Predicting organic structures directing agents for zeolites with conditional deep learning generative model

Abstract

Organic structures directing agents (OSDAs) are important components of the synthesis of zeolites, and their molecular properties play a significant role in the design of zeolite structures. As of yet, scientists have usually used priori knowledge to select suitable OSDAs for synthesis of specific structures of zeolites. We utilize a comprehensive database of OSDAs, zeolites, and gel chemistry to develop a computationally efficient model with fewer parameters based on self-attention mechanism and Long and Short-term Memory networks (LSTM) for modeling the interactions between OSDAs, zeolite structures and gel chemistry. Our model is able to learn the simplified molecules input line-entry system (SMILES) syntax and grasp molecular physicochemical properties by comparing the custom metric and weighted holistic invariant molecular (WHIM) descriptor after dimensionality reduction. Our model is capable of producing OSDA-like molecules under the conditions of zeolite structures and chemical gels, and predicting OSDA molecules for the specific structures of zeolites.