Machine learning-assisted designing of compounds with higher glass transition temperature. Chemical space visualization and synthetic accessibility determination

Abstract

Recently, advanced machine learning approaches have been implemented for the selection of suitable materials for various applications. In the present study, compounds are designed with higher glass transition temperature. 40 learning models were applied and tested to accurately predict the glass transition temperature and the hist gradient boosting (HGB) regressor was found to be the best model for the prediction (R-squared values of 0.983 and 0.866 for training and test set, respectively). 5000 compounds are designed and their glass transition temperature is predicted using the hist gradient boosting (HGB) regressor. The generated database of compounds is visualized using cluster analysis. Thirty compounds with the highest glass transition temperature are selected. The synthetic accessibility of the designed compounds is also predicted. Chemical similarity analysis is conducted to gain insights into the behavior of the compounds. For this purpose, clustering and heatmap techniques are employed. Our framework has the ability to screen efficient materials in a fast, cheaper and easy way.