Framework for discovering porous materials: Structural hybridization and Bayesian optimization of conditional generative adversarial network

Abstract

Although deep-learning-based materials discovery has attracted considerable research attention, the application of deep learning has been limited to discovery of materials within single material types such that the discovered materials are similar to existing ones. Thus, we developed an alternative approach for discovering porous materials. A materials discovery design space was configured using several key porous materials and a conditional generative adversarial network (CGAN), which structurally hybridizes them. Hybridization was controlled using a vector design variable input as a condition to the CGAN, and each design variable component represented a key porous material intensity, which was defined as a conceptual quantity for expressing materialness. Furthermore, by varying the vector latent variable input to the CGAN, any number of similar hybrid porous materials could be generated using the same design variable. The multiobjective Bayesian optimizer efficiently discovered hybrid porous materials constituting Pareto solutions in the objective function space of the material properties, which were mapped from the design space by the CGAN structural hybridization and computational fluid dynamics property evaluation. Tradeoff properties such as the pressure drop and filtration efficiency were multiobjectively optimized for key porous materials exhibiting random, packed-sphere, and sponge structures. By flexibly adopting their structural parameters depending on the required pressure drop and filtration efficiency, the discovered optimized hybrid porous materials outperformed the key porous ones. The results demonstrated the effectiveness of the proposed framework for discovering porous materials.

• Previous article in issue
• Next article in issue

Keywords
Porous material
Permeability
Filtration efficiency
Conditional generative adversarial network
Computational fluid dynamics simulation
Multiobjective Bayesian optimization

Description