Multi-Reward Reinforcement Learning Based Inter-Atomic Potential Models for Silica

Zeolites are a class of porous materials with wide range of applications, water purification, soil conditioning and catalysis. This inspired a large exploration of hypothetical zeolite configurations but only a handful have been confirmed with even less being synthesized. Owing to the diverse structures of zeolites, access to accurate models capable of predicting the time evolution of structures is necessary to accelerate the discovery and synthesis of zeolite polymorphs. Molecular Dynamics (MD) is a helpful tool to understand the dynamics and phase transformation of zeolite polymorphs but will need an accurate force-field in order to understand the configurational change induced by external stimulus. There are notable interatomic potentials like the buckingham type potentials (BKS, CHIK, Soules) along with the reactive force-field (ReaxFF) used to probe the molecular behaviour of silica. However, they are not capable of modelling both structure and energetics across several established configurations of silica. Existing force-fields are typically fitted to DFT data, and they failed to capture experimental observations. In this talk, we present our work on using a Multi-Reward Reinforcement Learning (RL) to navigate a high-dimensional parameter space and derive an improved set of BKS and Soules force-field parameters to accurately predict the energy, density and structure of 21 different silica configurations. We will discuss the limitations of existing pre-defined force-fields and ways to improve flexibility of the silica potentials.