A review published on ResearchGate by researchers at the University of Windsor analyzes how machine learning (ML) is being applied to direct energy deposition (DED) and wire arc additive manufacturing (WAAM). Covering studies published between 2010 and mid-2025, the review shows that research activity expanded rapidly after 2020, with methods such as deep learning, fuzzy logic, and physics-informed neural networks moving from isolated experiments to mainstream topics. Despite this progress, gaps remain in closed-loop process control, generalization across machines, and integration of location-specific mechanical effects.
The authors developed a Python script to automate searches through the Crossref bibliographic database. Publications were retrieved using combined keyword sets covering both DED processes and machine learning techniques, and duplicates were removed. Each entry was manually checked to ensure the study applied ML specifically to DED or WAAM. While this method captured a broad cross-section of the field, the dataset is limited to Crossref-indexed content, excluding proprietary databases such as Scopus or Web of Science. This introduces a degree of bias but still provides a representative overview of how artificial intelligence has entered this branch of metal 3D printing.
From early foundations to turning point
Research in the first half of the 2010s was sparse and exploratory. Initial studies tested fuzzy logic models to regulate scanning speed in laser cladding and applied simple neural networks to predict clad quality or optimize bead geometry. These projects showed the feasibility of data-driven methods but remained restricted to narrow parameter optimization tasks.
By 2016, unsupervised approaches appeared. One study categorized laser cladding beads using clustering, while another used neural networks to improve nozzle efficiency. Recurrent neural networks entered the field in 2018 to predict time-series temperature data. A year later, convolutional neural networks were introduced for image-based defect detection, setting the stage for more sophisticated vision-driven monitoring.
A sharp rise in activity followed in 2020. CNNs were trained on coaxial images to identify porosity in aluminum alloy deposits. Gaussian process regression models were developed to predict strain rates during deformation, and reinforcement learning frameworks were tested to optimize laser arm movement in multi-track deposition. This period marked a decisive shift from isolated demonstrations toward systematic application of advanced architectures.
Diversification after 2020
Research published after 2020 broadened the scope of machine learning in DED. Random forest classifiers were used to segment porosity in wire-based processes, while long short-term memory (LSTM) networks forecasted melt pool temperatures. Hybrid frameworks combining neural networks with finite element simulations enabled interpass temperature prediction, helping reduce thermal defects in WAAM. Physics-informed neural networks emerged by 2022, embedding governing equations into training to balance predictive accuracy with adherence to physical laws.
More recent studies have moved into hybrid and temporal models. In 2023, gated recurrent unit (GRU) networks outperformed CNNs and dense neural networks for predicting melt pool dynamics, showing that sequential architectures capture temporal dependencies more effectively. Semi-supervised approaches appeared in 2025, combining regression for labeled melt pool data with unsupervised clustering to extract hidden features from sensor inputs. These methods aim to address the shortage of high-quality labeled datasets without sacrificing model robustness.
Barriers to deployment
Despite the growth in methods and applications, several obstacles limit industrial adoption. Closed-loop control remains rare, with most models operating in open-loop configurations where sensor data informs predictions but does not drive real-time parameter adjustments. Effects linked to deposition location, such as stress accumulation at edges or distortion near corners, are also poorly represented, even though they strongly influence final part performance.
Data limitations are another obstacle. High-fidelity finite element models that capture coupled thermal and mechanical behavior are expensive to compute, restricting dataset size. Experimental data collection faces similar challenges, as labeling stress fields or microstructures with precision is technically demanding. These factors explain why supervised models dominate the literature, while unsupervised and semi-supervised approaches remain underdeveloped.
Methodological diversity also complicates the field. Regression, support vector machines, fuzzy logic, clustering, and deep learning all appear across studies, but few comparative works evaluate these approaches under identical conditions. Without benchmarks, best practices remain undefined, particularly for applications such as defect detection, melt pool monitoring, or residual stress prediction.
Outlook
The University of Windsor review identifies several priorities for future research. Location-aware modeling, capable of encoding deposition history and toolpath strategies, is essential for predicting anisotropy and part reliability. Real-time closed-loop control, integrating neural networks with sensor feedback, is viewed as a critical step toward adaptive manufacturing systems.
Physics-informed models continue to attract attention. By embedding physical constraints into data-driven architectures, they offer a balance of interpretability and predictive efficiency, particularly for thermo-mechanical simulations. Another underexplored area is uncertainty quantification. Probabilistic methods could provide confidence intervals for predictions, which is vital in safety-critical applications such as aerospace and defense.
The review also notes an imbalance in research focus. Process optimization and melt pool geometry dominate, while defect classification, multi-objective prediction, and real-time adaptive control remain less explored. Addressing these gaps will require larger, more representative datasets, comparative evaluations of methods, and integration of spatial and temporal variability into model design.
Limited spaces remain for AMA:Energy 2025. Register now to join the conversation on the future of energy and additive manufacturing.
Ready to discover who won the 2024 3D Printing Industry Awards?
Subscribe to the 3D Printing Industry newsletter and follow us on LinkedIn to stay updated with the latest news and insights.
Featured image shows trend in the number of published papers—the orange bar predicts the number of papers to be published from July 2025 until the end of the year. Image via University of Windsor.
link