Benchmarking and advancing machine-learning interatomic potentials for 3C-SiC across equilibrium and extreme conditions

arXiv:2510.06966 · cond-mat.mtrl-sci · Submitted 2025-10-08 · Read on arXiv

cond-mat.mtrl-sci

Submitted: 2025-10-08

Updated: 2026-11-03

Code: https://github.com/mcaroba/turbogap

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

The gist: Several machine-learning interatomic potentials (MLIPs) have been developed for cubic silicon carbide (3C-SiC).

Terminology

Abstract

Several machine-learning interatomic potentials (MLIPs) have been developed for cubic silicon carbide (3C-SiC). Here, we benchmark seven existing MLIPs against a common set of structural, thermal, defect, radiation-damage, disordered-phase, and computational-performance tests. We also introduce two complementary MLIPs trained on DFT-labeled crystalline, liquid, amorphous, and defective configurations. TGAP, a Gaussian approximation potential based on the soap turbo descriptor, provides consistently high accuracy across the sampled configurational space. MTP, a custom moment tensor potential, retains good accuracy across these regimes while offering substantially lower computational cost, GPU acceleration, and access to multi-million-atom simulations. As an application-level test, both potentials remain stable in displacement cascades spanning short-range repulsion, transient disorder, and recrystallization. Overall, MTP emerges as a particularly strong accuracy--cost compromise for large-scale and strongly non-equilibrium simulations of 3C-SiC.

Sources

Related papers