Reproduce “Distilling latent electrostatics from foundation machine learning interatomic potentials” on OpenResearch

About this paper

Xiaoyu Wang, Bingqing Cheng

Researchers from UC Berkeley developed a distillation workflow that creates computationally efficient, lightweight machine learning interatomic potentials (MLIPs) from complex foundation models by extracting explicit long-range electrostatics. This approach enables the prediction of electrical response properties like Born effective charges and infrared spectra, which are not directly available from the original foundation models.

Key insights

• The Latent Ewald Summation (LES) framework successfully distills long-range electrostatics, including Born effective charge (BEC) tensors, from diverse foundation MLIPs with high data efficiency.
• The underlying Density Functional Theory (DFT) level used to train the teacher foundation model is the primary determinant of the distilled electrical response and spectroscopic accuracy.
• The architecture of the teacher foundation MLIP generally plays a minor role in distillation, with the exception that an adequate receptive field is crucial for implicitly capturing long-range electrostatic information.

Summary by alphaXiv.