Navigating the empirical digital twin space of solar cells: Semantic compression incorporating j0 and Rs images for systematic defect analysis

In addition to current-voltage assessment, solar cells undergo an advanced characterization at the end-of-line based on inline measurements like electroluminescence (EL) and photoluminescence (PL) imaging or reflectance measurements. However, their full potential remains underutilized, as these measurements are typically analyzed separately, reduced to predefined defects, yet stored at high data costs. We introduce a machine learning model deriving empirical digital twins of solar cells, semantically compressing and summarizing these measurements. These digital twins categorize cells into clusters based on properties like defects, processing variations, or quality levels. To interpret these clusters, we propose a systematic approach assigning specific loss characteristics to each cluster. Applied to over 2800 industrial cells, we identify loss groups regarding resistance, reflectance and absorption at various wavelengths, and recombination. Furthermore, we analyze manufacturer processing data, finding that cluster variations could primarily originate from fluctuations in diffusion and passivation steps. The multi-modal deep learning model comprises EL, PL, and thermography images, along with reflectance curves, predicting (i) global parameters like efficiency and (ii) spatially resolved images of dark saturation current density ( j 0 ) and series resistance. Hence, the digital twin, extracted during the final compression step, links directly to quality information. Additionally, this approach allows inference of j 0 and R s images bypassing the actual measurements.