Agnostic calculation of atomic free energies with the descriptor density of states

TL;DR

This paper introduces a model-agnostic method for calculating atomic vibrational free energies using descriptor density of states, enabling rapid, accurate, and differentiable predictions suitable for uncertainty quantification and inverse design.

Contribution

The authors develop a descriptor-based, model-agnostic approach that estimates free energies without predefined interatomic potentials, utilizing score matching for efficient high-dimensional density estimation.

Findings

Accurate free energy predictions within 1-2 meV/atom for various phases.

Rapid computation of free energies in microseconds of CPU time.

Successful fine-tuning of phase transition temperatures via back-propagation.

Abstract

We present a new method to evaluate vibrational free energies of atomic systems without a priori specification of an interatomic potential. Our model-agnostic approach leverages descriptors, high-dimensional feature vectors of atomic structure. The entropy of a high-dimensional density, the descriptor density of states, is accurately estimated with conditional score matching. Casting interatomic potentials into a form extensive in descriptor features, we show free energies emerge as the Legendre-Fenchel conjugate of the descriptor entropy, avoiding all high-dimensional integration. The score matching campaign requires less resources than fixed-model sampling and is highly parallel, reducing wall time to a few minutes, with tensor compression schemes allowing lightweight storage. Our model-agnostic estimator returns differentiable free energy predictions over a broad range of potential parameters in microseconds of CPU effort, allowing rapid forward and back propagation of potential variations through finite temperature simulations, long desired for uncertainty quantification and inverse design. We test predictions against thermodynamic integration calculations over a broad range of models for BCC, FCC and A15 phases of W, Mo and Fe at high homologous temperatures. Predictions pass the stringent accuracy threshold of 1-2 meV/atom (1/40-1/20 kcal/mol) for phase prediction with propagated score uncertainties robustly bounding errors. We also demonstrate targeted fine-tuning, reducing the alpha-gamma transition temperature in a non-magnetic machine learning model of Fe from 2030 K to 1063 K through back-propagation, with no additional sampling. Applications to liquids and fine-tuning foundational models are discussed along with the many problems in computational science which estimate high-dimensional integrals.