Predicting the properties of perovskite materials by improved compositionally restricted attention-based networks and explainable machine learning

Abstract

Understanding the unique properties of perovskite materials is crucial in advancing solar energy technologies. Factors like heat of formation and bandgap significantly influence the light absorption capability and stability of perovskite solar cells. However, it is time-consuming and labor-intensive to obtain the properties of perovskites using traditional experimental or high-throughput computational methods. As a prospective method, machine learning can find regularities in the given training data and give accurate prediction results. In this article, we use deep learning models based on attention mechanisms and elemental features to predict the heat of formation and bandgap of perovskite materials. Random Forest and Gradient Boosted Regression Tree models have also been used for interpretable predictions of properties. The compositionally restricted attention-based network was improved by introducing a densely connected network and optimizing the network structure to increase data processing capabilities. The experiment results show that the mean absolute errors of the heat of formation and bandgap on the test sets are decreased by 5.77% and 3.37% respectively. The optimized model also shows better performance when used for classification tasks. In addition, we use the Gradient Boosting Regression Tree model and the Shapley Additive Explanations tool to conduct an interpretable analysis, explaining the impact of different features on the predictions of the properties.

Predicting photovoltaic parameters of perovskite solar cells using machine learning

“…Whereas the r-value is flat or decreases, we infer that this is due to the different bandgap values for the same perovskite composition in the dataset. In our recently published work, we used deep learning to predict properties such as bandgap of perovskite materials, and since the data sources were calculated by density functional theory simulations, the accuracy of the predictions was slightly less accurate due to intrinsic errors in the dataset [42]. However, all the bandgap values herein are obtained experimentally, contributing to the accuracy of predictions and making the error relatively low [43].…”