Computer vision-based method for quantifying iron-related defects in silicon solar cells

Abstract This study demonstrates the feasibility of employing transfer learning from pre-trained computer vision (CV) models to predict the iron concentration in silicon solar cells, even when the available training datasets are extremely limited. The predictions were based on the kinetic dependencies of the short-circuit current following FeB pair dissociation, which was converted into images using wavelet transformation. The performance of various combinations of CV models and regression algorithms was systematically analyzed. Specifically, several state-of-the-art CV architectures, including EfficientNetB7, MobileNetV2, NASNetLarge, ResNet152V2, Xception, and YOLOv4, were utilized either as classifiers or as feature extractors. Regression models, namely random forest, gradient boosting (GB), eXtreme GB, support vector regression (SVR), and deep neural networks (DNNs), were trained to predict iron concentration from the extracted features. Training and testing were performed using both simulated and experimental datasets. In both cases, EfficientNetB7 and NASNetLarge provided the most informative features for subsequent regression. Among the regression algorithms, SVR and DNNs were identified as the most effective. These models achieved mean squared error, mean absolute percentage error, median absolute percentage error, and R 2 values of up to 0.001, 6%, 4%, and 0.999, respectively, for the simulated data, and 0.008, 10%, 5%, and 0.996 for the experimental data.