Application of artificial neural networks for modeling of electronic excitation dynamics in 2D lattice: Direct and inverse problems

Machine learning (ML) approaches are attracting wide interest in the chemical physics community since a trained ML system can predict numerical properties of various molecular systems with a small computational cost. In this work, we analyze the applicability of deep, sequential, and fully connected neural networks (NNs) to predict the excitation decay kinetics of a simple two-dimensional lattice model, which can be adapted to describe numerous real-life systems, such as aggregates of photosynthetic molecular complexes. After choosing a suitable loss function for NN training, we have achieved excellent accuracy for a direct problem-predictions of lattice excitation decay kinetics from the model parameter values. For an inverse problem-prediction of the model parameter values from the kinetics-we found that even though the kinetics obtained from estimated values differ from the actual ones, the values themselves are predicted with a reasonable accuracy. Finally, we discuss possibilities for applications of NNs for solving global optimization problems that are related to the need to fit experimental data using similar models.

Automated summary

Machine learning approaches have gained significant attention in the chemical physics community for their ability to predict numerical properties of molecular systems with minimal computational cost. In this study, we evaluated the suitability of deep, sequential, and fully connected neural networks for predicting the excitation decay kinetics of a simple two-dimensional lattice model. We achieved excellent accuracy in predicting lattice excitation decay kinetics from model parameter values, and reasonably accurate predictions of the model parameter values from the kinetics themselves. This research also explores the potential of neural networks in solving global optimization problems related to experimental data fitting using similar models.