Deep Learning Approach to Inline Quality Rating and Mapping of Multi-Crystalline Si-Wafers
This work shows the first successful application of convolutional neural networks (CNN) for material characterization and process control in solar cell production. We present a fully data-driven machine learning approach for inline quality rating and quality mapping of as-cut multi-crystalline Silicon (mc-Si) wafers. We use Photoluminescence (PL) images to image crystallization related defects in the wafers. We show that we can learn how to quantify these defect patterns based on empirical data and derive a meaningful wafer representation directly from the high-resolution input images by means of deep CNNs. This end-to-end regression model predicts solar cell efficiencies with mean errors of 0.12% for materials of bricks not presented in the training set, which is 25%rel better than our classical methods based on feature engineering. Moreover, we visualize the expected quality distribution for each sample within a spatially resolved activation map. The mapping procedure gives an insight into the ""black box"" neural network and shows that the quality distribution is in accordance to the expectations of domain experts and similar to spatially resolved quality data like the image of the dark saturation current density (j0). More details on learning and mapping will be reported within two studies elsewhere.