Towards end-to-end Bayesian forward models in global 21-cm cosmology: surrogate modelling and marginalisation of beam uncertainty
Robust statistical inference in global 21-cm cosmology requires end-to-end uncertainty quantification that jointly handles the highly degenerate cosmological signal, foreground emission, and instrumental response. Although electromagnetic simulations capture physical antenna properties in a parametrised way, multi-hour runtimes make their integration within likelihood-based sampling frameworks infeasible. Most existing approaches therefore assume a single precomputed beam, a fragile assumption given our demonstration that realistic mismatches can severely bias the recovered cosmological and foreground parameters. To address this, we present an accelerated and differentiable Bayesian framework that incorporates an informed surrogate representation of chromatic beam uncertainty directly into a forward-modelling pipeline. Treating the physical antenna properties as nuisance quantities, we apply a two-stage decomposition directly to simulated directivity patterns, reducing the instrumental parameterisation by two orders of magnitude while retaining the angular and spectral structure required for accurate beam reconstruction. Exploiting the linearity of the resulting surrogate, we use analytical marginalisation to allow the continuous instrumental uncertainty to be propagated into the final posteriors and Bayesian evidence without directly sampling the beam space. Testing the framework against a suite of unseen beams and cosmological signals, we recover the true inputs at approximately the instrumental-noise level. We further show that, for the uncertainty considered here, as few as 100 electromagnetic simulations are sufficient to construct an effective surrogate, substantially reducing the simulation burden for future analyses. This framework provides a scalable, statistically rigorous route towards hardware-accelerated uncertainty quantification in global 21-cm cosmology.