Accelerating first principles inorganic materials design using machine-learned surrogate models

Transition metal complexes are a promising class of tunable materials with applications in sensors, molecular electronics and homogeneous catalysis. However, choosing metal and ligand field combinations from a discrete pool of candidates naturally leads to a combinatorial material design problem posed over an enormous chemical space. While first principles simulation can help screen this design space for target properties, even density functional theory (DFT) calculations are sufficiently expensive to severely limit the range of materials that can be studied. This is confounded by uncertainty about appropriate functional choice and DFT reliability as well as the incompatibility of notions of organic chemical similarity with inorganic systems. Machine learning (ML) has the potential to overcome these challenges by providing ultra-low cost approximate models that connect novel materials to properties at accuracies competitive with the baseline uncertainty in first-principals calculations. By supplementing high-throughput DFT screening with data-driven surrogate models, we can increase drastically the variety of chemical configurations that can be studied. However, the application of ML models to material design across highly diverse chemical spaces requires understanding of the transferability of these models to materials very unlike their training data. We obtain reasonable estimates of model uncertainty and domain of applicability by considering relative similarity to training data, providing a simple strategy of intelligently allocating DFT calculations either for lead extraction or activate learning. Equipped with uncertainty-aware surrogate models, we are able to adapt a simple genetic algorithm for inorganic materials design, and show good performance in the design of materials with targeted quantum mechanical properties from large design spaces: 1) identification of spin cross-over materials with <2% training data, with ~2/3 leads being validated at a DFT level and 2) design of complexes with targeted frontier orbital energies with good agreement with DFT with <6% training data coverage.