. Similarly, accurately determining radiation intensity and analyzing melt spots and their impact on part quality using machine learning techniques was shown to be essential [136]. The microstructure of laser cladding was a subject for ML methods in predicting microhardness [137], induced porosity [138], grain structure [139], and morphological anomaly detection [140].

In a 2022 review article, Chadha et al. [141] showcased AI’s ability to minimize human intervention, thereby enhancing the production of DED-fabricated parts.

The application of neural networks in laser cladding processes involves passive training, where a large, time-independent dataset with predefined inputs and outputs is available [142,143,144,145,146,147,148,149]. These models have served as response surface functions for optimization purposes [150,151,152]. However, the computation-intensive training of large datasets in the DED field has shifted towards physics-informed neural networks (PINNs), where governing partial differential equations (PDEs) are incorporated into the training process. This approach allows the network’s loss model to be based on physics and adhere to physical constraints [153,154]. Nevertheless, time-dependent data have favored the use of CNN models [155,156,157]. Yet, passive back-propagation approaches for real-time melting pool temperature prediction have been proposed [158,159].

A broader range of data from in situ signal monitoring, including melt pool dynamics, crack propagation, and pore formation, was utilized in deep learning models to develop acoustic-based defect detection for the DED process [160,161]. Deep learning was also utilized in learning abnormal powder feeding trained based on melt zone images [162]. Probabilistic deep learning models served as tools in determining uncertainties in the melt pool depths [163,164]. Deep learning U-net networks showed potential in identifying dendrites based on laser cladding layers [165] [in the Chinese language].

A categorization of the reviewed papers for the year 2022 is shown in

Table 9. Categorization of reviewed papers based on main topics for 2022.

Topic	Papers (Reference Numbers)
Process parameter optimization	[111,113,114,115,123,132]
Real-time monitoring and control	[103,105,117,155,160,161]
Defect detection and classification	[120,124,125,126,127,137,138,140,160,161]
Melt pool and bead geometry prediction	[112,118,128,136,156,159]
Surface property and microstructure prediction	[104,137,138,139,143]
Deep learning applications	[122,135,155,156,157,159,160,161,162,163,164,165]
Fuzzy logic-based applications	[106,107,108,109,110]
Regression modeling	[111,112,113,114,115,116,118,119,120,129,130]
Physics-informed ML models	[153,154]

. The studies show that the trend in applications moved towards deep learning models. Deep learning, regression predictions and real-time monitoring made up the majority of the studies.

3.9. Year 2023

Du et al. [166] conducted a study that combines statistical analysis to identify the most influential process parameter (scanning speed) with a fuzzy inference-based decision-making system and gray correlation analysis to optimize the fabrication process for multiple objectives, including clad hardness, thickness, dilution percentage, and surface roughness. Xv et al. [167] concluded a similar analysis by showing the scanning speed has a higher impact on the coat thickness than the feed rate, while the overlap rate exhibited the highest influence on the dilution rate for high-scanning-speed laser cladding coating processes on an Fe-Cr-Ni-based alloy. The analytic hierarchy process and fuzzy comprehensive evaluation method suggested the optimum process parameter quantitative values.

In addition to decision-making approaches, regression models have remained a focal point of research interest. Tendere et al. [168] applied the analysis of variance regression method to consider the optimum microhardness and dilution rate based on the process parameters. A new approach of real-time melt pool image processing to provide predictor models within a range of process parameters and melt pool dimensional data was employed by Lyu et al. [169], concluding that the Gaussian process regression model was the best-performing model. Support vector regression (SRV) has shown potential in making surface functions to predict the properties of specimens in direct energy deposition based on the process parameters, providing a rapid prediction to reduce porosity [170]. The application of regression models was extended to twin-wire-arc-sprayed coatings based on different compositions of metal wires for optimizing microstructure properties [171]. Bohrani et al. [172] applied linear regression and the response surface method (RSM) to optimize the laser cladding process for Inconel 625, a widely used nickel-based superalloy. A review article by Wang et al. in 2023 [173] explored various optimization techniques for laser cladding, including Taguchi methods, the response surface methodology (RSM), and machine learning-based approaches, indicating the necessity of intelligent methods to optimize the laser cladding DED processes. An integration of numerical simulations with the response surface method (RSM) to predict the geometrical characteristics of laser-cladded coatings demonstrated that RSM-based predictive models can reduce experimental costs while ensuring high accuracy in clad layer formation and property prediction [174]. Design of experiment and regression modeling were implemented on laser metal deposition (LMD) processes of NiCoCr alloys (VDM alloy 780) to reduce the residual stresses [175]. The statistical design of experiment (DOE) and ANOVA based on infrared thermography showed a correlation of thermal variation effects with potential defects [176].

