Transition metal complexes (TMCs) are promising homogeneous catalysts, offering both higher selectivity and better metal atom economy than their heterogeneous counterparts. However, due to the wide space of potential ligands, there is a combinatorially large design space of TMCs that is experimentally intractable. To identify the most promising candidates, computational workflows have been developed, many of which rely on density functional theory (DFT). DFT is a workhorse method in the community, but its accuracy is highly dependent on the choice and parameterization of density functional approximation (DFA) used. DFA errors on TMCs can be significant across a range of properties from barrier heights to dissociation energies, but prediction of ground state spin, which is highly relevant for catalysis, is especially problematic. In this work, we develop machine learning (ML) models to predict the optimal amount of Hartree-Fock exchange (HFX) to include in a DFA such that it approximates the spin-state ordering of DLPNO-CCSD(T), our chosen correlated wavefunction theory reference. We develop new datasets examining the dependence of spin-splitting energy on HFX, named VSS-452-HFX and CSD-76-HFX, and demonstrate that system-specific reparameterization of HFX allows for both the PBE and SCAN functionals to reproduce DLPNO-CCSD(T) references in almost all cases for CSD-76-HFX, a set of experimentally synthesized TMCs. Finally, we train ML models to predict the amount of HFX to include in both PBE and SCAN such that they approximate DLPNO-CCSD(T). We find that models utilizing the electron density can predict optimal HFX values well enough to outperform any single DFA, and are much more generalizable and accurate than simpler models based on TMC connectivity and atomic features or matching orbital energies to ionization potentials and electron affinities. These models enable DLPNO-CCSD(T) level accuracy at DFT-level cost, allowing for higher accuracy computations to be used in computational discovery workflows.
