(699g) Accelerating Inorganic Discovery with Machine Learning and Automation
Computation has emerged as a powerful tool for accelerating the discovery of new materials and molecules: first through first-principles simulation in high throughput screening and very recently even further with machine learning (ML). Unique challenges remain for the accelerated discovery and design of inorganic complexes, despite the highly tunable electronic structure properties that make these materials so compelling for applications in energy storage and catalysis. Density functional theory (DFT) is the method of choice in most computational screening, but DFT is simultaneously too computationally costly and inaccurate for the predictive design of open-shell transition metal complex properties. I will describe our recent efforts to overcome both of these challenges through developing ML models (e.g., artificial neural networks) that predict key first-principles energetic (e.g., spin-state ordering, redox potential, and binding energies) and geometric properties to within the accuracy of the underlying simulation method that provides the training data. I will describe our efforts to use these ML models for extrapolative applications (i.e., design) by incorporating heuristics for ML model uncertainty in a genetic algorithm to discover new inorganic compounds. I will also describe how we are using ML models to overcome increased variability and limited applicability of linear free energy relationships (LFERs), which are widely employed in the solid state, between energies of key intermediates or transition states in a catalytic cycle for single-site and molecular catalysis. Our work on representative catalytic reaction steps in oxidation catalysis demonstrates how our strategies first demonstrated on equilibrium properties such as spin state ordering can be extended to catalysis as well.