(435b) Using Machine Learning to Overcome Limitations in Electronic Structure Methodology for Chemical Discovery

Machine learning (ML)-accelerated discovery of materials, i.e., via surrogate models paired with efficient optimization algorithms, holds immense promise to overcome the conventional limitations of computational cost of first-principles electronic structure calculations. Nevertheless, surrogate models inherit the bias of the underlying electronic structure method. In most cases, the electronic structure method of choice is Kohn-Sham density functional theory (DFT), which suffers simultaneously from both self-interaction error or density delocalization error and static correlation error, to varying degrees depending on the density functional approximation. When novel and challenging materials, such as open shell transition metal complexes, is the target of a discovery campaign, few benchmarks are liable to be available for selecting the optimal electronic structure method or DFT functional. Furthermore, investigation of large regions of chemical space (e.g., by varying metal, coordination environment, or oxidation state in a transition metal complex) will likely lead to the conclusion that different electronic structure methods are more suitable for some compounds than others. I will first describe our analysis of the sources of error from the underlying DFT functional. I will describe how we have used consensus from functionals to make more robust predictions of spin crossover behavior, as validated by experimental datasets, and I will describe how we have incorporated both prediction of multi- reference character and DFT consensus to find method-insensitive light harvesting chromophores. Next, I will describe our development of ML-informed density functional models, including a recommender that can identify which DFT functional or parameterization is most predictive to obtain accurate properties of transition metal complexes. I will describe our recent efforts to extend this regress-then-classify strategy to a purely regression-based model that acts in its own right as a density functional. Finally, time permitting, I will describe our efforts in predicting, detecting, and correcting for high multireference character in transition metal complexes as well as some of the datasets we have curated for testing electronic structure methods.