(202h) Machine Learning Zeolite Synthesis Conditions

Zeolites are commonly used as catalysts and absorbents for various chemical applications, but the conditions to synthesize zeolites with target structures and thus properties remain elusive given the large synthesize design space. Machine learning (ML) methods provide a viable way to elucidate the relationships between synthesis conditions and the output zeolite structures and properties.

This study investigates the use of ML methods, particularly random forest and XGBoost, in discovering the optimal synthesis conditions for zeolite EMT. Using experimental datasets of EMT synthesis we first conducted a feature importance analysis using Shapley Additive exPlanations (SHAP) to rank all synthesis parameters. By perturbing the datasets, the model effectively pinpointed specific synthesis conditions conducive to EMT formation. Additionally, we utilized principal component analysis (PCA) and autoencoder for dimensionality reduction, which allowed us to retain three components for the development of a new predictive model. The model was employed to predict further synthesis conditions favorable for EMT formation. This method led to the identification of multiple synthesis conditions of producing EMT different from the experimental datasets and the synthesis conditions were confirmed using experiment to confirm the model prediction validity.