BENCHMARKING MACHINE LEARNING ALGORITHMS FOR BANDGAP PREDICTION IN HALIDE PEROVSKITES: A COMPARATIVE STUDY OF 540 COMPOSITIONS

Аннотация

This study presents a systematic benchmarking of seven distinct machine learning
algorithms for predicting the bandgap of halide perovskites. Using an adapted dataset of 540
perovskite compositions with the general formula ABX3, the predictive performance of models
ranging from linear methods to ensemble techniques was evaluated. The analysis reveals that
ensemble methods, particularly XGBoost, significantly outperform linear and kernel-based
approaches, achieving a coefficient of determination R2 of 0.92 and a root mean square error (RMSE)
of 0.35 eV on test data. Feature importance analysis identified the Goldschmidt tolerance factor and
B-X electronegativity difference as the most critical predictors. This work demonstrates the superior
capability of gradient-boosted decision trees for capturing complex structure-property relationships
in perovskite materials.