Machine Learning Prediction of Transition-State Energies in Small-Pore Zeolites Combining Acidity and Topology Descriptors

Descriptor-based rational design approach has been extensively developed and applied for the computational discovery of high-performance catalysts, while it is still in its infancy in computational zeolite catalysis. Using toluene methylation with methanol to p-xylene as a model reaction, herein, we developed machine learning (ML) models for the prediction of transition-state (TS) energies in small-pore zeolites, taking both acidity and framework topology into account. General scaling relations between TS energies and the adsorption enthalpy of ammonia (ΔHNH3) as a descriptor of acid strength can be well established within individual zeolite frameworks, while acidity sensitivity varies across framework topologies. In order to capture the effect of the framework, we introduced several readily available zeolite topology descriptors to directly estimate TS energies using different ML models. The tree-based models combining only one acidity descriptor and one topology descriptor, trained on the random-selected data sets, could be used to predict TS energies, but they fail for new topologies. We demonstrate that it is necessary to take the linear regression models and at least four descriptors (e.g., adsorption enthalpy of ammonia, framework density, diameter of the largest included sphere, and accessible volume of zeolites) to achieve an accurate energy prediction on new topologies. The consistency of the coefficients and the robustness of the model are validated. This work thus reveals that no single topology descriptor could quantitatively capture the confinement effect of zeolites, and the proposed scheme in this work combining simple descriptors and ML models could be employed for the rational design of zeolite catalysts.