Efficient modeling of dynamic properties in K3C60 using machine learning force fields

Abstract

Dynamic properties of the alkali-doped molecular crystal K${}_3$C₆₀ are investigated using machine learning force field (MLFF). Our trained force fields successfully reproduce phonons comparable to those from density functional theory 10-times faster in computational efficiency. Specific heat with MLFF also show good agreement with experimental data, demonstrating its reliability for thermodynamic analysis. Two descriptor schemes, Smooth Overlap of Atomic Positions and Atomic Cluster Expansion, are employed and systematically compared. This study represents the first detailed exploration of dynamic properties in molecular crystals using MLFF, highlighting MLFF’s potential in more complex molecular crystals, such as C₆₀ disorder in fullerides, molecular melting.