Machine Learning for Materials Evaluation in PV Production

Silicon-Photovoltaics

New technology in machine learning first surpassed human performance in classifying objects in 2012. As the capacity of production lines increases, purely data-based procedures for production control are gaining in importance, which can be implemented very efficiently with this technology. As part of its digitalization initiative, Fraunhofer ISE is addressing the transfer of deep learning procedures to various links of the PV value chain. To evaluate multicrystalline silicon wafers, a sufficiently broad range of data was gathered in the »Q-Wafer«A and »Q-Crystal«B projects to allow the expected solar cell quality to be predicted already from inline measurements before solar cell production.

© Fraunhofer ISE
Direct processing of the photoluminescence (PL) image of a multicrystalline Si wafer to predict the current-voltage (IV) parameters of the solar cell by application of a convolutional neural network.

In addition to predicting Machine Learning for Materials Evaluation in PV Production the efficiency accurately with a mean error of 0.12%abs, the process offers further advantages: In our application example, a convolutional neural network learns the direct connection between the 2D photoluminescence images of the wafer and the current-voltage parameters of the solar cell. No human specifications were needed to evaluate the input images. The model can be easily extended with additional processing and input parameters and prediction requires only a few milliseconds. Correct prediction for wafers from unknown bricks demonstrates the high degree of generalizability, so that deviations between predicted and measured efficiency can be used as indicators for processing errors and material anomalies, which can be specified by reference methods such as »modulum« that are available at Fraunhofer ISE.

© Fraunhofer ISE
Prediction of the open circuit voltage on the basis of photoluminescence images of the wafer, obtained by applying a convolutional neural network.

A Supported by the German Federal Ministry for Education and Research (BMBF)

B Supported by the German Federal Ministry for Economic Affairs and Energy (BMWi)