About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Chemical Basis Sets and Average-atom Potentials for Disordered Multicomponent Alloys |
| Author(s) |
Lorenzo Piersante, Deepak Somani, Anirudh Raju Natarajan |
| On-Site Speaker (Planned) |
Lorenzo Piersante |
| Abstract Scope |
Computational materials scientists routinely use machine-learning interatomic potentials (MLIPs) to study atomic structure. In large-scale simulations, bespoke models are essential owing to their computational efficiency. The Atomic Cluster Expansion (ACE) is a popular framework for fitting custom MLIPs. Although prior studies examined factors like self-interactions and regularization, the effect of alternative chemical basis sets, such as those used in traditional configurational cluster expansions, remains unexplored. We examine how the chemical basis affects the ACE learning performance. Explicitly treating the chemical degrees of freedom also enables the formulation of the Average-atom ACE, a mean-field approximation that regards the material as perfectly disordered. Such potentials help characterize the baseline thermodynamic, kinetic, and mechanical behavior of high-entropy alloys. We show that any pre-fitted ACE can be rigorously transformed into an average-atom ACE. Case studies on dilute and concentrated alloys demonstrate this transformation and establish the chemical site basis as a critical design choice. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, Computational Materials Science & Engineering |