AI applications in WAAM processes have advanced defect detection, real-time feedback generation, and the development of innovative control strategies, such as reinforcement learning, to address challenges related to process non-linearity and uncertainties [177]. The applications have shown the feasibility of extension to pre-design, and online and offline control optimization [178]. Machine learning classification based on optical emission spectroscopy (OES) sensors demonstrates potential to optimize the quality and integrity of functionally graded materials (FGMs), reducing material waste and improving microstructural consistency [179]. Clustering methods have identified distinct clusters corresponding to specific stress and distortion behaviors and patterns within a range of different geometries [180].

A comparison of regression-based neural networks and random forest classification was studied for predicting mechanical microstructure properties based on process parameters, showing that the random forest model outperformed deep learning NN in accuracy [181]. The effectiveness of porosity detection in wire laser metal deposition (WLMD) by the data-driven ANN [182] and thermal modeling by CNN [183] were explored. The study demonstrated successful knowledge transfer at various levels, including data representation, model architecture, and model parameters based on the data-driven model adapted from LPBF to learn about DED processes [184].

Deep learning was applied to WAAM copper surface oxidation defect online prediction by extending the one-dimensional sensor voltage to 2D time-wavelet data combined with deep learning, where the abnormalities of voltages indicate bead defects [185]. A study on WAAM optimization leveraged convolutional Bi-directional Long Short-Term Memory to apply sensor-extracted spatial feature data into convolutional learning while using BiLSTM’s capability to capture temporal correlations [186]. The improved model shows promising accuracy in the control of melt pool temperatures. The control of the dilution rate on the multi-pass laser cladding process using CNN deep learning based on experimentally calibrated FEM generated input data. Then, the model was implemented in optimization [157].

Further studies encompass applications of regression models for classification of experimental data for bead geometry prediction [187]; ANN models to predict bead geometry based on the process parameters [188,189,190,191,192,193,194]; bead geometry feature reduction and real-time incremental learning to predict the bead height and width based on electric sensor data [195], which reduces computational costs compared to CNNs; tensile deformation predictor [196,197]; merging sensor data and imaging for enhanced detect classification [198]; real-time neural sensor for nozzle clogging defect [199]; clad quality prediction based on multiple materials with several learning methods [200], suggesting that LSTM outperformed ANN and XGBoost; porosity defect classification [201]; bead height consistent deposition using imaged-based CNN and SVR in-process hybrid approach [202]; fatigue life prediction using machine learning based on microstructure plasticity data [203]; enhancing crystallographic analysis in metal AM using conditional generative adversarial networks (c-GANs) to reduce noise in electron backscatter diffraction (EBSD) images [204]; finding correlations between composition and microstructure based on ANN models [205]; physics-informed temperature prediction [206]; process-defect modeling without compromising proprietary design information [207]; quality prediction [208]; residual stress prediction [9,209]; melt-pool dynamics prediction based on gas–powder interaction [210]. By 2023, the challenges of applying AI models from different perspectives in WAAM and DED were reviewed by [18,211].

Additional studies in 2023 applied artificial neural networks in laser metal additive manufacturing. Soloviov et al. focused on utilizing neural networks for real-time monitoring and control of flux-cored wire arc surfacing, aiming to improve process stability and defect reduction [212]. Cui et al. [213] applied a Sparrow Search Algorithm (SSA) optimized back-propagation (BP) neural network to predict the cross-sectional morphology of laser-cladded layers, helping optimize process parameters. Gao et al. [214] combined neural networks with genetic algorithms to achieve multi-objective optimization of NiTi shape memory alloys fabricated through laser metal deposition (LMD), enhancing material performance. Dong et al. [215,216] introduced multi-modal CNNs for predicting LMD height and cross-section geometry, utilizing multi-sensor data fusion to enhance precision. Li et al. [217] proposed a physics-informed neural network (PINN) framework that predicts 3D temperature fields in LMD without requiring labeled data, enabling real-time process control. Perani et al. [218] integrated on-line artificial vision systems with deep neural networks to predict track geometry in LMD, significantly improving process automation and quality assurance. Diehl et al. [219] developed a physics-guided neural network to address imbalanced datasets in predicting melt pool width variations, improving the accuracy of process modeling. Yang et al. presented two studies: one using laser line scanners, vision cameras, and artificial neural networks for real-time corner height estimation in multi-layer DED [220]; and another employing domain-adaptive neural networks for layer height estimation, enhancing process control [221]. Chen et al. [222] introduced a blurry inpainting network to assess surface quality in curved DED parts, aiding in automated quality assurance. Patil et al. [223] proposed CNN deep learning frameworks for defect classification and detection in laser DED, reducing manual inspection needs. Chen et al. [224] applied multi-modal sensor fusion to enable real-time defect detection, providing location-specific insights for process optimization, using a hybrid CNN model. Kong et al. [225] developed a melt-pool monitoring system based on irregularity degree analysis and a CNN model, which enhances defect diagnosis during DED fabrication.

