Nanoporous materials discovery for energy and environmental applications using machine learning

Abstract

Nanoporous materials such as zeolites, metal-organic frameworks (MOFs), and covalent organic structures (COFs) are widely used in the fields of gas separation/storage, sensor, catalyst, battery due to their high surface area and tunability. Because of the importance of the porous materials, millions of porous structures have been synthesized experimentally and virtually in recent decades. Various studies have been conducted in the past few years to use machine learning in material science due to the rapid advances of machine learning. However, while machine learning is actively applied to the field of drugs and inorganic solids, The study on machine learning for porous materials remains relatively early stage. In this thesis, we have studied the methodology to develop porous materials that can be used for energy and environmental applications by using molecular simulation and machine learning. In the first, we predict the limits of methane adsorption capacity of zeolite by using the energy shape obtained from molecular simulation and the generative adversarial network (GAN). The study showed that the GAN can predict the performance limits of methane adsorption in zeolite without the information of high performing zeolites. In the second, we have generated zeolite structures by using energy/material shape calculated from molecular simulation and GAN. The study showed that the GAN can generate appropriate zeolite structures and also conditionally generate zeolite structures with the user-desired properties.