Data-driven performance prediction of polymer solar cells using multi-layer perceptron with Bayesian hyperparameter tuning

Abstract

Polymer solar cells (PSCs) rely on labor-intensive trial-and-error experiments to optimize power conversion efficiency (PCE), due to complex coupling of molecular electronic properties, blend morphology and fabrication parameters. This work develops an interpretable prediction framework integrating multi-layer perceptron (MLP) and Bayesian optimization (BO) to map multi-dimensional material descriptors to device PCE. A dataset of 1842 PSC records with 48 material and process features is preprocessed via missing value imputation, standardization and one-hot encoding. Dropout and L2 regularization are applied to suppress overfitting, and BO auto-tunes model hyperparameters within 50 trials. Comparative tests against SVR, random forest and shallow CNN show the MLP-BO model achieves a test R2 of 0.962 and MAPE below 3%, outperforming all baselines. Permutation feature importance identifies optical bandgap, donor–acceptor ratio and LUMO offset as dominant efficiency factors, consistent with photovoltaic fundamentals. The framework retains high accuracy under variable light and temperature, and supports inverse design to screen optimal blend systems for target PCE, offering an efficient tool for rapid organic photovoltaic material discovery.