Additional studies explored optimization, path planning, and real-time monitoring in laser-based DED and WAAM through machine learning, reinforcement learning, and deep learning techniques. Liu et al. [226] employed PCA and grey relational analysis (GRA) for multi-objective optimization of Ni-based laser cladding, improving material properties. Petrik et al. [227] used reinforcement learning for path planning in thin-walled WAAM structures, demonstrating an adaptive approach to improve deposition accuracy. Shi et al. [228] extended reinforcement learning applications by integrating deep reinforcement learning to optimize laser-DED parameters, improving process efficiency and quality. Wu et al. [229] proposed a semi-supervised multi-label feature selection algorithm for online monitoring of laser metal deposition quality using multi-regression and feature reduction approaches to facilitate real-time defect detection. Lastly, Pandiyan et al. [230] integrated coaxial imaging with a self-supervised deep learning framework for real-time monitoring and quality assurance in laser-based DED using a CNN model.

A categorization of the reviewed papers for the year 2023 is shown in

Table 10. Categorization of reviewed papers in 2023 based on main topics.

Topic	Papers (Reference Numbers)
Process parameter optimization	[166,167,170,171,173,188,214,226,227,228,230]
Real-time monitoring and control	[169,177,179,195,220,221,229,230]
Defect detection and classification	[182,185,198,201,207,208,222,223,224,225]
Melt pool and bead geometry prediction	[157,186,187,191,195,210,225]
Surface property and microstructure prediction	[181,196,197,205,209]
Deep learning applications	[157,186,202,204,212,213,215,216,218,219,221,222,223,224,225,230]
Fuzzy logic-based applications	[166,167]
Reinforcement learning applications	[227,228]
Regression modeling	[168,172,174,175,176,181,184,189,190,192,199,200,203]
Physics-informed ML models	[206,217]
Unsupervised learning and clustering	[180]

. There was a notable surge in interest toward defect detection methods in 2023.

3.10. Year 2024

Some review studies in 2024 explored the integration of AI and hybrid materials in WAAM, laser cladding, and DED, focusing on process optimization, automation, and sustainability advancements in additive manufacturing [231,232].

Regarding clustering, a few papers were introduced in 2024. A study by Pazireh et al. [233] utilized linear mixed-effects (LME) models, principal component analysis (PCA), and self-organizing map (SOM) clustering to analyze and provide insights into the mechanical properties influenced by toolpath patterns, geometry types, and layering effects. Real-time control of laser power, guided by accurate extraction and clustering of molten pool geometric features [234], outlier detection [235], and defect prediction using melt images and clustering model [236] were among the rest of the clustering model implementations in DEDs.

The laser cladding and wire arc approaches utilized fuzzy logic methods. With a novel approach, a fuzzy quality function was studied to relate the cost/property quality with several manufacturing strategies to find out the crisp values of the strategies on the qualities [237]. Among these, further fuzzy applications found in the literature are mechanical property prediction [238], surface roughness prediction [239], and implementation of an adaptive fuzzy controller for the morphology of powder laser metal deposition [240].

Two papers are found to have used unsupervised learning in DED fabrication. PCA, a feature reduction approach, was applied over acoustic and current signals to explore the dominant features in porosity detection [241]. The practicality of frequency-domain analysis in evaluating the quality of the WAAM process via unsupervised learning techniques for different processes and the applicability for different materials was considered by wavelet data analysis [242].

Data-driven methods found a broader range of applications in the WAAM and DED AM processes. The applications comprise process variability reduction in WAAM [243], using Deep Deterministic Policy Gradient Controller to enhance the WAAM control process [244] (corrected publication), and generating a data-driven ANN model for temperature and distortion based on FEM simulations [245].

The application of deep learning was explored in various DED manufacturing areas. The fluctuation of laser current and speed on the reliability of the manufacturing process was considered using a stochastic approach and applying Monte Carlo to study the uncertainty in WAAM processes [246]. The Delphi method suggested applying AI tools in WAAM based on melting mark identification, showing potential for metallurgical improvement [247]. The extended study reviewed defect feature definition and implementation in CNN models [248], extracting deposition height features from coaxial molten pool images and applying a CNN model for process stability [249], temperature field prediction using RNN and controlling the process [250], analyzing melting morphology of melting zone for real-time control [251], and applying FixConvNeXt model classification for defect classification through improved image recognition [252]. A deep learning model was studied to capture real-time microstructure properties to address weak-bonding issues [253].

