To accelerate the optimization process of solar cells, a deep understanding of material systems is essential. This is particularly important for emerging materials such as organic semiconductors and halide perovskites, which offer a larger degree of variation with respect to the stoichiometry and crystallinity. The downside of this chemical versatility is the larger amount of unknown optical and electronic properties of these emerging optoelectronic materials that triggers the need for more extensive efforts towards material and device characterization as compared to e.g. crystalline silicon. However, the strong correlation between material properties makes it challenging to isolate individual parameters using experimental data.

In the past, numerical simulations such as drift-diffusion simulations have enabled researchers to reconstruct characterization data using known material parameters (e.g., mobilities, recombination coefficients, band gaps) and compare the results with experimental data to analyze solar cells fabricated in a lab.^[1-3] Nevertheless, the traditional fitting routine for inferring parameters from one or several measurements is time-consuming and relies on a deterministic approach that lacks any quantification of the confidence in the uniqueness of the resulting fit.

To address these limitations of traditional fitting of numerical models to experimental data in the context of photovoltaics, we applied a methodology from the field of machine learning. We use a neural network as a surrogate model for the device simulation. The role of the neural network is to speed up (by a factor of 10⁵) the process of device simulation within a range of predefined parameters.^{[4, 5]} To include information on the confidence in and the uniqueness of the resulting fits, we employ the framework of Bayesian Parameter Estimation (BPE)^[6-8] methods. In this work, we have adapted this workflow to the parameter inference problem in organic solar cells. By leveraging the power of machine learning, we can efficiently explore the complex parameter space and provide a more accurate and robust characterization of organic solar cells. By analyzing JV curves of a certain blend of organic solar cells at different light intensities and for different thicknesses, we were able to infer electronic properties such as mobilities, recombination coefficients, and defect densities. Moreover, this approach provides a posterior probability distribution after observing an JV curve, allowing us to quantify the information gained per additional unit of experimental information considered for the inference process.