Proof-Carrying Materials Framework Detects 93% of Unstable Interatomic Potentials

Proof-Carrying Materials (PCM) uses adversarial falsification and Lean 4 formal certification to address reliability gaps in machine-learned interatomic potentials. Benchmarks show existing models like CHGNet and MACE miss 93% of density functional theory-stable materials, while PCM's risk model achieves an AUC-ROC of 0.938 in predicting failures.