Predicting Structural Properties of Pure Silica Zeolites Using Deep Neural Network Potentials

Abstract

Machine learning potentials (MLPs) capable of accurately describing complex ab initio potential energy surfaces (PES) have revolutionized the field of multiscale atomistic modeling. In this work, using an extensive density functional theory (DFT) dataset (denoted as Si-ZEO22) consisting of 187 unique silica topologies found in the International Zeolite Association (IZA) database, we have trained a DeePMD-kit MLP to model the dynamics of silica frameworks. The performance of our model is evaluated by calculating various properties that probe the accuracy of the energy and force predictions. This MLP demonstrates impressive agreement with DFT for predicting zeolite structural properties, energy-volume trends, and phonon density of states. Furthermore, our model achieves reasonable predictions for stress-strain relationships without including DFT stress data during training. These results highlight the ability of MLPs to capture the flexibility of zeolite frameworks and motivates further MLP development for nanoporous materials with near-ab initio accuracy.

Quantifying the Limits of Methane Activation in Cu-exchanged Zeolites using Reactive and Interpretable Machine Learning based Potentials

“…Specifically, this step samples the forces on the oxygen from [CuOCu] 2+ (denoted as f_O) -the histogram in Fig.7bshows that the model now has seen a much wider distribution of forces than what was available during the DFT/NEB. Additionally, the DFT/cMD also samples the zeolite backbone, which is shown previously by Sours et al79 .The remaining steps are more obvious. From right to left (or bottom to top), we have the initial states sampled by DP/MD at stage C1 (colored in blue), followed by the reactive region sampled by DP/NEB in stage C2 for a single site and D for all other MOR sites (colored in green) and then ends with the orange region for the final states (also sampled by DP/MD at stage C1).…”

“…The resulting dataset, denoted as rCuZEO23 (detailed in the SI, Section S2), is iteratively used to train a deep neural network-based potential. 79 We show that the resulting rMLP replaces the expensive DFT-based NEB calculations without any appreciable loss in accuracy -we evaluate C-H bond activation energies for 5,400 distinct sites across 52 zeolites within a couple minutes (per calculation) and obtain 3,356 valid sites suitable for methane activation. Our predictions are within 0.07 eV of the DFT computed energy barriers.…”

“…From a computational perspective, we recognize that force field-based methods (often accelerated by ML) are now routinely used to predict adsorption isotherms and diffusion properties across large libraries of zeolites. [79][80][81] Similarly, the computational catalysis community has reported several ML models to predict adsorption energy-based descriptors to approximate the kinetics of surface-mediated reactions. 72,73 However, to the best of our knowledge, this work represents the first example of using reactive MLPs to identify the transition state geometries and screen the catalytic performance of thousands of zeolite-based active sites at DFT accuracies.…”

“…1 and 2, this requires obtaining sufficient DFT training data from the relevant regions of the PES. Although MLPs trained on molecular dynamics have been widely used to predict equilibrium properties and obtain geometry optimized structures 79 , the development of reactive MLPs (denoted as rMLPs) for describing chemical reactions within configurationally complex materials (such as zeolites) has not yet been reported.…”