Detecting Microcracks in Photovoltaics Silicon Wafers using Varitional Autoencoder

Microcracks in silicon wafers significantly reduce the wafer fracture strength, reducing yield, and increasing $/W cost. In this study, we demonstrate a high-throughput prototype that relies on an edge-on illumination source of a near-infrared laser, and a high-framerate linescan camera to detect submillimeter cracks near the wafer edges. To fully achieve the automatic micro-crack detection, we use a variational autoencoder using the principle of unsupervised anomaly detection. We evaluate the different error metrics for crack detection. At the optimal selection of the error metric and threshold, our preliminary results demonstrate the detection precision of 0.83 and the detection recall of 0.72.