Predicting the band gap values of ABX3 type perovskite materials by using machine learning

Abstract

Perovskite material has recently attracted a lot of attention due to the increased demand for renewable energy. Metal halide perovskite material is useful in a variety of sectors. Nonetheless, there remains a requirement for the implementation of a machine learning approach to guide the enhancement of perovskite materials' performance. This study employs machine learning to build design methodologies, anticipate perovskite material performance, and improve material composition. By employing a genetic algorithm and accounting for both the tolerance factor and the octahedral factor, ten stable perovskite structures were determined by evaluating the substitution effects of various alternatives among the seventy-five combinations. Crossover and mutation techniques were employed to ascertain the ten stable perovskite compounds identified at room temperature. With regard to previously evaluated papers, 120 data points were carefully chosen to create the machine-learning models. These models direct the design of contemporary perovskite structures. Utilizing the Support Vector Regression (SVR) and Extreme Gradient Boosting Regression (XG boosting) techniques. Novel compositions of perovskite materials are created using machine learning techniques, and these compositions are subsequently synthesized to assess the effectiveness of the developed models.. It shows a high-stability perovskite composition in iodide halide with Pb2+. Furthermore, NH4GeCl3gives the band gap value as 1.6eV, and NH4GeBr3 gives 1.8eV. These two combinations are newly introduced by this research. A common validation split would allocate 60% of the data for training, 20% for validation, and 20% for testing. Every model‟s prediction accuracy on the test set is assessed using the explained variance (R2 coefficient) and root mean square error (RMSE).The XG boosting model provides higher values of RMSE and R2 coefficients of 0.48 and 0.88, respectively, than the SVR model. The SVR model exhibited accuracy rates of 59%, whereas the XG Boosting model exhibited accuracy rates of 87%. Therefore, according to the results, the SVR model can be considered relatively weak in comparison to the XG Boosting regression model.