Abstract
Crack detection in powder compaction is commonly addressed either through physics-based finite element modeling or through data-driven classifiers operating on sensor signals obtained from the press. In this work, a unified perspective is introduced by comparing the outcomes of a physics-based stress-path indicator and a data-driven force-signal-based classifier for crack detection on identical industrial specimens in a fixed-die multi-level pressing process. Crack initiation is assessed based on a physical model through violations of a Drucker–Prager Cap (DPC) yield envelope, and a data-driven approach that evaluates statistical features extracted from inter-level force differentials. The physics-based DPC indicator achieves 75% accuracy while the data-driven ML classifier achieves 93% accuracy with sub-second inference time. Stress-path violations correspond quantitatively to force-differential anomalies, establishing an explicit mechanical correspondence between outcomes of the physics-based and the data-driven approach. By addressing crack detection with both physics-based and data-driven approaches, this study clarifies their complementary roles and supports scalable, interpretable crack detection strategies for powder metallurgy.