Machine Learning Aided Design of Polymer with Targeted Band Gap Based on DFT Computation

Abstract

Polymer band gap is one of the most important properties associated with electric conductivity. In this work, the machine learning model called support vector regression (SVR) was developed to predict the polymer band gap, where the training data of the polymer band gap were obtained from DFT computation while the descriptors were generated from Dragon. After feature selection with the maximum relevance minimum redundancy, the SVR model using 16 key features as inputs gave the optimal performance for predicting polymer band gaps. The determination coefficient (R2) of the SVR model between the DFT computations and SVR predictions of polymer band gaps reached as high as 0.824 for the leave-one-out cross-validation and 0.925 for the independent test. Besides, the 16 key features were explored through correlation analysis and sensitivity analysis. The available model can be used to screen out the polymers with targeted band gaps before experiments, which is very helpful for rapid design of new polymers.