An interpretable machine learning-based optimization framework for the optimal design of carbon dioxide to methane process

The conversion of carbon dioxide into methane is widely recognized as an effective approach to address the challenges caused by climate change and carbon emissions. However, this process is highly intricate and susceptible to multiple influencing factors, making it challenging to optimize and determine through conventional trial-and-error methods simultaneously. Therefore, an interpretable machine learning-based optimization framework, which integrates the merits of process simulation, exergy analysis, and artificial intelligence approaches and tools, is developed for the optimal design of this process. The bottleneck analysis is conducted through exergy analysis based on the simulated material and energy results of the entire carbon dioxide (CO2) to methane process, revealing that its exergy efficiency is about 80.29 %. Furthermore, it is found that the CO2 methanation reactor exhibits the highest exergy destruction ratio, accounting for 60.57 % of the total destructions. Therefore, this study develops three types of machine learning models for enhancing the performance of the reaction process effectively. Compared with the random forest and deep neural network algorithms, the extreme gradient boosting model has the highest prediction accuracy on the CO2 conversion ratio, methane selectivity, and exergy efficiency of the reactor (with a coefficient of determination > 0.916). The Shapley additive explanations and partial dependence plots analysis are conducted to further identify the most important parameters for improving performance and analyze their impact mechanisms. The comparison with other input parameters highlights that the performance of CO2 methanation systems is primarily influenced by reaction conditions (accounting for 56.3 %) and catalyst conditions, particularly temperature. Finally, the CO(2 )conversion ratio, methane yield, and exergy efficiency of the CO2 to methane process are improved by 3.97 %, 3.06 %, and 1.46 % through multi-objective optimization.