Bio-Digital Catalyst Design: Generative Deep Learning for Multi-Objective Optimization and Chemical Insights in CO<sub>2</sub> Methanation
Runjie Bao, Zhao Wang, Qiwen Guo, Xiaoyu Wu, Qingchun Yang
Abstract
To address the limitations of traditional trial-and-error experimental methods and deep learning algorithms in the design of CO 2 methanation catalysts, this study proposes a dual-driven deep learning (BODAL) strategy. This approach aims to enhance prediction accuracy by leveraging bioinspired optimization algorithms for adaptive hyperparameter configuration and employing generative techniques for data augmentation. By systematically comparing six bioinspired optimization algorithms, the particle swarm optimization algorithm is identified as the most effectively automatic hyperparameter configuration tool and combined with the FT-Transformer (FTT) deep learning model to achieve high-precision prediction of the CO 2 conversion ratio and CH 4 yield. Compared with numerous generative data augmentation strategies, the generation of synthetic data by the Tabular Variational AutoEncoder approach is a more effective approach. It significantly alleviates the limitations of small sample data by iteratively enhancing the preferred FTT model, achieving a test set R 2 of 0.9591. The global and local interpretability analysis using SHAP and partial dependence plots reveals the contributions of 23 input variables to the model’s predictive performance, as well as the regulatory mechanisms of the most significant discrete and continuous variables on the catalytic performance of the CO 2 methanation reaction. The optimized FTT model is further integrated with the multiobjective particle swarm optimization algorithm, successfully optimizing the content of the active component and the reaction temperature for the widely used Ni/Al 2 O 3 catalyst. Additionally, it facilitated the prediction of six novel, highly efficient Ni-based catalysts. Notably, Ni–Y/ZrO 2 –La 2 O 3 achieves a CH 4 yield of 90.30% with an active component content of approximately 5.35%, while Ni–Gd/Al 2 O 3 –Pr 2 O 3 demonstrated a high CH 4 yield (59.89%) at a low temperature of 210 °C, surpassing the experimentally reported value of 52.44%.