A. Natarajan

Data scarcity in machine-learned interatomic potentials can be largely bypassed simply by switching from standard one-hot chemical encodings to an occupational site basis, which uniquely unlocks exact analytical average-atom potentials for disordered alloys.

Sampling rare vacancy fluctuations is completely unnecessary for multicomponent alloy diffusion: a single-vacancy partition function truncation slashes Monte Carlo compute by 90% while analytically isolating the true vacancy tracer diffusion coefficient.