The frontier molecular orbitals of organic semiconductor materials play a crucial role in the performance of photoelectric devices, including organic photovoltaics (OPVs), organic light-emitting diodes (OLEDs), and organic photodetectors (OPDs). In this work, a model for predicting frontier molecular orbital of organic materials, including HOMO and LUMO levels, is established with the extreme gradient boosting algorithm and Klekota-Roth fingerprints. The correlation coefficients of HOMO or LUMO energy levels in the testing set are 0.75 and 0.84 in the transfer model from 11,626 DFT data in Harvard Energy database to 1198 experimental data in literature. The difference between the ML predicted value and the experimental value is smaller than the difference between ML prediction and DFT calculation, always less than 10%. Moreover, based on correlation and SHAP interpretability analysis, 13 key structural fragments influencing energy levels are selected to further verify the effective regulation of the frontier molecular orbital by the key structural fragments in practical applications. Considering the completely opposite regulatory functions of key structural fragments on HOMO and LUMO energy levels, four new Y6 derivatives, Y-PCP, Y-P6F, Y-PCF, and Y-P4FC, are designed to flexibly modify the HOMO and LUMO energy levels. The prediction trends of ML align closely with the computational trends from DFT. It is worth noting that the accuracy of LUMO energy level prediction by the prediction model makes up for the instability of DFT calculation on LUMO energy level. This work offers a cost-effective method to accelerate the acquisition of electronic properties of organic materials.

The preparation parameters of PM6:Y6 non-fullerene organic solar cells (OSCs) have significant influence on the power conversion efficiency (PCE). Herein, machine learning (ML) models are applied for analyzing the quantitative effects of the PCE on PM6:Y6 non-fullerene OSCs from the perspective of fabrication parameters. Random Forest (RF) model has the best evaluation performance and is considered as the best model among the six different algorithms. The Pearson correlation coefficient, coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error of the test set in the RF model are 0.836, 0.668, 0.695, 0.538, and 0.035, respectively. In addition, the most important preparation parameters for PCE are the effective area of OSCs, type of solvent and solvent additives, spin-coating speed, total concentration, annealing temperature, annealing time from SHAP analysis. The optimized PCE can be obtained in PM6:Y6 non-fullerene OSCs when the effective area is between 4 and 6 mm2, the spin-coating speed is 3000 rpm, and the total concentration is 18 mg ml−1, respectively. The contour map further reflects that the optimal annealing temperature are 80 or 105 °C, and the corresponding annealing time are 5 or 9 min, respectively. Based on the above preparation conditions, we verified through the experimental design that the optimal PCE is 16.91%, and the relative error with the ML prediction is only 1.24%. Our research provides low time and cost guidance on preparation parameter matching for the development of high-performance the PM6:Y6 non-fullerene OSCs.