An explainable generative machine learning framework for accelerating CO2 methanation catalyst discovery

Machine learning (ML) is driving a paradigm shift in the development of catalysts for CO 2 methanation. However, traditional ML models lack sufficient internal interpretability, making it difficult to extract mechanistic insights from model predictions; meanwhile, the severe distribution mismatch between the virtual candidate space and the real catalyst domain limits their reliability in the screening and optimization of catalysts. To overcome these limitations, this study develops a framework (EA-GAN) that integrates attention mechanisms with generative adversarial networks (GAN), establishing a closed-loop workflow encompassing performance prediction, candidate catalyst generation, screening, and experimental validation. EA-GAN demonstrates excellent predictive performance in the CO 2 methanation task (R 2 = 0.9049，MAE = 0.0613). By utilizing GAN, this study constructed a virtual library containing 20,000 candidate components, effectively bridging the physical gap between the computational chemical space and the actual experimental space. Subsequent experimental validation of nine representative catalysts demonstrated a high degree of consistency between theoretical predictions and experimental performance. In-depth analysis uncovered a nonlinear relationship between reaction temperature (T) and catalytic performance governed by the law of diminishing marginal returns, identifying support 1 type (S 1 ) as a critical factor regulating marginal effects and high-temperature stability. Based on these findings, this study proposes a dual-design strategy for achieving high CO 2 conversion at low temperatures: the synergistic optimization of intrinsic catalytic activity and the regulation of hydrogen partial pressure. This research not only establishes a robust general framework for catalyst development but also provides a critical reference for the rational design of high-performance methanation catalysts.