Performance Evaluation of Machine Learning Algorithms for Predicting Organic Photovoltaic Efficiency

This study forecasts the power conversion efficiency (PCE) of organic solar cells using data from experiments with donors and non-fullerene acceptor materials. We built a dataset that includes both numerical and categorical features by using standard scaling and one-hot encoding. We developed and compared several machine learning (ML) models, including multilayer perceptron, random forest, XGBoost, multiple linear regression, and partial least squares. The modified XGBoost model performed best, achieving a root mean squared error (RMSE) of 0.564, a mean absolute error (MAE) of 0.446, and a coefficient of determination (R2) of 0.980 on the test set. We also assessed the model’s ability to generalize and its reliability by examining learning curve trends, calibration curve analysis, and residual distribution. Plots of feature correlation and permutation importance showed that ionization potential and electron affinity were key predictors. The results demonstrate that with proper tuning, gradient boosting methods can provide highly accurate and easy-to-understand predictions of organic solar cell efficiency. This work establishes a repeatable machine learning process to quickly screen and thoughtfully design high-efficiency photovoltaic materials.