Visualizing nexus of porous architecture and reactive transport in heterogeneous catalysis by deep learning computer vision and transfer learning

Reactive transport in porous media is the key to heterogeneous catalysis, which is the central process in both natural and engineered systems. Elucidating nexus between porous architecture and reactive transport is of importance, but remains a challenge. Conventional text-based approach relies on quantitative structural features (QSFs; porosity, tortuosity, and connectivity), which fails to identify key reaction regions and predict local reaction rate for anisotropic architecture due to isotropic assumption. To address these issues, this study reports a data-driven deep learning computer vision (DLCV) method for visualizing nexus between porous architecture and reactive transport in heterogeneous catalysis. Here, we show that the 3D local reaction rate can be inferred from 2D lateral images of anisotropic porous catalysts using Conditional Generative Adversarial Network and feature representation transfer learning (cGAN-FRT). Efficiency and generalizability are validated by rapid and accurate prediction of reaction rate for heterogeneous electrocatalysis. Based on feature importance generated by cGAN-FRT, pore throat, curved flow channel, and their combined structures are identified to be the dominant factors that affect nonlinear variation of porous reactive transport, which can be interpreted by physical field synergy. This study realizes visualizing nexus between anisotropic porous architecture and local reactive transport powered by artificial intelligence.