CC BY 4.0Kunze, PhilippPhilippKunzeKwapil, WolframWolframKwapilRein, StefanStefanReinDemant, MatthiasMatthiasDemant2025-08-192025-08-192025https://publica.fraunhofer.de/handle/publica/490749https://doi.org/10.24406/publica-509110.1002/solr.20250025310.24406/publica-50912-s2.0-105012481748This work focuses on estimating spatially resolved (Formula presented.) parameters from inline measurements of solar cells, aiming to provide detailed and localized cell characterization suitable for inline application. Existing methods achieving this, such as total cell analysis, often rely on laborious offline measurements and are therefore sophisticated yet limited by their time-consuming nature and specialization requirements. To address these limitations, we propose a semisupervised machine learning model estimating (Formula presented.) parameter maps by first predicting global (Formula presented.) parameters and subsequently inferring their local distribution across the solar cell. Using industry-standard measurement techniques like electroluminescence, thermography, and reflectance, the model predicts global parameters, e.g., efficiency or open-circuit voltage. This forms the basis for approximating the local distribution of (Formula presented.) parameters, referred to as quality activation maps. Through comparison with the total cell analysis as physical reference, we investigate their accuracy, verifying that known defect structures affect the correct (Formula presented.) parameters locally. For instance, finger interruptions reduce the fill factor at their location while leaving the pseudo-fill factor unaffected. Ultimately, we present an application in which quality activation maps are employed to quantify the efficiency losses induced by individual defects, thereby assigning the primary (Formula presented.) losses to their respective sources.enfalseartificial intelligencecharacterizationdeep learningmachine learningphotovoltaicsolar celljournal article