Delving into machine learning modeling of catalytic reactor system: a case study of steam methane reforming

Navigating the complexity of reactions is one of the fundamental tasks in the field of chemical engineering. Within this domain, identifying reactions in catalytic systems is one of the challenges due to the intrinsic interplay between process variables and reaction mechanisms. In recent years, an increasing number of researchers have examined the application of machine learning in the development of prediction models for identifying reaction mechanisms. Although, their applications in reaction modeling involving kinetics have remained confined, leading to unreliable predictions and limited insights into reaction systems. To address these challenges, we propose a graph ensemble deep learning approach for multiscale modeling that predicts catalytic reaction systems with process variables (e.g., pressure, temperature, reactor size, and flow rate). Our approach includes a tree-based deep learning model to predict conversion of reaction systems governed by distinct mechanisms. To distinguish the mechanism, our model is ensembled with graphical neural networks for inferring correlation between feature and target, such as chemical distribution and reaction condition profiles. We demonstrated the approach through a case study of Ni-based steam methane reforming under varying conditions using a process simulation dataset derived from kinetic equations. The proposed model has optimized hyperparameters by nested k-fold cross-validation. The prediction result of reaction conversion and selectivity shows that our approach can estimate outcomes and effectively explore undiscovered reaction spaces. Furthermore, our preliminary findings illustrate that the proposed approach is applicable to a wide range of reactions involving complex mechanisms without requiring extended experiments for kinetic study.