Linear-response time-dependent density functional theory (LR-TDDFT) provides reliable predictions of excited-state properties but remains computationally expensive for large molecules and high-throughput screening. In contrast, its semi-empirical alternatives like the simplified Tamm−Dancoff approximation (sTDA) offer substantial efficiency gains but suffer from reduced accuracy, which is largely attributed to the globally fitted and fixed parameters. In this work, we systematically show that tuning the specific Fock-exchange mixing parameter for each molecule across the extensive QCDGE dataset significantly improves the sTDA accuracy for both singlet and triplet excitations. To eliminate the individual parameter optimization for each molecule, we developed a machine learning (ML) model to predict optimized parameters directly from molecular structure using multiple fingerprint features. The ML model achieves high predictive accuracy (mean absolute error R2 > 0.96), and further validation on a GDB-17 subset confirms its robustness and generalizability, highlighting its practical utility in excited-state calculations.
Wang et al. (2026) studied this question.