Machine learning-driven in situ defect monitoring and real-time process control in directed energy deposition: Techniques, challenges, and future prospects

Defect formation remains a primary barrier to the qualification of Directed Energy Deposition (DED) for safety-critical components. Existing inspection strategies are predominantly ex-situ and fail to provide actionable information during fabrication. This review critically evaluates machine-learning-enabled in-situ sensing frameworks for defect detection and process regulation in Direct Energy Deposition. Reported studies indicate that supervised learning applied to melt-pool imaging and defect classification yields accuracies between 87 and 99.3%, while semi-supervised and multi-modal approaches integrating optical, thermal, and acoustic signals achieve 92.5-97% accuracy. Inference latencies range from 4 to 200 ms, with high-speed implementations supporting control frequencies exceeding 250 Hz, enabling intra-layer corrective intervention. Defects are systematically classified according to geometric, morphological, and microstructural attributes, and learning paradigms spanning supervised, unsupervised, and reinforcement strategies are assessed across thermal, optical, acoustic, and spectroscopic sensing modalities. Despite promising performance, deployment is constrained by limited data fidelity, poor generalization across material–process spaces, and latency bottlenecks that impede closed-loop operation. The analysis identifies methodological deficiencies, including the absence of standardized benchmarks, fragmented sensor-fusion practices, and minimal adoption of adaptive feedback control. Emerging approaches such as contrastive representation learning, Bayesian uncertainty quantification, and physics-informed neural networks, which report R² values near 0.75, demonstrate potential for improved reliability. Transfer learning is shown to reduce labeled data requirements by up to 95%, while edge-based inference enables millisecond-scale execution. The review argues that progress toward industrial deployment depends on resolving foundational data and validation issues and on transitioning from offline defect detection to fully integrated, real-time control architectures.