Learning Quality Rating of As-Cut me-Si Wafers via Convolutional Regression Networks

This paper investigates deep convolutional neural networks (CNNs) for the assessment of defects in multicrystalline silicon (me-Si) and high-performance me-Si wafers for solar cell production based on photoluminescence (PL) images. We identify and train a CNN regression model to forecast the T-V parameters of passivated emitter and rear cells from given PI, images of the as-cut wafers. The presented end-to-end model directly processes the PL image and does not rely on the human-designed image feature. Domain knowledge is replaced by a model based on a huge variety of empirical data. The comprehensive dataset allows for the evaluation of the generalizability of the model with test wafers from bricks and manufacturers not presented in the training set. We achieve mean absolute prediction errors as low as 0.11%(abs) in efficiency for test wafers from unknown bricks, which improves handcrafted feature-based methods by 35%(rel) at simultaneously lower computational costs for prediction. Samples with high prediction errors are investigated in detail showing an increased iron point defect concentration.