Machine learning-guided scalable manufacturing of high-efficiency perovskite solar modules

Based on an XGBoost multi-target prediction model, this code optimizes the device parameters for perovskite solar cells. By performing a grid search across the combinatorial space of cation (Cs/FA/MA) and halide (Br/I/Cl) components, it identifies the optimal formulation while keeping the device structural parameters fixed; furthermore, it applies a correction function to calibrate the predicted efficiency based on specific processing conditions. Additionally, for a given fabrication process and fixed composition, the code scans and plots the efficiency response curve as a function of device area. This code constructs a four-target prediction model for perovskite solar cells based on XGBoost. The data preprocessing module normalizes the cationic (Cs/FA/MA) and halide anionic (Br/I/Cl) components, applies one-hot encoding to categorical variables—such as additives and passivators—and imputes missing values in numerical features using the median. The model employs an XGBoost regressor with fixed hyperparameters for training. For efficiency prediction, the code reads new device parameters from an external file, automatically imputes any missing feature columns, and then invokes the PCE model to perform inference, thereby enabling the rapid prediction of the Power Conversion Efficiency (PCE) for unknown samples. This code constructs a multi-objective predictive model for perovskite solar cells based on XGBoost, performing regression predictions for PCE, Jsc, and FF, and a three-class classification prediction for Voc. It mitigates data imbalance issues through 2D joint sample weighting and enhances predictive accuracy by combining RandomizedSearchCV-based hyperparameter optimization with isotonic regression post-processing. Furthermore, SHAP analysis is introduced to quantify the feature contributions for the four target variables, while a material-grouping SHAP analysis is employed to reveal the influence patterns of the materials within each functional layer.