What question did this study set out to answer?

This review critically assesses machine learning methodologies for process optimization and defect detection in metal additive manufacturing.

June 3, 2026Open Access

Machine Learning for Process Optimization and Defect Detection in Metal Additive Manufacturing: A Critical Review of Algorithms, In-Situ Monitoring Strategies, and Quality Assurance Frameworks

Key Points

This review critically assesses machine learning methodologies for process optimization and defect detection in metal additive manufacturing.
Reviews over 130 peer-reviewed studies from 2016 to 2025
Evaluates algorithms like CNNs for anomaly detection, LSTMs for monitoring, and reinforcement learning for parameter governance
Surveys in-situ monitoring technologies and quality assurance architectures for defect localization.
Identifies gaps in large annotated datasets and cross-domain generalization
Highlights the computational challenges of real-time deployment in industrial settings
Proposes a roadmap focusing on physics-informed neural networks and federated learning for enhanced ML quality assurance.

Abstract

Among the many manufacturing technologies to emerge in the last three decades, metal additive manufacturing (AM)—spanning laser powder bed fusion (LPBF), directed energy deposition (DED), and electron beam melting (EBM)—stands out for the speed and breadth of its industrial adoption. Once confined to producing plastic concept models, it now fabricates flight-critical aerospace structures, load-bearing skeletal implants, and thermally demanding heat management components. Yet persistent process-induced defects—among them gas porosity, lack-of-fusion voids, hot cracking, and inter-layer delamination—continue to jeopardize part integrity and create formidable barriers to certification in regulated sectors. Machine learning (ML) and deep learning (DL) have attracted growing attention as remedies, offering the capacity to extract decision-relevant patterns from the rich multi-modal data streams that flow continuously during metal AM builds. This review undertakes a critical examination of the current state of ML-driven process optimization and defect detection for metal AM. Convolutional neural networks (CNNs) are evaluated for image-based anomaly identification; long short-term memory (LSTM) networks and transformer architectures are assessed for temporal process monitoring; reinforcement learning is examined for closed-loop parameter governance; and generative adversarial networks are considered as instruments for training data augmentation. In-situ monitoring hardware—encompassing melt-pool cameras, pyrometers, acoustic emission sensors, and photodiode arrays—is surveyed alongside the sensor fusion strategies that sharpen defect localization performance. Closed-loop quality assurance architectures, digital twin integration, and ML-guided parameter optimization receive detailed critical evaluation. Drawing on a systematic synthesis of more than 130 peer-reviewed studies spanning 2016 to 2025, we identify structural gaps: a shortage of large annotated benchmark datasets, insufficient cross-domain generalization, and the computational demands of real-time industrial deployment. A forward-looking research roadmap—centred on physics-informed neural networks, federated learning architectures, and edge-optimized inference—is advanced as the route toward certifiable, industrially viable ML quality assurance systems for metal AM.

Mark Helpful

Bookmark

Relay

View Full Paper