The majority of papers with the keyword “machine learning” in DED studies in 2024 focus on five main categories in which machine learning models have been utilized in DED research. (1) Optimizing process parameters: melt pool data-based ANN [254], thin-wall geometry data-based ANN [255], global energy distribution constituent data [256], catchment efficiency [257], bead cross-section area prediction [258], optimized layer deposition [259], multi-objective process optimization [260], weld geometric prediction [261], and energy efficiency considerations [262]. (2) Real-time monitoring and anomaly detection: surface tension transfer method for anomaly detection [263], methodologies of defect detection [264], in situ monitoring for machine learning training in defect detection [265,266,267,268], IoT-driven intelligent system for WAAM [269], gas-shield flow rate model for defect avoidance [270], microstructure composition prediction [271], time-series dimensional error prediction [272], and porosity knowledge gain [273]. (3) Melt pool and bead geometry: bead geometry prediction based on process parameters [274,275,276,277,278], surface roughness prediction [279], melt pool prediction [280,281,282], and functionally graded material optimum design for cost reduction [283]. (4) Wear and surface property prediction: wear prediction and surface roughness improvement [284,285,286,287], microstructural and mechanical prediction and optimization [288], corrosion study [289], fatigue failure prediction [290], crack prediction [291], clad and geometric characteristic prediction [292,293], interlayer defect identification [294], and predicting attenuation of electromagnetic interference for electromagnetic shielding [295]. (5) Process simulations and hybrid physics-informed models: transient thermal and thermo-fluid physics-informed simulations [296,297,298], sulfur content effect on the DED process [299], optimal process exploration [300], process instability detection based on physics-informed model [301], predicting composite alloys [302], and developing simulation models based on digitalization methods [303].

Regarding the keywords “artificial neural networks”, there are seventeen published research articles found for 2024. The ANNs comprise of monitoring melt pools [304], tensile property predictions [305], a microstructure study and dimension predictions [306], metal wire parameter optimization [307], CNN-based classification for acoustic emission data [308], camera-captured image data classification [309], computer-vision ANN architectures for defect detection [310], a feed-forward back-propagation multi-layer perceptron ANN model based on temperature history input data for residual stress prediction using FEM models [311], CNN-based models applied on in situ image data for predicting temperature [312], and geometry accuracy [313]. Further ANN studies have explored depth prediction using group-incorporating ANN [314], RNN and LSTM models for melt-pool geometry predictions [315], a CNN approach for segmented melt pool zones [316], time–frequency acoustic wave data for detect identification [317], and an MLP model for predicting bead geometry as a virtual sensor [318].

Other types of regression models were applied to DED and WAAM processes for regression purposes. For the sake of optimization the following were used: a surface response model and post-treatment approach [319], material change comparison [320], optimized bead width to minimize the defects [321], combination of modeling and experimental data for regression and optimization [322], and mixing the melt pool data of different materials for Gaussian regression models to improve predictions [323]. The second category of regression models comprises predictive modelings for quality estimation. A coat quality analysis based on two random forest (RF) and adaptive boosting (AB) illustrated that of AB outperformed RF, as AB shows high sensitivity to anomalies [324]. A hybrid Least Squares Support Vector Regression (LSSVR) and an optimization approach improved potential in predicting clad geometry [325]. The influence of a Stellite 6 coating on SS316L based on an experimental-based second-order regression model was studied by [326]. Linear regression models for predicting clad morphological characteristics (e.g., width, melt area, penetration rate, etc.) from process parameters were considered by [327]. Weld quality prediction from the substrate temperature and the process parameters was investigated using linear regression models [328]. A paper in 2024 showed the study of a regression model for predicting residual stresses using the process parameters [329]. Regression-based models have also been implemented for hybrid manufacturing and repairing processes [330,331].

The exploration of ML application in DED and WAAM in 2024 demonstrates the use of reinforcement learning in intelligent process parameter optimization [332] and laser power planning [333].

A categorization of the reviewed papers for the year 2024 is shown in

Table 11. Categorization of AI/ML papers in DED and WAAM for 2024.

