Molecular structure–redox potential relationship for organic electrode materials: density functional theory–Machine learning approach

Abstract

In this study we develop a high-throughput screening method by employing a density functional theory (DFT) - machine learning (ML) framework for the design of novel organic electrode materials. For this purpose, DFT modeling is performed to calculate basic electronic properties of various organic compounds, namely redox potential, electron affinity, highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO), which are used in conjunction with basic molecular descriptors to train three machine learning models (ML): artificial neural networks (ANN), gradient-boosting regression (GBR), and kernel ridge regression (KRR) through three different protocols. These three protocols, or pipelines, are developed in order to enhance each model's capability to learn the data and make predictions. The first two pipelines utilize the original features only, while the third pipeline utilizes composite features which are screened by a least absolute shrinkage and selection operator (LASSO). Particularly, the second and third pipelines employ a Pearson correlation analysis in conjunction with recursive feature elimination (RFE). From this study, the most important features to predict redox potential are identified as the electron affinity and the number of bound Li atoms. After optimizing machine learning models in each pipeline, it is found that KRR predicts the redox potential with the highest accuracy.