WorksPhysics and Astronomy
A GPU-Native Nested Sampling Framework for Finite-Temperature Materials Modelling with Machine-Learned Interatomic Potentials
Abstract
Modern equivariant machine-learned interatomic potentials (MLIPs) have dramatically improved the accuracy of and the scales accessible to atomistic simulation, yet the thermodynamic sampling methods needed to turn those potentials into finite-temperature phase stability predictions have not kept pace computationally. We present the GPU-native, massively parallel nested sampling framework JAXRENS, implemented on top of JAX, which removes the practical bottlenecks preventing atomistic nested sampling from being combined with state-of-the-art MLIPs. The framework provides distributed multi-GPU execution, vectorised ensembles of walkers, and direct interoperability with modern force fields. We demonstrate the highly beneficial scaling of JAXRENS on hardware accelerators, and apply it to the simulation of the ZnS phase diagram based on a MACE foundation model, a procedure previously computationally inaccessible. Despite requiring in the order of 10^10 model evaluations, the final converged simulation completed within six days on four GPUs, marking an important step towards routine, first-principles-quality predictions of finite-temperature phase stability using machine-learned interatomic potentials.