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

Fullerides forms a big familiy of molecular crystals exhibiting various useful electro- and magnetochemical properties. Therefore, an efficient in silico method is desirable to look into such chemical system. To alleviate the burden of computationally-heavy first-principles calculations, here we present the successful attempts of using Machine Learning Force Field (MLFF) in predicting dynamical properties of the alkali-doped fulleride K₃C₆₀. Two on-the-fly Gaussian-Process-MLFF schemes based on different atomistic descriptors, Smooth Overlap of Atomic Position (SOAP) and Atomic Cluster Expansion (ACE), have been experimented. The performance of generated K₃C₆₀ MLFFs are validated by accurate prediction on energy and forces of 1,000 randomly disturbed K₃C₆₀ structures compared to respective DFT results. Several other dynamical properties including phonon dispersion, elastic moduli and heat capacity obtained by MLFFs have also shown good agreement with results from DFT calculations, whose computational cost on such post-processing is several times more expensive than training a single MLFF that performs effortless post-processing. This shows the potential of applying MLFF to complex molecular crystal systems and pave the way to the investigation of much more intricate fulleride properties.