Abstract
Leveraging our recent development, which incorporates hole and particle information into the multi-channel molecular orbital image (MolOrbImage), to generate exceptional accuracy (mean absolute error, MAE < 0.1 eV) in predicting excited-state energies of practical photofunctional materials containing several hundred atoms, we have advanced the implementation of a new approach to overcome the high computational cost of mean-field ground-state calculations that limits its application in high-throughput materials discovery. In this work, low-cost approaches for generating approximate orbitals, including the superposition of atomic densities technique and the semiempirical tight-binding method, have been employed to construct cost-effective multi-channel MolOrbImages. By connecting with a convolutional neural network, the performance of our model is evaluated for both small organic molecules (MAE < 0.1 eV) and practical photofunctional materials (MAE < 0.14 eV). Perturbation analysis of MolOrbImages highlights the importance of frontier orbital energies, which further motivates the adoption of transfer learning techniques to reduce prediction errors in excited-state energies.