Machine Learning Parameters of Optimally Tuned Range-Separated Hybrid Functionals for Transition Metal Complexes

In this work, we present a machine learning (ML) approach for predicting the optimal range separation parameters in transition metal complexes (TMCs), aiming to reduce the computational cost associated with optimally tuned range-separated hybrid (OT-RSH) functionals while preserving their accuracy. A data set containing 4380 TMCs was constructed by screening the tmQM database, with each TMC represented by a 62 087-dimensional multiple-fingerprint feature (MFF) vector and labeled with its optimally tuned range separation parameter. Multiple regression models were applied to train the prediction model, and the support vector machine (SVM) model yielded the best performance. The resulting machine-learned functional, ML-LC-PBE0*, achieves ionization potentials comparable to those obtained from fully tuned LC-PBE0* calculations but at a reduced computational cost. Furthermore, comparisons with DLPNO-CCSD(T) reference values confirm that ML-LC-PBE0* consistently outperforms conventional global and fixed parameter RSH functionals. This work provides an efficient and accurate framework for high-throughput screening and electronic structure prediction of TMCs, paving the way for the discovery of advanced TMC-based optoelectronic materials.