Graph Convolution Neural Network for predicting Photo-excited State Properties of Indole C8H7N Absorption Spectra

Abstract

Absorption spectroscopy is a very important tool in chemistry which predicts which part of the electromagnetic radiation is absorbed by a molecule, giving a unique signature for each molecule. All photo-responsive biological phenomena (i.e., photosynthesis, photocatalysis, UV absorption of DNA leading to skin cancer, and fluorescence in biological systems) are initiated by absorption of light, so many recent efforts have been shifted towards accurately modeling the spectra of molecules, where the collective vibrational effects have been identified to significantly affect the spectra shape in electronic spectroscopy. With the growing promise of deep learning in predicting properties in photo-excited molecules, this research explores the feasibility of applying the SchNet deep learning model to predict excited states properties of a molecule with added complexity outside the popular QM9 and MD17 datasets. We have found the graph convolutional structure of SchNet to account for the molecular dynamics requirements and reproduce the spectra with significantly less time and computational resources.