The Empirical Digital Twin: Representation Learning on Solar Cell Images and Efficient Defect Detection with Human-in-the-Loop
Measurement images of solar cells provide information beyond current-voltage characteristics regarding process and material quality in a spatially resolved manner. However, this information is only partially used because algorithms search for human-defined defects and structures. These labels can be inaccurate and incomplete, a relevance in terms of electrical quality is not necessarily given. Thus, we propose a purely data-based approach to derive a comprehensive representation from the measured images that is meaningful in terms of electrical quality and show how it can be used for efficient defect detection. We call this representation the empirical digital twin. For its calculation, we design a convolutional neural network combining multiple measurement images by correlating them with quality variables. The digital twin is an intermediate representation of the network and summarizes quality-related defect signatures that are visible in the images. We show how this representation can be used to derive sorting criteria for quality inspection within an efficient human-in-the-loop approach detecting defects such as finger interruptions, shunts, etc. The human-in-the-loop method not only needs fewer training samples and thus fewer labels but also improved the 𝐹1-Score detection rate by about 2% on average.