Monitoring the electrical quality of printed structures based on optical measurements using deep learning

As the structural size of contact fingers in solar cells decreases to reduce silver consumption and shading losses, their quality inspection becomes increasingly relevant and challenging. Existing techniques to determine the electrical quality of contact fingers are either destructive and slow or restricted to electrical properties between busbars, thus not resolving individual finger properties. This work presents a machine learning approach that allows for monitoring electrical quality based on optical images of the corresponding finger geometry measured during production. The optical measurements are capable of deriving structural information from the shadow of the printed structure cast by laser illumination. A comprehensive dataset of 7250 samples with varying print structures was created to train and validate the model. As ground truth, the electrical resistance of each contact finger is measured. The multivariate network design allows for both the regression of resistance values and the classification of defective structures. It successfully predicts the resistance of contact fingers, achieving a correlation coefficient of 0.98. Defective structures are detected with a precision of 0.92 and a recall of 0.85. Finally, integrating the model into a computer vision pipeline for full cell analysis allows for the visualization of the distribution of electrical quality parameters for the identification of process- or screen-related defect patterns and systematic production errors.