Phase stability in multicomponent solid solutions depends on configurational entropy beyond the ideal mixing limit, but capturing it together with vibrational entropy within the same atomistic framework remains challenging. Here, we extend non-equilibrium thermodynamic integration to composition-dependent transformations through an alchemical interpolation of the interactions, combined with Monte Carlo identity exchange moves and molecular dynamics that sample the vibrational and non-ideal configurational entropy along the integration path. We apply the framework to the Au-Cu binary alloy using Atomic Cluster Expansion potentials trained on density functional theory data using the LDA, PBE, and r2SCAN functionals, and construct composition-temperature phase diagrams directly from atomistic free energies. We find that explicit configurational sampling lowers the AuCu order-disorder transition temperature predicted by the ACE potential trained on LDA data from approximately 810 K to 710 K, closer to the experimental value of 683 K, and substantially widens the stability range of the solid solution. At the same time, the much larger sensitivity to the exchange-correlation functional shows that this level of agreement should not be interpreted as general predictive accuracy. Non-ideal configurational entropy must therefore be sampled explicitly, alongside a careful choice of functional, for a reliable atomistic description of binary phase diagrams.
Sarath Menon, M. Poul, T. Hickel et al.· 0 citations
The construction of machine-learned interatomic potentials (MLIPs) is often limited by the cost of generating large density-functional-theory (DFT) training datasets. For systematically generated structure pools such as ASSYST, a central practical question is how many configurations must be labeled to achieve reliable accuracy. Here we assess geometry-based, label-free subset selection for training linear Atomic Cluster Expansion (ACE) potentials. Using statistical leverage scores and CUR-type sampling, we compare leverage-guided selection against random, energy-based, and force-based baselines under controlled iterative protocols. Elemental Al provides the primary benchmark, with Cu and Al-Cu alloys used for transfer validation. Leverage-guided subsets recover plateau-level energy and force accuracy using substantially smaller labeled fractions (approximately 30-40%) than random sampling, corresponding to an effective 2-3x reduction in DFT labeling for the systems studied. In alloy tests, defect energetics remain comparable across strategies once sufficient chemical diversity is included, while leverage selection maintains competitive accuracy at reduced training size. These results demonstrate that descriptor-space-guided, label-free subsampling can significantly reduce DFT workload for linear ACE models trained on ASSYST structure pools without degrading defect-level fidelity.
Aynour Khosravi, M. Poul, J. Neugebauer et al.· 0 citations