Topic	Papers (Reference Numbers)
Process parameter optimization	[260,261,288,319,320,321,322,330,332,334,335,336,337,338,339,340]
Real-time monitoring and control	[234,235,236,265,266,267,268,269,270,271,273,301,309]
Defect detection and classification	[236,252,264,265,266,267,268,270,273,308,309,310]
Melt pool and bead geometry prediction	[258,259,274,275,276,277,279,280,281,282,283,304,314,318]
Surface property and microstructure prediction	[286,287,289,290,291,292,293,294,295,305,306]
Deep learning applications	[246,247,248,249,250,251,252,253,297,308,309,310,312,313,314,315,316,317]
Fuzzy logic-based applications	[237,238,239,240]
Reinforcement learning applications	[244,332,333]
Regression modeling	[319,320,321,322,323,324,325,326,327,328,329,330,331]
Physics-informed ML models	[282,296,298,299,301]
Unsupervised learning and clustering	[233,236,241,242]

. The 2024 trend demonstrates a significant increase in both the number and diversity of studies, highlighting a shift from basic prediction models toward more advanced, integrated approaches such as real-time control, physics-informed simulations, and reinforcement learning.

3.11. Year 2025

The review was conducted at the end of July 2025, and the papers included were published up to that date.

In 2025, data-driven approaches for DED processes encompassed multi-objective optimization applied to regression models [334], fast scanning pattern selection using deep regression models [341], wire thickness adjustment based on arc voltage drop [342], and improving bead uniformity and deposition efficiency in 3D printing 17-4PH stainless steel in WAAM with ANN and genetic algorithm [343].

Two unsupervised studies were published. Mattera et al. [344] decomposed welding signal waves into distinct frequency components and applied clustering techniques to categorize Inconel 718 thin-walled structures. Li et al. [345] utilized video-captured data to detect anomalies by employing a Vector Quantization Variational Convolutional Autoencoder (VQ-VCAE) model, which was trained on defect-free data, and used reconstruction errors during online fabrication to identify quality defects.

A fuzzy approach was found among the published research for 2025. Abdulrazaq et al. assessed the fuzzy logic in optimizing the head angle and position when the head of the deposition head was not horizontal using rule-based fuzzy logic [346].

In the context of deep learning, real-time measurement using a deep neural network to find bead dimensions using online data [347] in the WAAM process, online pore detection under varying light/image conditions [348], hump defect detection based on melt data using CNN models [349], morphology prediction based on melt zone data based on a U-Shaped Network [350,351], light strip image-based defect identification using a YOLOv5 neural network [351], sensor-based temperature data for temperature prediction using multi-modal deep learning [352], automatic recognition and segmentation of defects [353], and PINN-based stress prediction based on thermo-mechanical simulation data [354] were introduced.

The wider applications of machine learning have been applied to geometry predictions [335,355,356,357], melt pool modeling using ML [358,359,360,361], defect/anomaly detection [362,363,364,365], process optimization [336,366,367,368,369], residual stress and cracking mitigation [370,371], and hardness predictions [372,373,374]. Ref. [374] showed that the uncertainty arising from thermal conditions in the build process needs to be quantified to address the tensile and porosity challenges. A study developed random forest, decision tree, and XGBoost machine learning methods to accurately and efficiently predict the friction of special alloy coatings in the laser cladding process, avoiding costly experiments and helping improve coating design [369].

The further in-depth contributions cover temporal data RNN-based models [375], temporal melt zone image-based CNN models [376,377], a back-propagation optimized model [337], a physics-informed fatigue life prediction model [378], and a physics-informed neural network to predict temperature to control microstructure quality [379]. Bayat et al. [360] have shown that gated recurrent unit (GRU) models are superior in terms of prediction accuracy over 1D-CNNs, dense neural networks, and RNNs.

Non-deep learning regression models were used to consider melt geometry predictions using Ant Colony Optimization (ACO) using the SVR regression model [380], process parameter optimization [338,339,340], statistical analysis of variability [381], weld morphology prediction [382], strength characteristic prediction in WAAM [383], laser cladding quality characteristic prediction [384], and the control of scan speed and laser power [385].

A semi-supervised ML was introduced to use a regression model (supervised) for the melt pool while an unsupervised learning was leveraged to gain more knowledge of the melt zone features [386].

A categorization of the reviewed papers for the year 2025 is shown in

Table 12. Categorization of AI/ML papers in DED and WAAM for the year 2025.

Topic	Papers (Reference Numbers)
Process parameter optimization	[334,338,339,340,341,342,343,368,369]
Real-time monitoring and control	[347,348,349,350,351,352,385]
Defect detection and classification	[345,348,349,351,353,362,363,364,365,376,377,378]
Melt pool and bead geometry prediction	[335,337,355,356,357,358,359,360,361,375,380]
Surface property and microstructure prediction	[369,370,371,372,373,374,378,383,384]
Deep learning applications	[337,347,348,349,350,351,352,360,375,376,377]
Fuzzy logic-based applications	[346]
Regression modeling	[334,338,339,340,341,366,380,381,382,385,386]
Physics-informed ML models	[354,359,378,379]
Unsupervised learning and clustering	[344,345]

. Compared to 2024, the 2025 studies show a continued diversification of AI/ML approaches in DED and WAAM, with deeper integration of unsupervised learning, semi-supervised models, and advanced temporal architectures such as RNNs and GRUs, marking a shift toward real-time, adaptive, and hybrid intelligence systems for in situ monitoring, defect mitigation, and predictive modeling.

