AI Toolkit for Accelerated Nanoporous Material Discovery

Abstract

Nanoporous materials are promising candidates for carbon capture, due to their large surface area and highly porous structure. There is a vast number of constituent materials which can be synthesised, enabling precise tuning of various chemical and geometric properties. In this work, we present an AI toolkit aimed at improving the discovery of materials such as zeolites, metal- organic frameworks, and zeolitic imidazolate frameworks with good adsorption of greenhouse gases, such as CO2. The toolkit consists of 4 software packages that respectively implement: (1) a linear regression pre- screening of candidate materials, (2) a crystalline graph convolutional neural network that acts as a computationally cheaper surrogate model for in-silico adsorption simulations, (3) symbolic regression techniques for learning the adsorption isotherm as a function of pressure and temperature, and (4) a black-box optimisation technique for efficiently discovering a near-optimal material within a large database.