(202j) Integrated Use of Explainable AI and Constrained Generative Adversarial Networks for Novel Design of Redox-Active Organic Molecules for CO2 Capture

With the climate crisis more imminent as ever, various methods for mitigating CO₂ is in great demand. While amine-based CO₂ capture has been considered as an efficient method for capturing CO2 from fluegas, it has limitations in use for direct air capture (DAC) owing to volatility of the amines and degradation. As an alternative, solid sorbents such as amine-functionalized sorbents and metal-organic frameworks are being used as materials for DAC, but significant amount of energy is required at the current technology level, resulting in very high costs making it less profitable for companies to fully invest in it. Since DAC is a net-negative technology in terms of CO₂, it is considered a key technology for meeting the CO₂ emission goals, and a breakthrough is required to contribute to actual CO2 mitigation.

Recent studies have made use of redox-active organic materials (ROMs) as electrochemical CO₂ capture agents. Such studies have made use of ROMs such as NeutralRed and 4,4-Azopyridine for capture of CO₂, showing regeneration energy of 35 kJ/mol of CO₂, which is comparable to state-of-the-art amine-based capture systems. While these ROMs have shown potential for DAC, ROMs tend to be sensitive towards oxygen, which limits their usage. To make use of ROMs for DAC, various characteristics need to be met, such redox potential, solubility, CO₂ affinity and O₂ insensitivity. However, a ROM meeting all of these criteria is yet to be found, and efficient analysis of the molecular characteristics and synthesis of these materials is an essential field of study to expand the usage of ROMs.

In this study, explainable AI is paired with constrained generative adversarial networks (cGAN) to analyze the characteristics of various ROMs, and to synthesize novel molecules with preferred properties. Based on the Murcko fragmentation rule, the attributions of the substructures of the ROMs included within the RedDB with respect to DAC-preferred characteristics are analyzed by substructure masking. This involves the development of a graphical isomorphism network for predicting the characteristic values of a ROM, and masking of the substructures to evaluate its effect on the ROM characteristic. After defining the essential substructures of ROM for effective DAC, a cGAN model is developed to generate novel ROM molecules. Upon acquiring the ROM with optimal characteristics for DAC, the ROM is synthesized and tested out for DAC performance. Air is flown through an H-cell with the synthesized ROM, and energy values for electrochemical stripping of CO₂ is evaluated.