Ensemble Machine Learning for Thickness‐Dependent Bandgap Prediction in Perovskite Solar Cell Thin Films
TL;DR
Feature attribution via SHAP, MDI, and permutation importance (PI) consistently identifies perovskite thickness, the engineered effective absorbance term (BG × Thickness × PCE), and the Br/I halide mixing ratio as the dominant predictors.
Abstract
This report presents a comprehensive ensemble machine learning study for predicting the Bandgap × Thickness product (eV·nm), a key opto‐electronic figure of merit for perovskite solar cell absorber layers. A dataset of device records is featurized using physics‐informed descriptors encoding device architecture‐related composition fractions with layer‐stack interactions, RDKit molecular descriptors for perovskite organic A‐site cations (methylammonium, formamidinium) with halide anions as fraction‐weighted averages, and Mordred extended descriptor vectors for the same chemical components. Six ensemble regressors are trained and benchmarked as Random Forest, Extra Trees, XGBoost, LightGBM, a Voting Ensemble, and a Stacking Ensemble with a Ridge meta‐learner. Among them, XGBoost achieves the highest test R2 of 0.98 (RMSE = 14.28 eV·nm and MAE = 5.98 eV·nm), while the Stacking Ensemble delivers the lowest MAE of 4.79 eV·nm. Feature attribution via SHAP, MDI, and permutation importance (PI) consistently identifies perovskite thickness, the engineered effective absorbance term (BG × Thickness × PCE), and the Br/I halide mixing ratio as the dominant predictors. Mordred descriptors contributed ~20.5% of the total MDI importance to demonstrate molecular fingerprinting at device‐level datasets.