Predicting the multiple parameters of organic acceptors through machine learning using RDkit descriptors: An easy and fast pipeline

Abstract

Machine learning (ML) analysis has gained huge importance among researchers for predicting multiple parameters and designing efficient donor and acceptor materials without experimentation. Data are collected from literature and subsequently used for predicting impactful properties of organic solar cells such as power conversion efficiency (PCE) and energy levels (HOMO/LUMO). Importantly, out of various tested models, hist gradient boosting (HGB) and the light gradient boosting (LGBM) regression models revealed better predictive capabilities. To achieve the prediction effectively, the selected (best) ML regression models are further tuned. For the prediction of PCE (test set), the LGBM shows the coefficient of determination (R-2) value of 0.787, which is higher than HGB (R-2 = 0.680). For the prediction of HOMO (test set), the LGBM shows R-2 value of 0.566, which is higher than HGB (R-2 = 0.563). However, for the prediction of LUMO (test set), the LGBM shows R-2 value of 0.605, which is lower than HGB (R-2 = 0.606). Among the three predicted properties, prediction ability is higher for PCE. These models help to predict the efficient acceptors in a short time and less computational cost.