Efficient and Transferable Machine Learning Potentials for Zn-Coordinated Zeolitic Imidazolate Frameworks Through Structural Diversity Sampling and Transfer Learning

Abstract

Zeolitic Imidazolate Frameworks (ZIFs) exhibit remarkable structural flexibility that governs many of their functional properties, yet accurately describing their dynamic behavior remains a significant computational challenge. Classical force fields often fail to capture framework flexibility, whereas first-principles methods are prohibitively expensive. In this work, high-fidelity machine learning potentials (MLPs) are first developed for ZIF-8 using the equivariant Allegro architecture. To reduce the cost associated with the development of MLPs, an active-learning-inspired structural selection strategy based on the smooth overlap of atomic positions (SOAP) is implemented, allowing DFT-level accuracy to be achieved with fewer training configurations. Benchmarking against currently available models demonstrates that, while those models also offer reasonable energy predictions, their force errors remain 1 order of magnitude larger than the specialized MLP developed herein with a force mean absolute error of as small as 0.016 eV/Å. Building upon these findings, a transferable MLP spanning diverse Zn-based ZIFs is further constructed. Moreover, a hierarchical transfer learning workflow is also established to enable rapid adaptation to unseen topologies with improved accuracy using only a limited number of additional configurations. As a demonstration, MD simulations powered by the developed MLP successfully capture methane-induced gate-opening in ZIF-8, resolving the structural transition at subnanosecond resolution. Overall, this work provides a transferable, DFT-accurate machine learning force field for Zn-based ZIFs.