Machine Learning for Designing Undesignable Metal-Organic Frameworks

Abstract

Many crucial processes are too complex for computational modeling, requiring experimentation to identify promising materials. Here, a methodology for material design is presented, while photocatalysis is presented as a specific case-study. Metal-Organic Frameworks (MOFs) are a subset of highly promising porous nanomaterials, used in a variety of unmodellable applications. Reinforcement learning generated 60,000 novel MOFs optimized for CO/H20 selectivity. A predictor funnel system was created, iteratively removing low-scoring MOFs to 10,986 potential candidates, improving computational efficiency by 276%. While trained Crystal Graph Convolutional Neural Network (CGCNN) models predicted features for creating a fitness function incorporating stability, catalytic ability, material cost, sustainability, and adsorption while allowing the inclusion of application specific design criterion. This designed function provides a computational method to model photocatalytic performance- and filtered down to two promising MOFs which each pass a myriad of synthesis criteria, first a Cr-based MOF with photocatalyst score 230% higher than the control. Second, a Zn-based MOF outperforms the best control across all relevant metrics, demonstrating robustness against variable fitness functions. This work designed 20 materials, each 125% better than the control for this application. Furthermore, analysis revealed insightful design patterns, such as the significant influence of metal cluster N262 on catalytic performance, providing a method for future work to narrow the chemical space. By incorporating industrially applicable features such as cost or stability of the material, this work successfully designs industrially promising materials in otherwise unmodellable processes such as drug delivery, while paving a method for multi-objective optimization incorporating 260% more features than prior work.