Modeling Diffusion in Metal-Organic Frameworks Using On-the-fly Probability Enhanced Sampling-Based Machine Learning...

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Machine learning potentials (MLPs) enable accurate simulations of nanoporous materials but struggle with rare events due to limited training data from near-equilibrium states.
• Studying high-energy, off-equilibrium configurations is crucial for understanding rare events but is computationally expensive with traditional methods.

Purpose of the Study:

• To develop and demonstrate a novel active learning strategy for training MLPs capable of simulating rare events in nanoporous materials.
• To overcome data scarcity issues in training MLPs for complex physicochemical phenomena.

Main Methods:

• Implemented an active learning curriculum utilizing the On-the-fly Probability Enhanced Sampling (OPES) method.
• Applied time-dependent OPES biases with collective variables to systematically explore the potential energy surface of imidazole diffusion in the SALEM-2 metal-organic framework (MOF).
• Performed molecular dynamics (MD) simulations with near-density functional theory (DFT) accuracy.

Main Results:

• Successfully captured imidazole diffusion through both 6-membered and a previously unreported 4-membered window in the MOF.
• Observed a novel ring-opening mechanism involving transient Zn-N bond dissociation, which is beyond the scope of classical force fields and DFT.
• Achieved nanosecond-scale MD simulations with high accuracy.

Conclusions:

• Enhanced sampling methods, like OPES, are effective in overcoming data scarcity for training MLPs to study rare events in nanoporous materials.
• The developed approach enables the accurate simulation of complex, off-equilibrium processes crucial for materials discovery and design.