Diverse Sharpness-Aware Minimization for Machine Learning Force Fields in Semiconductor Simulations

Dong-Hee Shin, Young-Han Son, Tae-Eui Kam2025 · DOI

Abstract

The growing demand for advanced semiconductor chips has driven the need for rapid and accurate material property prediction. While density functional theory (DFT) provides high accuracy, its computational cost limits practical applications. Machine learning force fields (MLFFs) offer faster alternatives but struggle with out-of-distribution (OOD) generalization when encountering unseen atomic structures. In this work, we investigate the use of sharpness-aware minimization (SAM) and its variants to improve MLFF generalization performance. Experiments on the SAMD23 benchmark demonstrate that SAM and its variants significantly enhance generalization on OOD datasets, with physics-informed SAM achieving the best performance.
BibTeX
@misc{shin2025diverse,
  author = {Dong-Hee Shin and Young-Han Son and Tae-Eui Kam},
  title = {{Diverse Sharpness-Aware Minimization for Machine Learning Force Fields in Semiconductor Simulations}},
  year = {2025},
  doi = {10.1109/isocc66390.2025.11330061},
}

Loading related content...

Figure