Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Visualizing Material Quality and Similarity of mc-Si Wafers Learned by Convolutional Regression Networks

: Demant, M.; Virtue, P.; Kovvali, A.; Yu, S.X.; Rein, S.

Postprint urn:nbn:de:0011-n-5584754 (3.2 MByte PDF)
MD5 Fingerprint: efa59984bb1e89575f8b6270a7375c2d
© IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.
Erstellt am: 21.11.2019

IEEE Journal of Photovoltaics 9 (2019), Nr.4, S.1073-1080
ISSN: 2156-3381
ISSN: 2156-3403
Zeitschriftenaufsatz, Elektronische Publikation
Fraunhofer ISE ()
Inline-Wafer-/Prozessanalytik und Produktionskontrolle; Photovoltaik; Silicium-Photovoltaik; Charakterisierung von Prozess- und Silicium-Materialien; Messtechnik und Produktionskontrolle; CNN; regression; DenseNet; mapping; mc-si

Convolutional neural networks can be trained to assess the material quality of multicrystalline silicon wafers. A successful rating model has been presented in a related work, which directly evaluates the photoluminescence (PL) image of the wafer to predict the current-voltage parameters after solar cell production. This paper presents the results of two visualization techniques to understand what has been learned in the network. First, we reveal what has been learned in the PL image by visualizing the spatial quality distribution of the wafers based on the activation maps of the network. The method is denoted as regression activation mapping. We compare regression activation maps with j0 images of solar cells to show the semantically meaningful representation of the trained features. Second, we show what has been learned in the data by mapping the learned network representation of all wafers into a low-dimensional subspace. Visualizations reveal the smoothness of our representation with respect to the PL input and measured quality. This technique can be used to detect material anomalies or process faults for samples with high prediction errors.