Abstract
Crack formation during metal powder compaction is a critical manufacturing challenge that compromises component quality and efficiency. Traditional non-destructive testing is costly, slow, and sampling-based, underscoring the need for predictive simulation and in line quality assessment. This thesis develops an integrated physics-based and data-driven framework for crack prediction in multilevel powder compaction, combining open-source finite element modelling with machine learning classification. A strain-dependent finite element model was implemented in Code_Aster, embedding an exponential elastic modulus law to capture elastic modulus evolution during densification. Compared to constant modulus baselines, this reduced force-displacement errors to ≤26%. The model was then coupled with a multi-body dynamics representation of the hydraulic press written in Python, including servo-valve flow, chamber pressure dynamics, and PID control with velocity feed-forward. The coupled simulation achieved <3% open-loop and <0.1% closed-loop trajectory errors against industrial data. Further, two complementary crack detection strategies were evaluated. A Drucker-Prager Cap-based indicator assessed stress paths in the p–q plane, yielding 75% accuracy for crack prediction. In parallel, ensemble machine learning classifiers trained on force-based features exceeded 90% accuracy in distinguishing cracked from non-cracked compacts. Validation was performed on multi-level iron powder parts with controlled crack induction and metallographic verification. Results highlight the interpretability of the physics-based DPC predictor and the high performance of the ML classifier. Together, they provide a robust foundation for digital twin implementations that support real-time crack prevention, reduce scrap, and minimize reliance on costly NDT.