betaFeedback

WorksPhysics and Astronomy

Distilling an atomistic foundation model is all you need for large-scale molecular dynamics of semiconductor interfaces

Abstract

State-of-the-art atomistic foundation models provide broadly transferable energy and force predictions, but their direct deployment in large-scale molecular dynamics remains computationally costly. Here we examine whether their knowledge can be distilled into compact potentials while retaining the physical responses of heterogeneous semiconductor interfaces. From 128 pretrained models spanning 22 model families, we select a fixed EquiformerV3 teacher and train atomic cluster expansion (ACE) and neuroevolution potential (NEP) students without target-specific teacher fine-tuning, with both simulation speed and accessible system size improved by several orders of magnitude relative to the teacher model. Matched training configurations and architectures enable direct comparisons with students trained on density-functional-theory (DFT) reference labels. For 79,920-atom BAs/GaN interfaces, direct distillation preserves termination-dependent thermal conductance and dominant spectral features, with conductance differing by at most 6.6% from DFT-trained counterparts. In 29,792-atom Cu/diamond specimens loaded at 500 K, tensile strength and loading work differ by 2.8% and 1.5%, respectively. Exploratory SiC/a-C/SiC simulations retain the principal thermal bottleneck and the direction of the annealing response. Furthermore, a four-model committee retrieves 42-56% of the highest-error teacher configurations with a 10% selection budget and improves high-error force fitting, but does not consistently improve downstream properties. These findings demonstrate that direct distillation can preserve key interfacial thermal and mechanical responses in compact interatomic potentials, opening a path towards large-scale, high-throughput simulations of heterogeneous interfaces.

Paper details

vlnc:W_snbypqfx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Generative AI tools

View statement
Looking for related papers…