Machine Learning Potentials Unveils Rare Events During Diffusion of Organic Compounds in Metal-Organic Frameworks

Machine learning potentials (MLPs) bridge the gap between high-fidelity, short-time ab initio Density Functional Theory (DFT) simulations and long-time classical Molecular Dynamics (MD) simulations for functional nanoporous materials ^[1]. A key challenge is developing accurate MLPs, often achieved with active learning based on model ensemble uncertainty. However, traditional exploration strategies using MD simulations primarily sample configurations near local minima on the potential energy surface, limiting the MLP's ability to predict high-energy configurations (rare events). To overcome this limitation, we introduce an active learning framework utilizing the "On-the-fly-Probability-Enhanced-Sampling" (OPES) method for systematic exploration of high-energy configurations ^[2].<br/>This work showcases the effectiveness of the OPES-based active learning framework by modeling imidazole diffusion in functionalized ZIF-8 Metal-Organic Frameworks (MOFs) as a prototypical example. We employ a time-dependent OPES bias along expanded collective variables (ECVs) for temperature and distance-based CVs during model development. This enables extended MD simulations (up to 10 ns) with ab initio accuracy in large supercells using the trained MLPs, allowing detailed observation of the diffusion process. Intriguingly, our simulations reveal a previously unconsidered phenomenon: ring-opening events within the MOF structure during imidazole diffusion across four-membered rings. Classical potentials (e.g., UFF), lack the flexibility to represent these complex, large-scale structural rearrangements that involve breaking and reforming bonds within the MOF framework and hence cannot capture this emergent behavior even at long-time simulations. This discovery unlocks exciting possibilities for designing MOFs with novel functionalities by strategically modifying linkers to exploit this ring-opening process. Additionally, we investigate the impact of OPES on optimal training set selection and its transferability across diverse structures and chemistries.<br/>1. Guo, J.; Sours, T.; Holton, S.; Sun, C.; Kulkarni, A. R. Screening Cu-Zeolites for Methane Activation Using Curriculum-Based Training. ACS Catal. 2024, 14 (3), 1232–1242. https://doi.org/10.1021/acscatal.3c05275.<br/>2. Invernizzi, M.; Piaggi, P. M.; Parrinello, M. Unified Approach to Enhanced Sampling. Phys. Rev. X 2020, 10 (4), 041034. https://doi.org/10.1103/physrevx.10.041034.