(195c) Machine Learning for the Prediction of Electronic Couplings and Charge Transport Calculations in P3HT

Predicting which organic chemical blends will make desirable solar cell materials requires a suite of computational chemistry tools including molecular dynamics simulations, quantum chemical calculations, and kinetic Monte Carlo simulations. In this work, we focus on ameliorating a quantum chemical calculation bottleneck in the pipeline between chemical ingredients and the charge mobilities that results from the self-assembled structure in these soft materials. Here we focus on the benchmark polymer poly(3-hexylthiophene) (P3HT) and calculating the dependence of charge mobility on P3HT nanostructure with kinetic Monte Carlo simulations that take charge hopping rates between chromophores as inputs. We compare five machine learning techniques for correlating structural features of P3HT monomers with the electronic couplings calculated with Zerner's method of intermediate neglect of differential overlap for spectroscopy (ZINDO/S). We find that random forests are an easy-to-implement, accurate predictor of couplings that finds the separation and orientation of bonded and non-bonded P3HT monomers chiefly determine their coupling. We discuss tests informing the minimal training set needed to generate random forests with 0.02 mean absolute error that is 400 times faster than performing new ZINDO/S calculations. We find that the charge mobilities predicted with the random-forest-generated couplings to be in agreement with the ZINDO/S mobilities. We conclude with an overview of the pipeline in which these random forests are used and how they are enabling broader screening of organic photovoltaic candidate chemistries.