Machine learning in directed energy deposition: A systematic literature review

Among metal additive manufacturing (AM) processes, directed energy deposition (DED) is a versatile route for fabricating, repairing, and coating large components. Advances in machine learning (ML) are reshaping DED via real-time monitoring, predictive modeling, and adaptive control, yet the literature is fragmented across sensing modalities, methods, and applications. This systematic review synthesizes 148 peer-reviewed studies (2016-2025), organizing them by application domain (defect metrics, geometric prediction, mechanical properties, melt-pool characterization, thermal-field modeling, and emerging topics) and by data-driven approach. Quantitative analysis shows that supervised learning predominates; classical regression remains in use but is declining; reinforcement learning is rarely explored; and multimodal fusion is growing but still a minority. Approximately 75% of studies use experimental data, with simulation-based and hybrid studies forming smaller fractions. Reported accuracies are often high (e.g., R 2 > 0.95 or classification > 90%), but persistent challenges limit deployment: scarce public datasets, high annotation cost, inconsistent definitions and metrics, sensor and modality constraints, weak generalization and cross-machine transfer, limited real-time implementations, and scalability issues. We conclude with priorities for the field: open benchmarks and standardized taxonomies, annotation-efficient pipelines, sensoragnostic multimodal learning, physics-guided and uncertainty-aware modeling, and robust real-time control. The review provides a unified map linking manufacturing objectives to data modalities and learning paradigms, guiding the development of reliable, high-performance DED systems.