Machine learning Frenkel Hamiltonian parameters to accelerate simulations of exciton dynamics

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Site-based models are crucial for understanding exciton dynamics in condensed phase systems.
• Parameterizing these models accurately requires detailed molecular information, often computationally intensive to obtain.
• Existing methods struggle to efficiently encode complex molecular morphologies into model parameters.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for accelerating the parameterization of site-based exciton dynamics models.
• To investigate efficient strategies for predicting molecular parameters from all-atom configurations of sexithiophene systems.
• To improve the accuracy and generalizability of predicting excitation energies and intermolecular couplings.

Main Methods:

• Development of multiple machine learning models, including kernel ridge regression (KRR) and deep neural networks.
• Utilizing Coulomb matrix featurization for predicting excitation energies.
• Employing bi-molecular featurization and unimolecular transition densities for predicting intermolecular couplings.
• Analysis of single-molecule excited-state electronic-structure calculations.

Main Results:

• A KRR model with Coulomb matrix featurization accurately predicts excitation energies.
• A deep neural network model for direct prediction of intermolecular couplings performed poorly.
• An indirect ML approach using KRR to predict unimolecular transition densities for coupling computation showed excellent performance.
• The indirect approach requires a smaller feature space and integrates molecular physics insights.

Conclusions:

• Machine learning can significantly accelerate the parameterization of exciton dynamics models.
• Indirect prediction of bimolecular properties via higher-order unimolecular properties is a highly effective ML strategy.
• This approach offers a generalizable and computationally efficient pathway for modeling complex molecular systems.