Molecular descriptor-assisted interpretable machine learning: A scheme for guiding the synthesis of zeolites with target structures

Abstract

Zeolites, with their ordered channel structure, find extensive applications in the petroleum industry and environmental protection. However, the complex nucleation and crystallization processes of zeolites pose challenges for the efficient synthesis of novel zeolites such as the extra-large pore size zeolites (ELPZ). Due to the potential of germanosilicate zeolites to synthesize ELPZ and low framework density (FD) zeolites, we construct machine learning (ML) models for pore size classification and FD prediction. We present a comprehensive and efficient OSDA featurization using weighted holistic invariant molecular (WHIM) descriptors, which better links the synthesis conditions to the structures of germanosilicate zeolites. By employing different interpretable machine learning methods, we elucidate the influence of synthetic descriptors on zeolite structure and determine key experimental conditions conducive to the synthesis of ELPZ and low-FD zeolite. Furthermore, we introduce an assignment method to extend SHapley Additive exPlanations (SHAP) to the molecular properties described by WHIM, thereby enabling the understanding of the impact of OSDA structural characteristics on resulting zeolites. We provide targeted optimization suggestions for a single experimental condition through a comparison of local interpretations for different samples, which are verified by the predictions of the model.