Provenance records predict which multi-scale conclusions a foundation-potential release can change in SiGeC semiconductors

Multi-scale simulations of semiconductor alloys rest on foundation interatomic potentials (machine-learned force models trained on large density-functional databases), re-released as training data grow. A release changes defect energies and phonon stiffness and, through them, the recipes and design verdicts downstream; recomputing everything is costly, and recomputing nothing leaves conclusions that may no longer hold. Workflow records list every result downstream of a calculation, not which could change a decision. We tested whether the conclusions a release can change are predictable from a campaign’s provenance record (a record of which calculation fed which result) before any rerun and settled by recomputing only the predicted units. The chain served a SiGeC design question (germanium content, carbon content and SiC particle size for low lattice thermal conductivity), with MACE-MP-0 replaced by MACE-MPA-0 and all predictions and decision rules fixed in advance. From the record, 49 of 128 record entries were reached, 24 conclusions were classified by their margin to a decision threshold, and four recompute units were named. Recomputing those units (70 relaxations, 91 molecular-dynamics runs) settled 20 conclusions: none changed outside the predicted set (exact 95% upper bound on the miss rate 0.24), eight of nine flagged as at risk held, and one, Ge attraction to the first shell around a SiC particle, reversed by less than a millielectronvolt past its threshold, while the conclusion it served (no germanium-depleted zone forms) held. Magnitudes moved more than assumed: the energy carbon gains by precipitating as SiC grew by 0.65–0.75 eV against an assumed 0.3 eV; the recipe and the germanium saving held. Recomputation cost 3.5 CPU-task-hours and 354 GPU-hours, 2.2 times the GPU estimate scaled from the recorded runs. Decisions whose margins exceed a measured release shift can be carried across a release with a recorded check; magnitudes need recomputation, and explicit records of each result’s inputs and code would make the prediction a lookup.