Q-Wafer – Development of an Inline-Applicable Quality Assessment for Multi-Crystalline Silicon Wafers for Solar Cell Production

Duration: August 2011 - December 2015
Contracting Authority/ Sponsors: German Federal Ministry of Education and Research (BMBF)
Project Partners: PV Crystalox
Project Focus:
© Fraunhofer ISE
Fig. 1: Appearance of structural crystal defects in Si wafers.
© Fraunhofer ISE
Fig. 2: Predicted results of the voltage of PERC and AI-BSF solar cells based on the assessment of the as-cut PL images for the material of two unknown manufacturers, whose material was not included in the training data record.

The quality assessment of multi-crystalline (mc-Si) and high-performance multi-crystalline (HPM) wafers in the incoming tests of industrial solar cell productions requires a reproducible description of the relevant material defects and classification schemes that can also assess materials of unknown manufacturers. For this purpose, an image processing software was developed at Fraunhofer ISE, which allows the detection of different crystallization-related defects in photoluminescence (PL) images and their quantitative classification. Here, the relevancy of the individual image properties is determined via their importance for the prediction of the voltage (Voc) of the finished solar cell (AI-BSF and PERC). The created classification scheme was successfully applied to 7500 wafers that cover almost the complete spectrum of commercially available materials.


The purpose of his project was to establish a rating and prediction model for solar cell quality based on the assessment of PL images of as-cut wafers. The voltage (Voc) of the solar cell was used as target variable. A large wafer set of more than 7500 wafers was used that was purposefully compiled from 72 bricks of nine different manufacturers. Thereof, 1050 wafers were processed to Al-BSF solar cells and 6450 wafers to PERC solar cells. We captured PL images for the assessment, extracted and quantified the defects using pattern recognition techniques, and finally trained different assessment models using machine learning methods. Fig. 1 shows PL images of HPM wafers and mc-Si wafers with detected defects.

The measurement results were divided into a training and test data set for the development of the assessment models. This division can be random, while the most difficult case occurs if wafers of a manufacturer need to be rated, whose material was not included in the training data set. The simplest assessment model is the “2 feature model”, which only contains the weighting of dislocation lines and edge areas. The more complex models developed here contain more image properties and a more complex connection of information. However, this may lead to over-adjustment to the training data. A balanced model was achieved with the “elastic-net” regression method.

The wafer assessment results for an unknown manufacturer are shown in Fig. 2. The superiority of the “elastic-net” model developed here over the “2-feature” model is indicated by the significantly better correlation between the predicted and measured Voc values for PERC solar cells with a mean prediction error of 2.5 mV compared to 5.1 mV. The poor results of the “2-feature” model show that a reliable quality assessment for wafers for the PERC solar cell production urgently requires assessment models beyond the current industry standard.