Physics-Based Machine Learning Electroluminescence Models for Fast yet Accurate Solar Cell Characterization

Electroluminescence analyses of solar cells and modules allow for fast, cost-effective, and nondestructive spatial characterization of devices at different stages of their development and use. Voltage-dependent electroluminescence (ELV) measurements have been shown to mimic diode voltage–current characteristics. A derived physical model enables the determination of two local pseudoparameters from ELV data measured on silicon solar cells: a pseudorecombination current (Formula presented.) and a pseudoseries resistance (Formula presented.). Local characteristics of the solar cells, such as the series resistance (Formula presented.) or the dark saturation current (Formula presented.), can be deduced from these pseudoparameters. ELV measurements are stored in large data cubes, typically containing a few hundred thousand pixels. Pixel-wise regression is commonly achieved through nonlinear least squares (NLLS) minimization; knowing that a luminescence image of a 6^′′ silicon solar cell contains about 1 Mpix, this method is time-consuming, necessitating a trade-off between sample size, spatial resolution, fitting accuracy, and computation duration. We hence propose to replace NLLS fitting with machine learning (ML) techniques, known for their efficiency in rapidly processing large datasets. We compare the regression performances of a multilayer perceptron (MLP) with the ones of a convolutional neural network (CNN) called modified U-NET (mU-NET). The first ML model conducts a pixel-wise analysis of the data cube and the second processes the entire data cube in a single step. We present a comprehensive characterization of prediction accuracy, objectively assessing the advantages and limitations of the proposed techniques. Our first step is to ensure that the prediction precision is sufficient for a valid comparison of the analysis duration. The deviation of accuracy of these models compared to NLLS is almost negligible for MLP and of 3.1% when employing mU-NET, demonstrating their relevancy for operational application. Both ML models are fast and efficient: the time required for regression decreases by a factor of 240 with the MLP and by a factor of 1200 with the mU-NET, compared to the NLLS method.