UV-Vis absorption spectra using semi-empirical Hamiltonian plus a machine learning model

Machine learning (ML) is becoming a powerful tool for obtaining accurate properties of materials at a reduced computational cost. This work shows that semi-empirical Hamiltonian, INDO/s + machine learning models (INDO/s+ML) could give high-level first principle results. The machine learning models were used to add corrections to INDO/s absorption spectra to produce TDDFT absorption spectra. Excitation energy and oscillator strength corrections for 16k+ organic molecules were learned using the Kernel Ridge Regression (KRR) and Neural Network (NN) models. The INDO/s+ML predicts TDDFT excitation energies and oscillator strengths within MAE/RMSE of 0.15/0.23 eV and 0.04/0.11, respectively, for 3600 organic molecules not included in the training. The calculated spectra for these molecules with INDO/s+ML agree well with those from TDDFT.