(569x) Monte Carlo Molecular Dynamics Simulations to Study Supported Nanoparticles

As machine learning interatomic potentials (MLIP) enable the calculations of systems larger than typical quantum chemical investigations, new methods are required to properly utilize MLIPs. In this work, we compare hybrid Monte Carlo Molecular Dynamics (MCMD) with Simulated Annealing (SA) with the goal of finding low energy supported nanoparticle configurations utilizing MLIPs trained at the DFT level of theory. We develop an MPI parallelized MC evaluation implementation that can be deployed on a range of CPU and GPU hardware utilizing NequIP, ASE, and MPI4Py. Evaluating images in parallel allows for stricter acceptance criteria for finding low energy structures without slowing down the real time simulation speed and faster simulations are possible at a given acceptance criteria for MC steps. The careful design of reasonable trial steps is also shown to be critical in this heterogenous system and compared towards simpler MCMD methods. The MCMD simulations are shown to be more efficient than SA simulations in providing the lowest energy structure and provides a baseline for our future work to extend this tool to more complex chemical environments (oxides, carbides, zeolites, etc.)