Bridging Simulation and Experiment using Bayesian and Deep Learning Approaches to Uncover Atomistic Insights in Heterogeneous Catalysis and Nanomechanics

Abstract

Existing experimental and computational methods have not allowed for precise design and control of material structure and performance given trade-offs in cost and atomistic resolution. Addressing these limitations, several applications of machine learned force fields (MLFFs) are presented here for understanding and predicting the behavior of materials in the context of metallurgy, nanomaterials, and heterogeneous catalysis. By combining the atomistic resolution and quantum mechanical accuracy of MLFFs at increased length- and time-scales compared to ab initio methods like density functional theory (DFT), we uncover previously inaccessible insights into material processes; such as, surface reconstruction on metals, responses of nanoparticles (NPs) to reactive adsorbates, and the identification of active sites and deactivation mechanisms in complex multi-component catalytic systems. We establish confidence in each of these MLFFs by using unbiased molecular dynamics simulations which are then compared directly to experimental observations, permitting confidence in their predictive abilities, moving the fields of heterogeneous catalysis and metallurgy towards in silico design and control of material structure and performance. Key areas of focus include assessing the boundaries of accuracy for MLFF architectures where both Gaussian processes and equivariant neural networks are validated against ground-truth DFT labels for transition metals, the investigation of defects in bulk copper and their impact on material plasticity, and the simulation of mesoscopic surface reconstructions of gold. We also uncover critical components of the nanoscale alloying behavior of GaPt and morphological response of Pt NPs under varying conditions, highlighting phenomena such as `mutual catalysis' and the impact of reactive chemisorption on nanoparticle shape, specifically influenced by H2 and CO. The versatility of MLFFs is further exemplified in the exploration of Pd-Au alloy NP catalysts, where the trained model successfully replicates various properties as predicted by DFT, and ultimately allows for the unrestricted study of multi-component catalyst NPs without previous limitations of length- and time-scale. Most important here is the ability of the MLFF to reliably capturing the coupled phenomena of alloying, activity, and catalyst deactivation as a response to environmental conditions of the catalyst. I also briefly cover some ongoing work pertaining to the use of MLFF for the study of catalyst support materials and their associated quasi-particles in the case of ceria. I then conclude with a brief outlook, wherein I discuss the potential of MLFFs in advancing the design of materials for specific functionalities across diverse fields, paving the way for more controlled and efficient material processing and utilization to address increasing energy demands and catalyst recyclability issues in chemical production.