Automated quantitative diagnosis of GaAs solar cells using the CBAM-MS-1DCNN model on absolute electroluminescence imaging and distributed circuit modeling.

Three-dimensional distributed circuit modeling based on SPICE software can simulate the electrical performance and the uniformity of electroluminescence (EL) intensity. To solve the problem that traditional methods require manual iteration to determine the device parameters of the simulation model, we proposed an efficient automated quantitative analysis method that can quickly diagnose the localized series resistor () of the dark-spot defects in GaAs solar cells via absolute EL images. This method employs a one-dimensional convolutional neural network model based on a multi-scale and convolutional block attention module (CBAM-MS-1DCNN). A medium dataset consisting of 200,000 defects from 250 simulated solar cells to train the CBAM-MS-1DCNN model, where the coefficient of determination ()>0.95 and the normalized root mean square error ()<5, indicating that the proposed model can predict well. Furthermore, the relative error of the predicted absolute EL intensity based on real samples can be controlled within 10% using the CBAM-MS-1DCNN model. The model trained on the simulated dataset has good prediction performance for real GaAs solar cells, which provides ideas for the problem of obtaining EL image datasets.