Recombination-active defects e.g. dislocations in multicrystalline silicon (mc-Si) wafers impact the quality of solar cells. These defects can be quantified during the incoming control of silicon wafers with Photoluminescence (PL) imaging and used to rate the solar cell quality. In this work, we analyze the relevance of defect patterns in PL images and grain-boundary (GB) data for current-voltage (IV) prediction by means of image processing algorithms. Based on a large set of empirical data of passivated emitter and rear cells (PERC), a sparse prediction model is trained for each IV-parameter. Our results include both, the prediction of different quality parameters and the relevance of the features extracted from PL images and GB images. We achieve mean absolute prediction errors as low as 2.72 mV and 0.18 mA/cm2 for open circuit voltage (Voc) and short circuit current density (Jsc) respectively, and 0.18% for efficiency as combined parameter. In this evaluation, the wafer data set is split into training group and test group. Therefore the results show the prediction of unknown material. This makes the prediction more challenging but represents a realistic use case for production. The comparative overview of the relevant feature set shows a difference between the prediction of short-circuit current and open-circuit voltage prediction.Recombination-active defects e.g. dislocations in multicrystalline silicon (mc-Si) wafers impact the quality of solar cells. These defects can be quantified during the incoming control of silicon wafers with Photoluminescence (PL) imaging and used to rate the solar cell quality. In this work, we analyze the relevance of defect patterns in PL images and grain-boundary (GB) data for current-voltage (IV) prediction by means of image processing algorithms. Based on a large set of empirical data of passivated emitter and rear cells (PERC), a sparse prediction model is trained for each IV-parameter. Our results include both, the prediction of different quality parameters and the relevance of the features extracted from PL images and GB images. We achieve mean absolute prediction errors as low as 2.72 mV and 0.18 mA/cm2 for open circuit voltage (Voc) and short circuit current density (Jsc) respectively, and 0.18% for efficiency as combined parameter. In this evaluation, the wafer data set is split into training ...
[1]
S. Rein,et al.
Appearance of rift structures created by acidic texturization and their impact on solar cell efficiency
,
2010,
2010 35th IEEE Photovoltaic Specialists Conference.
[2]
B. Lai,et al.
Recombination activity of grain boundaries in high-performance multicrystalline Si during solar cell processing
,
2018
.
[3]
M. Schubert,et al.
Photoluminescence imaging of silicon wafers
,
2006
.
[4]
Michael Felsberg,et al.
The monogenic signal
,
2001,
IEEE Trans. Signal Process..
[5]
A. Lorenz,et al.
Fast Photoluminescence Imaging of Silicon Wafers
,
2006,
2006 IEEE 4th World Conference on Photovoltaic Energy Conference.
[6]
Thomas Brox,et al.
Inline quality rating of multi‐crystalline wafers based on photoluminescence images
,
2016
.
[7]
Eicke R. Weber,et al.
Quality control of as-cut multicrystalline silicon wafers using photoluminescence imaging for solar cell production
,
2010
.