Applicability of MLIPs trained on unary/binary data to multicomponent alloys

Determine the applicability of machine learning interatomic potentials that are trained exclusively on unary and binary metallic systems to multicomponent metallic alloys, specifically whether such models can reliably generalize to alloys containing three or more elements across increasing compositional complexity.

Background

The study compares NEP and GRACE-FS models trained only on unary and binary configurations yet tested on multicomponent alloys. Despite demonstrating some extrapolative capability, the authors emphasize that the broader question of general applicability to multicomponent systems is not fully resolved.

To probe this issue, the paper introduces additional DFT benchmarks spanning 2–16 elements and evaluates model performance, finding clear architectural dependencies and showing that data augmentation helps but does not fully close the gap. This context motivates formally establishing how far unary/binary-trained MLIPs can be trusted in genuinely multicomponent regimes.