Physics-inspired semi-supervised multi-loss learning for internal defect monitoring in laser directed energy deposition

Abstract In laser directed energy deposition (LDED) additive manufacturing, monitoring internal defects based on melt pool features is critical for ensuring product quality, process efficiency, and operational safety. However, existing monitoring approaches remain constrained by two key limitations: (i) the scarcity of defect-labeled data, since defect characterization typically requires destructive testing; and (ii) the dominance of purely data-driven models, which often neglect the intrinsic connection between melt pool features and internal defects, consequently lack interpretability. To address these limitations, this paper proposes a physics-inspired semi-supervised defect monitoring (PISSDM) framework that integrates melt pool features with defect-related physical information within a multi-loss learning paradigm. Specifically, image processing extracts melt-pool width, length, area, perimeter and intensity. Correlation-derived feature weights define a class-specific consistency criterion for accepting high-confidence pseudo-labels, while crack and porosity penalties and a structural complexity term guide the selection of random-forest models. The forests are constructed using standard CART induction; the physics-inspired information acts through candidate-model selection and the composition of the augmented training set. Ablation and comparative studies on LDED data report an accuracy of 0.94, a support-weighted F1-score of 0.94 and a macro-F1 score of 0.87. The framework provides a means of incorporating defect-related knowledge into monitoring with limited annotations.