VROOM-SBI: A Fast Simulation-Based Bayesian Inference Methodology for QU-Fitting

Bayesian QU-fitting is among the most accurate approaches for line-of-sight Faraday inference, but its per-pixel computational cost has made survey-scale application infeasible. QU-fitting is an alternative to Faraday synthesis with comparable accuracy in recovering line-of-sight Faraday components, but it has historically been computationally prohibitive at survey scale. Fitting to the Stokes spectra in $Q$ and $U$ through Bayesian inference is effective but slow. We introduce $\mathtt{\text{VROOM-SBI}}$, which uses simulation-based inference, particularly neural posterior estimation, to speed up inference. Our results are comparable to both Faraday synthesis and QU-fitting, and deliver a speedup of $\sim$$500$ over classical QU-fitting implementations. We provide an open code repository and tools along with trained models via HuggingFace for the four standard depolarization models in common use, trained on VLA L-band frequency coverage.