3.12. Categorization of Reviewed Papers

To provide a clearer overview of the research focus areas identified in the reviewed literature,

Table 13. Categorization of reviewed papers based on main topics for 2010–July 2025.

Topic	Papers (Reference Numbers)
Process parameter optimization	[24,25,27,32,38,39,48,51,52,58,59,62,66,67,93,111,113,114,115,116,132,150,151,166,167,170,171,173,188,214,226,228,259,260,261,288,319,320,321,322,330,332,334,335,336,337,338,339,340,343,368,369]
Real-time monitoring and control	[22,28,30,33,42,55,65,76,80,84,89,98,103,105,117,121,159,161,169,176,177,179,195,220,221,229,230,234,240,251,252,266,267,270,271,301,303,309,342,363,385]
Defect detection and classification	[23,31,34,54,60,63,74,81,92,120,124,125,126,127,136,140,155,182,185,198,201,207,208,222,223,224,235,241,242,243,263,264,265,268,269,273,290,291,296,348,349,351,353,362,364,365,370,378]
Melt pool and bead geometry prediction	[36,41,53,56,72,82,83,88,96,100,112,118,128,131,134,138,158,186,187,191,210,225,254,274,275,276,277,279,280,281,323,327,328,350,356,357,358,360,361,380]
Surface property and microstructure prediction	[37,40,44,45,46,64,73,79,86,101,104,123,130,137,143,181,196,197,205,209,256,286,287,292,302,305,306,307,369,371,372,373,374,381,383,384]
Deep learning applications	[35,43,49,50,57,77,85,87,94,95,97,102,122,135,142,144,145,146,147,148,149,152,157,162,163,164,165,183,193,194,202,204,212,213,215,216,218,219,245,246,247,248,249,250,253,255,304,308,310,311,312,313,314,315,316,347,352,355,367,375,376,377]
Fuzzy logic-based applications	[20,21,70,71,106,107,108,109,110,237,238,239,346]
Reinforcement learning applications	[69,227,244,333]
Regression modeling	[26,29,61,68,78,99,119,129,133,141,156,168,172,174,175,178,184,189,190,192,199,200,203,211,231,232,257,258,262,272,283,284,285,289,293,297,300,317,318,324,325,326,329,331,341,366,382,385,386]
Physics-informed ML models	[47,139,153,154,206,217,282,298,299,354,359,379]
Unsupervised learning and clustering	[75,180,233,236,344,345]

categorizes the referenced papers according to their main application topics, including process parameter optimization, real-time monitoring, defect detection, melt pool prediction, surface property prediction, and various machine learning approaches such as deep learning, fuzzy logic, reinforcement learning, physics-informed modeling, and unsupervised learning.

4. Challenges

As DED technology advances, processes involving multi-material deposition and freeform surface fabrication are becoming increasingly common. Figure 6 illustrates two representative examples: one demonstrating separate toolpaths for distinct materials within a composite block, and another showing selective material transitions across layers in a curved geometry. These applications present significant challenges in both planning and control, specifically in complex geometries, thin walls, and metal composites (different materials within a layer and between layers), and so forth. Variations in geometry, thermal accumulation, and material properties must be considered simultaneously, making traditional process modeling approaches insufficient. Moreover, such cases demand location-aware control strategies and adaptable machine learning models capable of handling spatial, material, and temporal complexity. These scenarios exemplify the broader challenges discussed in this section.

One significant challenge in applying machine learning to DED is capturing the effects of geometric transients and in-process variations that arise solely from changes in toolpath strategies, even when all other process parameters remain constant. As illustrated in Figure 7, based on experimental work conducted in the authors’ research lab, substantial variations in microhardness are observed across different build strategies (5-axis, 3+2 axis, and rotary paths). These differences reflect complex thermal histories and solidification behaviors that are difficult to encode into conventional ML features. The working images underscore the challenge of capturing such localized effects, which are further compounded by the absence of standardized geometric representations for multi-axis DED builds.

Another key obstacle in developing robust ML models for DED is the presence of low-frequency geometric transients that subtly influence melt pool behavior and deposition characteristics over time. Figure 8, based on experimental work conducted in the authors’ research lab, illustrates how minor variations in power input, despite occurring at relatively low frequency, can result in significant fluctuations in bead width, an essential quality indicator. These transients are typically under-represented in standard feature engineering and can degrade ML model accuracy if not explicitly captured. Their subtle but cumulative effects highlight the need for temporal context-aware models capable of recognizing long-duration spatial or power perturbations in multi-track builds.

