Reinforcement Learning for Tailored Development of Aerogels

Abstract

Ever since Kistler developed the first ‘aerogels’, silica aerogels have been the interest of the scientific community due to their exceptional thermal insulation and lightweight characteristics and suitability for diverse applications [1]. Depending on the nature of synthesis and the application, ranging from thermal insulation in high-temperature applications to their application in lithium sulphur batteries, several intrinsic material characteristics may influence the structure-property relationships of the final aerogel product. However, designing aerogels for specific requirements remains a complex task due to the intricate and nanostructured morphology of the material. Given the recent advancements in the areas of materials research and artificial intelligence, deep reinforcement learning (DRL) provides a solution to such optimisation problems for developing aerogels for achieving targeted properties. With the ability to learn and extract complex patterns and relationships, it provides a data-driven approach to understand and optimise these materials. As such, an offline DRL approach in combination with a property predictor (surrogate model) is presented to optimise computationally designed aerogel microstructures for diverse application. The surrogate models act as intelligent digital twins, eliminating the requirement for iterative computational modelling and the subsequent post-processing. These computational microstructures are modelled with aggregation algorithms mimicking the sol-gel chemistry [2] and gaussian random field-based algorithms [3] to optimise the mechanical and the flow properties of the aerogel microstructures. References [1] M. A. Aegerter, N. Leventis and M. M. Koebel, Aerogels handbook, Springer Science & Business Media, 2011 [2] R. Abdusalamov, C. Scherdel, M. Itskov, B. Milow, A.Rege, J. Phys. Chem. B 2021, 125, 1944–1950. [3] C.J. Gommes, A.P. Roberts Phys. Rev. E, 2008, 77, 041403