Charge Transfer Simulations using Hamiltonian Elements and Forces from Neural Networks

TL;DR

This paper demonstrates that neural network-based Hamiltonians can accurately simulate charge transport in organic semiconductors, matching traditional methods while reducing computational costs.

Contribution

The study introduces neural network models that predict Hamiltonian elements and forces, enabling efficient charge transport simulations in organic semiconductors.

Findings

NN-based Hamiltonians accurately reproduce hole mobilities

Models trained on DFTB-quality data are effective

Significant reduction in computational cost achieved

Abstract

The trajectory surface hopping method has been widely used in the simulation of charge transport in organic semiconductors. In the present study, we employ the machine learning (ML) based Hamiltonian to simulate the charge transport in anthracene and pentacene. The neural network (NN) based models are able to predict not just site energies and couplings but also the gradients of the site energy as well as off-diagonal gradients necessary for forces. We train the models on DFTB-quality data for both anthracene and pentacene. By using the obtained models in propagation simulations, we evaluate their performance in reproducing hole mobilities in these materials in terms of both quality and computational cost. The results show that the charge mobilities obtained using the NN-based Hamiltonian are in very good agreements with the charge mobilities computed using the DFTB-based Hamiltonian.