9:15 AM - MT03.01.04/MT02.01.04
Automatic Microcrack Inspection in Photovoltaics Silicon Wafers by Unsupervised Anomaly Detection via Variational Auto-Encoder
Zhe Liu1,Felipe Oviedo1,Emanuel Sachs1,Tonio Buonassisi1
Massachusetts Institute of Technology1
The presence of microcracks in silicon wafers significantly reduces wafer strength, yielding wafer breakage during the manufacturing process, transportation and field operation. With the trend of decreasing wafer thickness for cost reduction purposes, thinner wafers are more prone to breakage in the presence of microcracks . To enable a smooth transition to thin wafers for even cheaper photovoltaic modules, we recently developed a high-throughput prototype for in-line crack detection for silicon wafers . This tool scans silicon wafer in the near-edge regions for micro-cracks and outputs linescan signals from a linescan camera, where no crack shows a smooth, undisrupted profile. As an in-line detection tool, it also requires a rapid and reliable algorithm that automatically identifies the presence of a micro-crack within a second after wafer scanning. In this work, we adopted an unsupervised machine learning method for anomaly detection, because the presence of microcracks above the critical length is typically a statistically rare event in the current PV production line (typically less than 5%). Specifically, a generative machine learning algorithm variational auto-encoder (VAE) is used to identify scans with microcracks . The working principle of this algorithm is that: (1) VAE encodes the linescan profiles into lower-dimension vectors of latent variables, and then the latent variables are reconstructed back to linescan profile with the goal of minimized error; (2) because of most linescan profiles are very similar smooth curves without any cracks, the VAE model is trained to be biased toward linescan without cracks; (3) whenever a linescan profile for a crack appears, the trained VAE model generates a vastly different profile with a significant reconstruction error; (4) the crack is then detected by monitoring anomalous reconstruction error. The advantage of this unsupervised VAE method over the previous neural network method  is that it does not require a large amount of labelled crack data with different crack shapes (which can be very difficult to obtain). We demonstrate successful crack detections with several different wafer types (e.g., multi, mono, as-cut, and textured) and crack shapes (e.g., line-shape, cross-star, L-shape). We show that, with statistical analysis, this VAE-based anomaly detection could be a reliable and versatile method to enable the rapid detection of microcracks in silicon wafers.
 S. Wieghold, Z. Liu, S. J. Raymond, L. T. Meyer, J. R. Williams, T. Buonassisi, and E. M. Sachs, “Detection of sub-500-μm cracks in multicrystalline silicon wafer using edge-illuminated dark-field imaging to enable thin solar cell manufacturing,” Solar Energy Materials and Solar Cells, vol. 196, pp. 70–77, Jul. 2019.
 Z. Liu, S. Wieghold, L. T. Meyer, L. K. Cavill, T. Buonassisi, and E. M. Sachs, “Design of a Submillimeter Crack-Detection Tool for Si Photovoltaic Wafers Using Vicinal Illumination and Dark-Field Scattering,” IEEE Journal of Photovoltaics, vol. 8, no. 6, pp. 1449–1456, Nov. 2018.
 H. Xu, W. Chen, N. Zhao, Z. Li, J. Bu, Z. Li, Y. Liu, Y. Zhao, D. Pei, Y. Feng, J. Chen, Z. Wang, and H. Qiao, “Unsupervised Anomaly Detection via Variational Auto-Encoder for Seasonal KPIs in Web Applications,” Feb. 2018.
 M. Demant, T. Welschehold, M. Oswald, S. Bartsch, T. Brox, S. Schoenfelder, and S. Rein, “Microcracks in Silicon Wafers I: Inline Detection and Implications of Crack Morphology on Wafer Strength,” IEEE Journal of Photovoltaics, vol. 6, no. 1, pp. 126–135, Jan. 2016.