Untangling linker dynamics in a metal–organic framework using ab initio molecular dynamics and physics-informed machine learning

Abstract

Organic linkers in metal–organic frameworks (MOFs) can display multiple degrees of freedom in their movement, and such complex dynamics are known to influence gas adsorption and diffusion, luminescence, and thermomechanical properties. Here, we have examined the torsional and transverse dynamics of 1,4-benezedicarboxylate (BDC) linkers in a Zn${}_2$(BDC)${}_2$ (1,4-diazobicyclo [2,2,2] octane) metal–organic framework (Zn-DMOF) by using ab initio molecular dynamics (AIMD) and physics-informed machine learning (ML). AIMD was performed at 200 and 300 K to generate time-resolved data for training ML models. Neural networks (NN), constrained with classical mechanics-based equation of motions, allowed us to extract harmonic and anharmonic torsion stiffness constants from AIMD data; while Gaussian process regression (GPR) was employed to probe for statistical correlations between torsional and transverse dynamics. These supervised ML approaches revealed that BDC linkers in pristine Zn-DMOF loose their resistance to torsional motion between 200 and 300 K, and that contributed to mechanical coupling between transverse and torsion modes at the higher temperature. As an extension of this effort, we utilized NN-derived torsion-time profiles and trained-GPR models to probe the effect of functionalizing BDC linkers with bulky non-polar groups. Our ML-based analysis confirm literature reports that such attachments will hinder torsion and transverse movements, and likely facilitate gas adsorption. Thus, our combined AIMD and physics-informed ML approach can potentially facilitate MOF design strategies that require engineering of linker dynamics for targeted applications.