OPES-Enhanced Machine Learning Potentials Capture Node-Assisted Diffusion and Ring-Opening in SALEM-2

Machine learning potentials (MLPs) can help bridge the length- and time-scale gaps required to study functional nanoporous materials at ab initio accuracy. MLPs are usually trained using quantum chemical data obtained from traditional molecular dynamics (MD) simulations. These MD simulations predominantly sample near-equilibrium configurations on the potential energy surface. This often limits the MLPs ability to accurately describe less favorable, high-energy configurations. We address this challenge by introducing an active learning framework based on the On-the-fly-Probability-Enhanced-Sampling (OPES) method. Using imidazole diffusion in the SALEM-2 metal-organic framework (MOF) as an example, our approach employs a time-dependent OPES bias with expanded temperature- and distance-based collective variables. This allows extended MD simulations of linker molecule diffusion over 5 ns with ab initio accuracy. Interestingly, our simulations reveal a novel ring-opening phenomenon during imidazole diffusion across the four-membered window. This process, which is not captured by classical potentials, suggests the advantage of using OPES-based training curricula to study diffusion in MOFs.