Learning composition-sensitive signatures in multi-material PBF-LB: a lightweight, modality-aware, explainable graph-attention sensor fusion framework for in-situ monitoring of graded 316L–CuCrZr alloys

Abstract Decoding composition-sensitive process signatures in Laser Powder Bed Fusion (PBF-LB) of multi-material structures remains challenging due to the complex and transient nature of melt-pool emissions across spatially varying material gradients. In this study, we propose a modality-aware learning framework that fuses acoustic emission (AE) and back-reflected optical emission (OE) at the laser wavelength for spatiotemporally resolved classification of local composition in graded 316L–CuCrZr alloys. The framework integrates learnable shapelet extraction with graph-based attention, where shapelets act as interpretable descriptors of modality-specific temporal behavior and enable compact, discriminative representation learning. The resulting dual-modality representations are organized as a temporal graph of signal segments, and a Graph Attention Network (GAT) with modality-specific attention heads adaptively prioritizes sensor streams based on their relevance to compositional variation. Across five Cu-containing concentrations (20%–100 wt.%), the model achieved up to 92% accuracy with only ~ 4,000 trainable parameters. Beyond accuracy, the results show that sensor fusion is imperative for multi-material PBF-LB, as no single modality consistently captures composition-dependent phenomena across the CuCrZr gradient. The learned attention and saliency trends further reveal that OE becomes increasingly influential at higher CuCrZr fractions, consistent with increased reflectivity at the laser wavelength. Overall, the proposed framework establishes a compact and interpretable approach for decoding composition-dependent multimodal process signatures in multi-material PBF-LB.