Rapid advances in organic photovoltaics (OPVs) have been driven by the tunability of molecular semiconductors and continuous improvements in device design. Yet, understanding key losses and creating a predictive model of emerging photovoltaic devices requires a better understand of the relation between variations in processes and chemistry of a sample and the material parameters that govern charge generation, transport, and recombination. The ideal scenario would be the ability to derive key material parameters directly from readily available characterization data such as current-voltage curves. However, this remains a major challenge due to the strong interdependence of parameters and the computational burden of traditional drift-diffusion–based fitting. Here, we present a machine-learning-assisted Bayesian framework that enables rapid, high-dimensional inference of material parameters directly from current–voltage (J–V) characteristics of organic solar cells [1]. Instead of fitting experimental data with repeated numerical simulations, we construct neural-network surrogate models trained on large datasets of drift-diffusion simulations covering the relevant parameter space. This approach accelerates evaluation of J–V curves by approximately three orders of magnitude, allowing comprehensive exploration of the posterior probability landscape associated with key physical parameters. We apply this framework to organic solar cells, demonstrating that the combination of surrogate modelling and Bayesian inference captures both the most probable material parameters and their uncertainties. By incorporating thickness- and illumination-dependent J–V curves, we quantify the information content carried by different experimental datasets and identify which measurements are most sensitive to transport, recombination, or interfacial parameters. Our analysis reveals, for example, that injection barriers remain poorly constrained by J–V curves alone, while charge-carrier mobility, direct recombination coefficients can be determined with higher confidence. Two-dimensional visualizations of the posterior parameter correlations further highlight strong couplings between mobility and direct recombination, illustrating the importance of probabilistic inference beyond single best-fit values. Overall, this work establishes a generalizable and computationally efficient route for extracting material parameters from routine OPV characterization. The combination of physics-based simulation, neural-network surrogate modelling, and Bayesian inference provides a powerful toolset for diagnosing loss mechanisms, guiding material optimization, and accelerating device development across emerging photovoltaic technologies.