Modeling of extreme thermal conductivity materials via machine-learned molecular dynamics

Machine-learned potentials (MLPs) are widely used to study atomic-scale thermal transport via the Green-Kubo (GK) formula, yet their reliable application remains challenging. For example, MLPs tend to underestimate the thermal conductivity (κ) of typical high-κ solids. Moreover, in low-κ superionic systems, atomic diffusion plays a major role so it is not a priori clear what energy gauge should be used. Here, we discuss machine-learned modeling of κ for these two cases. (i) We find that the fitting errors in machine-learned forces cause the underestimated κ, as they act as external perturbations to real interatomic forces. To address this issue, we introduce different levels of force noises and then extrapolate to the limit of zero force error. This correction yields good agreement with experimental data for all materials studied. (ii) We show that the κ values of superionic materials, computed via the conventional GK integral of energy flux, vary significantly across different MLP models. Further, we show that by using Onsager reciprocal relations, a robust κ can be obtained across a wide temperature range.