Presentation Details
| Physics-Informed Machine Learning for Solar Cell Diagnosis: Extracting a Physical Digital Twin from Inline Metrology Andreas Fell, Alexandra Wörnhör, Wilkin Wöhler, Ralf Preu, Stefan Rein, Matthias Demant. Fraunhofer Institute for Solar Energy Systems, Freiburg, Germany |
Abstract
Conventional inline measurements in solar cell production capture vast amounts of data but fall short to directly reveal critical physical parameters that cause performance losses. To bridge the gap between accessible measurement data and expert-level physical understanding, we introduce a physics-informed machine learning approach that extracts a physical digital twin (PDT)—a complete set of physical solar cell properties—from heterogeneous inline measurements within seconds. The PDT supports root-cause analysis of performance limitations, and identification of optimization potential of the current cell design. Superior to classical physics-informed ML models, which dominantly use supervised learning e.g. on pre-simulated datasets, our model fully integrates a complete physical simulation model directly into the deep-learning neural network architecture. It processes multimodal inline data (EL/PL images, JV-curves, reflectance spectra, and scalar metrics) through specialized feature extractors (2D CNNs for images, 1D CNNs for spectra, transformers for cross-modal fusion). The predicted PDT parameters feed into a hybrid simulation comprising a parallelized C++ optical model and a Gaussian process surrogate for the Quokka3-based electrical model. Crucially, optimization relies on reconstruction consistency between simulated and measured observables, eliminating the need for ground-truth PDT labels—which are prohibitively expensive to obtain—while enforcing physical plausibility through the simulation constraints. We validated the approach on 2,000 bifacial PERC cells featuring 15 systematic intentional process variations. The results demonstrate high model generalizability by correctly predicting measurable observables on a test dataset excluded from the training run. The extracted PDTs furthermore correctly capture expected trends of latent physical properties, like front contact resistivity and recombination parameters as a function of firing temperature and laser enhanced contact optimization (LECO). Our PDT model substantially reduces data storage requirements while unlocking actionable physical insights from every produced cell, a basis for continuous, automated quality assurance and process optimization in solar cell manufacturing.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.