Physics–informed multimodal deep learning for real-time defect detection and process monitoring in solid-state additive manufacturing

Defect-free fabrication remains a major challenge in realizing industrial adoption of solid-state additive manufacturing (SSAM) due to thermomechanical interactions prone to cause defects during layer-by-layer fabrication, some of which are non-trivial to capture with conventional process monitoring. Signal-based models can be insensitive to subtle defect precursors, while image-based approaches offer fine morphological details but are generally limited to offline validation steps. This work proposes MCAT-Phys, a physics-informed multimodal deep learning framework that learns the joint distribution of acoustic emission, torque, and infrared signals with microstructural details from FESEM and optical microscopy to identify process-induced defects. Specifically, the framework was compared against (i) conventional unimodal baseline approaches such as CNN, ResNet, and U-Net; (ii) advanced unimodal models such as a Temporal Transformer for process signals, and SegFormer for microstructural images; and (iii) the proposed MCAT-Phys hybrid model. The framework’s robustness was statistically analyzed using five-fold cross-validation, confidence-interval analysis, and paired t-tests. MCAT-Phys achieved an accuracy of 94.2%, and F1-score of 0.88, an AUROC of 0.97, and an AUPRC of 0.96, outperforming the evaluated unimodal baselines. Error analysis further indicated an approximately 30% reduction in false negatives and a 40% improvement in crack-boundary continuity. These results illustrate that physics-informed cross-modal fusion over process dynamics, microstructural information, and process meta-data can be realized. In summary, MCAT-Phys provides a promising yet interpretable basis for SSAM defect monitoring but requires further validation across machines and material and processing conditions before industrial implementation.