Recent applications of temporal machine learning models such as RNNs, GRUs, and CNNs for sequential data analysis in DED have revealed unique challenges. Studies have reported issues such as overfitting in short-sequence data, training instability, and limited generalizability across different geometries and build rates [9,317,360,375,376]. These difficulties emphasize the temporal complexity that ML frameworks must overcome to be effective in dynamic DED environments.

Based on this review, several key challenges with applying machine learning to DED and WAAM processes are identified. Research focused primarily on experimental studies or traditional modeling approaches prior to 2020, with little use of machine learning. The number of publications began to rise sharply in 2020, yet many studies used general terms such as “machine learning” without specifying particular methodologies like regression, support vector machines, or artificial neural networks. The emergence of advanced topics such as deep learning and fuzzy logic was only notable after 2021, indicating that these approaches are still in their early stages of exploration. A significantly greater amount of attention was paid to process parameter optimization and melt pool geometry prediction than to real-time control, defect classification, or physics-based machine learning. It is evident from these observations that while machine learning applications have grown rapidly in DED, a deeper methodological analysis and higher diversity of research topics are needed in order to strengthen the field.

There are still several gaps in the field, despite these developments: real-time closed-loop control remains underexplored; physics-based models have yet to gain traction; defect classification models lack interpretability; most studies focus on predicting single outcomes rather than integrating multiple outputs; and generalization across different machines, processes, and materials remains limited.

In addition to these challenges, another gap identified is the limited consideration of deposition location and its associated mechanical effects in existing machine learning models. Most current studies emphasize thermal history, melt pool size, or other thermally driven parameters, often using point cloud data or temperature field tracking. However, the mechanical conditions such as stress accumulation, strain localization, and distortion are strongly influenced by the specific location of deposition, including edges, corners, and central regions of parts. Studies have shown that microstructure and mechanical properties vary significantly depending on deposition position within a part [389,390]. This spatial variation plays a critical role in mechanical anisotropy and part performance but remains poorly integrated into ML frameworks. These factors play a crucial role in residual stress development and final part quality, yet they are seldom incorporated into machine learning frameworks. Addressing coupled thermo-mechanical behavior, particularly with location-specific modeling, represents a key direction for future research.

To summarize, several practical challenges emerge from the application of ML techniques in DED and WAAM environments:

• Geometric complexity and multi-material deposition: Freeform geometries, thin walls, and layer-wise material transitions introduce significant spatial and material complexity. These factors demand adaptive path planning and location-aware ML models that most current approaches do not accommodate.
• Lack of standardized geometric representations: Multi-axis toolpaths and curved geometries lack consistent representations, making it difficult to encode spatial features for ML training, especially for microstructure prediction and property mapping.
• Temporal and in-process transients: As shown in Figure 7 and Figure 8, subtle power fluctuations and changes in toolpaths result in dynamic variations (e.g., bead width, microhardness) that static feature models fail to capture. Temporal models like RNNs and CNNs struggle with short sequences, overfitting, and generalization.
• Spatially varying mechanical effects: Location-specific phenomena, such as stress accumulation, strain localization, and distortion, are often excluded from ML models. However, mechanical effects vary significantly based on deposition position (edges, corners, centers), affecting anisotropy and final part quality.
• High computational cost of training data: Generating labeled datasets using FEM simulations with thermal–mechanical coupling is time-consuming and resource-intensive, restricting the size and diversity of training datasets.
• Limited experimental data and labeling challenges: Acquiring high-quality labels (e.g., microstructure or stress maps) is experimentally challenging. As a result, unsupervised or semi-supervised methods are often employed, though they lack interpretability and robustness.
• Lack of consensus on ML techniques: A wide range of methods, including regression, neural networks, clustering, fuzzy logic, and deep learning, have been explored without a clear consensus. Few studies compare different ML methods under identical conditions to determine best practices.
• Underexplored research areas: Key applications such as real-time closed-loop control, defect classification, and multi-objective predictions (e.g., combining thermal and mechanical outcomes) are still in their infancy and deserve more focused investigation.

