Time-dependent dielectric density functional theory predictions for excited-state molecular properties using neural networks guided by lower-level quantum chemistry

Abstract

This paper discusses a neural network approach for the prediction of ground- and excited-state molecular properties. The time-dependent dielectric density functional theory was employed to handle the excited state of coumarin-152 dye. We demonstrated that lower-level quantum chemistry techniques, such as the semiempirical method and the Hartree–Fock approximations, can be used to improve the predictive performance of neural network models. The semiempirical quantum chemistry method provided excellent feature vectors (descriptors) on excited-state molecular properties in terms of computational cost and performance. In addition, a vector-based descriptor to improve molecular-force predictions is discussed.