Application of Neural Networks for Investigating Potential Energy Surfaces of Electronically Excited States and Internal Conversion Mechanisms of Organic Molecules

Based on non-empirical data obtained using a high-level multi-configurational quantum chemistry method, neural networks with the architectures of a multilayer perceptron and an E (3) -equivariant graph network were constructed and trained to predict the energies of the ground and the first two electronically excited states of the methaniminium cation, C{{H}₂}NH₂^ +. It is shown that the E (3) -equivariant graph neural network architecture demonstrates higher accuracy. Using the trained network, a segment of the cation’s potential energy surfaces near the region of the conical intersection between the first excited and the ground states was investigated; this region plays an important role in the mechanism of internal conversion and photoisomerization reactions. It is demonstrated that the neural network accurately reproduces the topography of the potential energy surfaces of the two electronic states in the region of their conical intersection.