Thermal Conductivity Predictions with Foundation Atomistic Models

Recent advances in machine learning have led to foundation models for atomistic materials chemistry, potentially enabling quantum-accurate descriptions of interatomic forces at reduced computational cost. These models are benchmarked by predicting materials' properties over large databases; however, these computationally intensive tests have been limited to basic quantities related to harmonic phonons, leaving uncertainty about the reliability for complex, technologically and experimentally relevant anharmonic heat-conduction properties. Here we present an automated framework that relies on foundation models to compute microscopic vibrational properties, and employs them within the Wigner formulation of heat transport to predict the macroscopic thermal conductivity in solids with arbitrary composition and structure. We apply this framework with the foundation models M3GNet, CHGNet, MACE-MP-0, and SevenNET to 103 diverse compounds, comparing predictions against first-principles references and introducing a benchmark metric based on conductivity. This framework paves the way for physics-aware, accurate predictions of vibrational and thermal properties, and for uncovering materials that violate semiclassical Boltzmann transport and feature exceptional heat-shielding or thermoelectric performance.