Hyperspectral spectral photoluminescence (HPL) imaging is a powerful, contactless, and spatially-resolved technique widely employed to characterize solar cells at various stages of their development and operation. However, this method is time-consuming due to the sequential acquisition across multiple wavelengths and the subsequent analysis of large, high-dimensional data cubes. We propose an advanced framework for HPL data analysis in perovskite solar cells. We first investigate the application of deep learning (DL) techniques to significantly reduce both acquisition and processing times, paving the way for faster and more scalable solar cell analyses, and then explore the correlation between electrical and optical characteristics of the cells.

We perform a DL-based local analysis of HPL data cubes to extract maps of Urbach energy (Eu), quasi-Fermi level splitting (QFLS) and bandgap energy (Eg). A multilayer perceptron (MLP) is trained on synthetic data generated using the generalized Planck’s law, combined with a logistic absorptance model, and simulated additive noise. The performances of the MLP are compared to non-linear least squares (NLLS) fitting applied to HPL spectra, following the approach introduced by Laot et al. [1].

Our results show that, while the MLP does not significantly increase the prediction error, it substantially reduces the computation time. The average absolute relative error (ARE) across more than 13 000 spectra - between the NLLS fitting results and the MLP prediction - is of 1.76% for Eu, 0.07% for Eg and 0.12% for QFLS. The total prediction time of for all three parameters across 13 000 voxels is under 1 second using MLP, compared to approximately 10 minutes using NLLS fitting.

We investigate how increasing the wavelength step size - thereby reducing the number of sampling points in the spectrum - impacts the prediction accuracy on the three variables of interest (Eg, Eu, QFLS). This approach enables a significant reduction in acquisition time, which not only accelerates the characterization process but also minimizes the degradation of perovskite materials that may occur during prolonged measurements. Furthermore, we apply transfer learning methods to adapt DL models trained on densely sampled spectra for use with sparsely sampled spectra, thereby reducing the training time required for the new models.

Using the theoretical framework proposed by Kirchartz et al. [2] we compute the local photocurrent, recombination current and open-circuit voltage, under the approximation that the quantum efficiency is equivalent to the previously calculated absorptance of the cells. We finally correlate the optical characteristics of the cells with their electrical parameters by analyzing a dataset of more than 100 p-i-n halide perovskite solar cells.

We explore how local parameter heterogeneity influences the global electrical figures of merit.