Reviews published on ResearchGate Researchers at the University of Windsor will analyze how machine learning (ML) is applied to direct energy deposition (DED) and wire arc additives (WAAMs). This review covers research published between 2010 and mid-2025, shows that research activities have expanded rapidly since 2020, and neural networks based on deep learning, fuzzy logic and physics have moved from isolated experiments to mainstream topics. Despite this advancement, gaps remain in closed-loop process control, machine generalization, and integration of location-specific mechanical effects.
The author has developed a Python script that automates searches through the CrossRef Bibliographic Database. Publications were obtained using a combination keyword set that covers both DED processes and machine learning techniques, and duplicates were removed. Each entry was manually checked to confirm that the study applied ML specifically to DED or WAAM. Although this method captured a wide range of cross sections of the field, the dataset is limited to CrossRef-Indexed content, with the exception of its own databases such as Scopus and Web of Science. This introduces some degree of bias, but provides a representative overview of how artificial intelligence has entered this metal 3D print branch.
From the early foundations to turning points
Research in the early 2010s was sparse and exploratory. In the initial studies, fuzzy logic models were tested to adjust the scan speed of laser cladding and apply simple neural networks to predict quality and optimize bead geometry. These projects demonstrated the feasibility of data-driven methods, but remained limited to narrow parameter optimization tasks.
By 2016, an unsupervised approach had emerged. In one study, clustering was used to classify laser cladding beads, while in another, neural networks were used to improve nozzle efficiency. Recurrent Neural Networks entered the field in 2018 and predicted time series temperature data. A year later, convolutional neural networks were introduced for image-based defect detection, setting up a more sophisticated vision-driven monitoring stage.
Activities continued to rise sharply in 2020. CNNs were trained on coaxial images to identify porosity in aluminum alloy deposits. The Gaussian process regression model was developed to predict strain rates during deformation and tested the reinforcement learning framework to optimize laser arm movement in multitrack deposition. This period marked a critical shift from isolated demonstrations to systematic application of advanced architectures.
Diversification since 2020
Research published since 2020 has expanded the scope of DED's machine learning. A random forest classifier was used to segment porosity in a wire-based process, while a long-term memory (LSTM) network predicted the melt pool temperature. The hybrid framework combining neural networks with finite element simulations has enabled interpass temperature prediction and helped reduce thermal defects in WAAMs. A physically-based neural network has emerged by 2022, embedding equations into training to balance prediction accuracy and adherence to physical laws.
More recent research has shifted to hybrid and temporal models. In 2023, Gate Recurrent Unit (GRU) networks outperform CNNs and dense neural networks to predict melt pool dynamics, indicating that sequential architectures capture time dependencies more effectively. In 2025, a semi-monitored approach emerged, combining regression and unsupervised clustering of labeled melt pool data to extract hidden features from sensor inputs. These methods aim to address the lack of high quality labeled datasets without sacrificing the robustness of the model.
Barriers to deployment
Despite the increasing number of methods and uses, some obstacles have limited the adoption of industry. Closed loop control is rare, and most models operate in open loop configurations that signal predictions to sensor data but do not facilitate adjustment of real-time parameters. Although its strong impact on the final performance is also less expressed, it is also not expressed in the effects linked to deposition locations such as stress accumulation at corner edges and corner distortion.
Data restriction is another obstacle. High fidelity finite element models that capture combined thermal and mechanical behavior are expensive to limit the dataset size. Experimental data collection faces similar challenges. This is because accurate labeling of stress fields and microstructures is technically demanding. These factors explain why supervised models dominate the literature, but the unsupervised semi-teacher approach remains underdeveloped.
Methodological diversity also complicates the field. Regression, support vector machines, fuzzy logic, clustering, and deep learning all appear throughout the study, but few comparative works assess these approaches under identical conditions. Without benchmarks, best practices remain undefined, especially in 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, which allows for encoding deposition history and toolpath strategies, is essential for predicting anisotropy and some reliability. Real-time closed-loop control that integrates neural networks and sensor feedback is considered an important step into adaptive manufacturing systems.
Physics-based models continue to attract attention. Embedding physical constraints into data-driven architectures provides a balance between interpretability and predictive efficiency, particularly for thermomechanical simulations. Another area of undrilling rigs is quantification of uncertainty. Probabilistic methods may provide confidence intervals for predictions. This is essential for safe applications such as aerospace and defense.
The review also points to imbalances in the focus of the research. Process optimization and melt pool geometry dominates, but defect classification, multi-objective prediction, and real-time adaptive control have not been much investigated. Addressing these gaps requires larger, more representative data sets, comparative evaluations of methods, and integration of spatial and temporal variation into model design.
There is a limited space left AMA: Energy 2025. Sign up now and join us in conversations about the future of energy and additives.
Ready to discover who won 2024 3D Printing Industry Award?
Subscribe to 3D Printing Industry Newsletter And follow us LinkedIn Stay up to date with the latest news and insights.
The featured images show trends in the number of published papers. Orange Bar predicts the number of papers to be published from July 2025 to the end of the year. Images via the University of Windsor.
