Spatially Resolved Material Quality Prediction Via Constrained Deep Learning

Novel material classes for solar cell production e.g. high performance multicrystalline silicon or epitaxially grown wafers have a huge impact on solar cell performance. A speedup of these developments calls for a rapid assessment of the material quality in the as-cut stage already. This work introduces a generic architecture for the material rating of wafers in terms of solar cell quality. Our approach allows for a simultaneous prediction of the open-circuit voltage of the solar cell and the image of the dark-saturation current density (j(0)) from photoluminescence images of as-cut wafers. In the sense of theory-guided data-analysis, we combine a data-driven machine learning approach with known physical constraints, here given by the one-diode equation. Due to this statistical optimization, our method derives the j(0) values for the occluded regions beneath the busbar. From the derived j(0) values, we also evaluate the impact of material-related defects and grid metallization structure on the dark-saturation current density.