| Abstract Scope |
Every interatomic potential, from the embedded-atom method to foundation machine-learning models, is fit to limited data and asked to transfer beyond it. Transferability, where and why a potential fails, has been the field's recurring question, and the community has answered it institution by institution: OpenKIM, the NIST Interatomic Potentials Repository, growing benchmark suites. The argument here is that failure itself is now a measurable object, and that characterizing it demands the field's own evidentiary standards. Aggregating prediction error across roughly 900 published potentials by random-effects meta-analysis, and identifying its drivers by causal inference, reveals that error concentrates on a low-dimensional hyper-ribbon, the same sloppy-model geometry seen elsewhere in physics. That geometry survives the classical-to-machine-learning transition across fourteen of fifteen benchmark metals, making transferability a property one can predict rather than discover case by case. Doing this honestly means falsifiable claims and public self-correction, not curated success stories. |