Distilling latent electrostatics from foundation machine learning interatomic potentials

A new method extracts latent electrostatic properties from computationally expensive machine learning interatomic potentials, yielding faster models that can predict electrical response and infrared spectra, according to a paper submitted on 12 June 2026 [1]. Foundation machine learning interatomic potentials (MLIPs) enable atomistic simulations across wide chemical spaces but often lack explicit electrostatics, limiting their use for systems governed by long-range interactions [1][2]. The technique, called Latent Ewald Summation (LES), learns latent atomic charges and long-range electrostatics from density functional theory (DFT) energy and force labels [1][2]. Researchers now apply LES to extract electrostatics latent in foundation models by using a teacher model’s predicted energies and forces to train a lightweight LES-augmented student MLIP, with optional fine-tuning on additional DFT data [1][2]. The resulting student models reduce computational cost while providing access to Born effective charge tensors and infrared spectra [1][2]. The team benchmarked student models distilled from a broad set of foundation MLIPs—including UMA, MACE, Orb, eSEN, GemNet-OC, PET, and EquiformerV2-based models—against experimental infrared spectra for liquid water, concentrated hydrochloric acid, and the anatase TiO2(101)-water interface [1][2]. Across these systems, electrostatic response could be extracted from most foundation MLIPs [1][2]. The benchmark further showed that the underlying DFT level and dataset used to train the teacher model play a larger role than architecture in determining electrostatic and spectroscopic accuracy [1][2]. For the TiO2-water interface, fine-tuning with a modest amount of higher-level DFT data improved structural and infrared predictions [1][2]. The work demonstrates that LES-based distillation provides a practical route for converting foundation MLIPs into efficient, electrically responsive models while also testing the physical fidelity encoded in foundation models [1][2]. Transfer learning and fine-tuning strategies have been explored in related computational chemistry contexts. Prior work examined how the OC20 dataset could aid the OC22 dataset via transfer learning or joint training, noting that the small-molecule and drug-discovery communities have used transfer learning to move between varying levels of electronic structure calculations [4]. The current LES distillation approach extends such ideas to the electrostatic domain, using teacher-student training rather than dataset-to-dataset transfer [1][2].