Deep ensemble inverse model for image-based estimation of solar cell parameters

M. Battaglia, E. Comi, T. Stadelmann, R. Hiestand, B. Ruhstaller, E. Knapp;

APL Mach. Learn. 1 September 2023; 1 (3): 036108.

https://doi.org/10.1063/5.0139707

A data-driven approach using deep learning predicts parameters of a solar cell model based on electroluminescence (EL) images. Using 75,000 synthetic EL images, the study employs a deep ensemble of 17 modified VGG19 neural networks to add uncertainty estimates. This approach, tested on four solar cell samples, bridges deep learning with engineering applications needing real-time physical model parameterizations with confidence intervals. The network's predictions showed an average deviation of 0.2% (max 10%) in junction voltage values, confirming the method's validity.

How Laoss was Used

Laoss, a simulation model, was used to model the measured solar cell samples and generate simulated electroluminescence (EL) images. These images were then used to train the inverse convolutional neural network (CNN) model. The Laoss parameterization used to generate the training data is discussed in the paper. The inverse model was trained on a set of simulated EL images, and the network architecture and training of the inverse CNN model are explained in detail. The results of the inverse model method are evaluated, and the changes in the CNN structure and training hyperparameters required to implement and train such a deep ensemble CNN inverse model are discussed.