5. Future Directions

Based on the review findings presented in this study, several topics of interest are identified for future research. One critical area is the influence of toolpath and deposition location history on the thermo-mechanical behavior of fabricated parts, which remains a fundamental research gap. Most existing ML-based control strategies rely on open-loop approaches driven by in situ data capture; however, the development of robust closed-loop control systems through the integration of machine learning and optimization techniques warrants further exploration. To address the identified gap in real-time closed-loop control, future research should explore system architectures that integrate in situ sensor feedback, machine learning models, and adaptive process control. For instance, real-time melt pool imaging data can be processed using CNNs to detect geometric or thermal anomalies [376], while acoustic emission signals may inform defect formation patterns through neural network classification [90,265]. A technically feasible framework would involve sensor streams feeding into trained ML models (e.g., LSTM for time-series prediction or CNNs for spatial detection), which in turn adjust parameters such as laser power or feed rate through a feedback controller. Developing such closed-loop systems requires synchronized hardware–software integration, fast inference capabilities, and robust model generalization under variable conditions that present both opportunities and challenges for future research.

Future studies should also emphasize the incorporation of physics-based constraints into ML and AI models, combined with data-driven approaches, to better address mechanical limitations. PIML models offer a promising approach by embedding governing physical equations directly into the learning framework, thereby reducing the need for computationally expensive numerical solvers. These models maintain consistency with known physical laws while enabling rapid predictions. When combined with experimental data, PIML approaches enhance model robustness and accuracy by leveraging the high fidelity of simulations alongside the real-world variability captured in experiments. This integration allows for efficient, scalable modeling that preserves both physical interpretability and data-driven prediction models for complex DED/WAAM processes in closed-loop control systems.

Additionally, there is a noticeable gap in the application of probabilistic and uncertainty quantification methods, which are essential to provide reliable and interpretable predictive outcomes. Similarly, future research should focus on developing robust temporal ML architectures that can handle varying sequence lengths and dynamic boundary conditions. Data augmentation and hybrid modeling strategies may help overcome the overfitting and instability issues noted in recent studies [9].

A summary for future machine learning integration and research prioritization in DED/WAAM is presented in

Table 14. A summary for future machine learning integration and research prioritization in DED/WAAM.

	Performance	Functionality	Design and Validation Considerations
System	Durability	Data collection and analytics (real-time metrics)	Low-to-medium-volume systems design
	Repeatability	Production planning and scheduling and adaptation to changes	Discrete event-/agent-based simulation tools (incl. heat management)
	Cycle time/throughput
	Stability
	Task complexity, line balancing, etc.
Process	Parameter optimization	Component variety, shape, surface, overhangs	Real-time process–part data collection
	Tool path optimization	Repair, new features, etc.	DED-AM specific performance metrics
	Durability, repeatability	Functionally graded materials (intra-/interlayer)
	Real-time monitoring, control
	Thermo-mechanical behavior	Integrated machining
Prediction Tools	Adaptable, extensible, rapid	Quasi-static strength and impact response	High-fidelity prediction tools
	Geometry and mechanical properties	Induced residual stresses	Adaptive control
	Real-time validation	Wear characteristics	DED-AM specific in-process metrics
	Automatic parameter adjustment	Bead, layer, and shape geometry
Design Tools	Design for DED-AM CAD and CAM	Data-driven design optimization (component + operations)	Automated geometry adjustments (e.g., topology optimization)
	Modified materials datasets
	ML-supported design tools
Personnel	Training: product and process design	Training: tool paths, planning	Inspection and analysis tools
	DED system interaction types	Mechanical characterization
	Calibration strategies

, which presents a structured framework capturing key research directions and considerations across system-level, process-level, predictive, and design-focused domains in DED and WAAM.

A roadmap, illustrated in Figure 9, summarizes the key areas where machine learning and artificial intelligence can support and enhance DED processes. It categorizes opportunities across five levels—system, process, prediction tools, design tools, and personnel—highlighting essential functions such as real-time monitoring, adaptive control, data-driven design, and training needs.

6. Conclusions

This review presented a comprehensive trend study of machine learning (ML) and artificial intelligence (AI) applications in direct energy deposition (DED) and wire arc additive manufacturing (WAAM) technologies. By employing an automated Python-based retrieval approach using the Crossref API, over 370 relevant publications were identified from 2010 to July 2025. The analysis reveals a clear acceleration in ML-related DED research beginning in 2020, with substantial growth in topics such as deep learning, fuzzy logic, reinforcement learning, and hybrid physics-informed models.

The literature was categorized into major themes, including process optimization, melt pool geometry prediction, defect detection, and real-time monitoring. While recent advances highlight the growing maturity of data-driven DED solutions, this review also identified existent challenges. These include the lack of real-time closed-loop control systems, minimal integration of mechanical stress/strain modeling, limited generalizability across different DED platforms, and insufficient use of uncertainty quantification to address the challenges of complex geometries, thin walls, freeform surfaces, and composite materials. These underscore the need for location-aware and thermo-mechanically coupled ML frameworks. This review provides a foundation for future research in intelligent additive manufacturing and highlights pathways toward robust, adaptive, and interpretable AI-driven solutions for DED and WAAM systems.