Deep Learning Methods for Catalyst Surface and Interface Structure Analysis

The increase in global energy demand and raised environmental concerns have motivated the design of novel materials for energy-related applications. However, the design of ever-complicating materials for emerging energy technologies is currently bottlenecked by limited resources to understand complex surface and interface structures and property relationships. In the first part of this thesis, we develop a tandem framework that combines a molecular thermodynamic theory and molecular dynamics simulations in an attempt to investigate solid interfacial phenomena and to discuss how deep learning methods can improve the framework as a next step. In the second part, we develop a set of deep learning methods that solve various materials and catalyst design problems including property, structure, and stability analysis. We present a graph neural networks architecture to learn the optimal representations of heterogeneous catalysis systems for the accurate prediction of adsorption/binding energies. Then we extended the approach to approximate ground-state structures of the catalysis systems by incorporating differentiable optimization methods into the graph neural networks architecture. We further develop a general deep reinforcement learning framework to identify the metastability of alloy catalyst surfaces by exploring possible surface reconstructions and their associated kinetic barriers under reaction conditions. With these advanced data-driven methods that understand the surface and interfacial phenomena, we open up new avenues for accelerated materials and catalyst discovery.