Application of Graph Methods in Computational Design of Kinematically Active Molecules.

We have developed a graph editing approach for automated computational design of organic molecules. This approach formalizes the design space as a decision tree which encompasses trillions of unique molecule designs after only a handful of graph editing choices. As this design space is too large to search using brute force computing, we are working to establish machine learning (ML) schemes for accelerated automated molecule design by forecasting the properties of molecules. We are specifically interested in predicting the mechanical and kinematic behavior of linker molecules for potential metal-organic frameworks. This poses a particular challenge because kinematic and mechanical behavior depend both on the local chemical moieties present in a molecule, and their globale geometric arrangement relative to loading, and thus schemes for encoding (embedding) the structure of molecules as input for ML methods must capture all relevant information. We present a comparison of the performance of five different embedding schemes for prediction of vibrational and kinematic behavior. To do this we used our automated design process to generate a pool of >50,000 dicarboxylate molecules as potential candidate ditopic ligands. Molecular dynamics simulations were performed to compute each candidate’s vibrational and kinematic behavior. The five embedding schemes were used to train and test various ML schemes' ability to predict molecules’ stiffness, folding, and vibrational behavior. Each of the embedding methods have strengths and weaknesses and no one embedding method is a silver bullet, but when used together the five methods provide a significant predictive ability.

__________________________________________________________________________________________________________

Biography

Dr. Greaney gained a MS in Metallurgy and the Science of Materials from Oxford University and his PhD in Materials Science and Engineering from UC Berkeley. He completed postdoctoral training at UC Berkeley and MIT before becoming an Assistant Professor at Oregon State University in 2011. He moved to UC Riverside in 2015. His research group uses theory and modeling to study the structure-property relations underpinning the mechanical, thermal and electronic properties of materials for application in energy conversion, energy storage and clean water.