Inline quality rating of multi‐crystalline wafers based on photoluminescence images

Abstract

The quality assessment of multi-crystalline and high-performance multi-crystalline silicon wafers during incoming inspection of solar cell production requires a reproducible description of the relevant material defects and a classification scheme that is capable to rate as-cut wafers from unknown manufacturers. Both needs are addressed in this work. We introduce an image processing framework that allows the various types of crystallization-related defects visible in photoluminescence images to be detected quantitatively and thus enables a complete wafer description in terms of defects. The importance of different features within this defect characteristic is weighted by predicting the open-circuit voltage of solar cells with aluminum back-surface field as well as passivated emitter and rear cells with a stepwise extension of the model. The resulting robust classification scheme is successfully evaluated on a set of 7500 wafers, which represents the whole spectrum of material types and qualities that are currently available at the market. A comparison of defect signatures in high-performance multi-crystalline and standard multi-crystalline silicon materials underlines the relevance of additional features. As a result of this paper, we show that a regularized version of multi-linear regression models for quality prediction can optimize simpler linear models by adding selected features to the defect characteristic leading to mean absolute prediction errors of 2.2 mV for solar cells with aluminum back-surface field and 2.9 mV for passivated emitter and rear cells solar cells in a true blind test

Explaining the Efficiencies of Mass‐Produced p‐Type Cz‐Si Solar Cells by Interpretable Machine Learning

“…The high prediction accuracy for the hourly moving averages (correlation coefficient of 0.99 and mean absolute error of only 0.025) is especially remarkable as 1) all time‐related quantities are removed beforehand, 2) as we deal with a large timeframe, usually harder to learn than on data from a daily basis, and 3) as there is only little material‐related influence (compare, e.g., previous studies [ 3,11 ] for the huge material‐related impacts in multicrystalline silicon solar cells). [ 3,11 ] The background of the last point is that to achieve high goodness of fits for multicrystalline silicon solar cells it is already sufficient to include inline measurements such as base resistivity and dislocation density from photoluminescence imaging to already explain big fractions of the absolutely much higher cell efficiency variance. As can be seen from the grayscale colorbar of the dots illustrating the model predictions in the same range as the raw actual or true data, even the most of the really poorly performing cells are explained by the model (e.g., the ones in the cluster in the beginning and in the efficiency drop after 105 h).…”

“…with a Pearson coefficient of correlation of 0.84 and a mean absolute error of 0.33 between the standardized true and predicted values. The high prediction accuracy for the hourly moving averages (correlation coefficient of 0.99 and mean absolute error of only 0.025) is especially remarkable as 1) all time‐related quantities are removed beforehand, 2) as we deal with a large timeframe, usually harder to learn than on data from a daily basis, and 3) as there is only little material‐related influence (compare, e.g., previous studies [ 3,11 ] for the huge material‐related impacts in multicrystalline silicon solar cells). [ 3,11 ] The background of the last point is that to achieve high goodness of fits for multicrystalline silicon solar cells it is already sufficient to include inline measurements such as base resistivity and dislocation density from photoluminescence imaging to already explain big fractions of the absolutely much higher cell efficiency variance.…”

Anomaly Detection and Automatic Labeling for Solar Cell Quality Inspection Based on Generative Adversarial Network

“…Today, different imaging techniques are used during PV module inspection to obtain images where defects appear highlighted, for example, Electroluminescence (EL) [ 4 , 5 ], Photoluminescence (PL) [ 6 , 7 ], or Thermography [ 8 , 9 ]. During the assembly stage, EL is one of the predominant techniques.…”

Anomaly detection and automatic labeling for solar cell quality inspection based on Generative Adversarial Network

“…Nowadays, different imaging techniques are used during PV module inspection to obtain images where defects appear highlighted, for example, Electroluminescence (EL) (Bartler et al (2018), Chen et al (2020)), Photoluminescence (PL) (Demant et al (2016), Nos et al (2016)), or Thermography (Pierdicca et al (2018), Vaněk et al (2016)). In industrial scenarios, the EL is one of the predominant techniques.…”