GAP vs. MACE: efficiency evaluation in a liquid electrolyte system

Abstract Machine learning interatomic potentials (MLIPs) have revolutionized molecular simulations, but as they evolve, so does the demand for advanced computing architectures, particularly graphics processing units (GPUs). However, the high cost of GPUs limits accessibility, making it crucial to compare GPU and central processing unit (CPU) based MLIPs under practical conditions. This study examines two popular MLIPs: the GPU-accelerated multi-atomic cluster expansion model and the CPU-based Gaussian approximation potentials model, applied to a battery electrolyte system known for its complex properties. By focusing on these models, our study evaluates differences in computational performance, resource efficiency, and accuracy in reproducing experimental properties. This rigorous benchmark provides insights into the trade-offs between GPU and CPU-based approaches in molecular simulations.