IRIS: A Bayesian Approach for Image Reconstruction in Radio Interferometry with expressive Score-Based priors

TL;DR

IRIS tackles the ill-posed problem of reconstructing sky brightness from incomplete and noisy interferometric visibilities by employing score-based priors trained on optical galaxy images within a Bayesian diffusion framework. The method combines a learned prior p(x) with a tractable likelihood p(V|x) via the Convolved Likelihood Approximation, enabling posterior sampling from p(x|V) using SDE-based samplers (VP and VE). It demonstrates robust image reconstructions of protoplanetary disks from ALMA/DSHARP data, producing calibrated posterior samples even under prior misspecification, and provides uncertainty quantification unavailable in traditional imaging algorithms like CLEAN and MPoL. While computationally intensive, IRIS offers a principled, uncertainty-aware alternative for high-fidelity interferometric imaging and lays groundwork for broader applications and improved forward models.

Abstract

Inferring sky surface brightness distributions from noisy interferometric data in a principled statistical framework has been a key challenge in radio astronomy. In this work, we introduce Imaging for Radio Interferometry with Score-based models (IRIS). We use score-based models trained on optical images of galaxies as an expressive prior in combination with a Gaussian likelihood in the uv-space to infer images of protoplanetary disks from visibility data of the DSHARP survey conducted by ALMA. We demonstrate the advantages of this framework compared with traditional radio interferometry imaging algorithms, showing that it produces plausible posterior samples despite the use of a misspecified galaxy prior. Through coverage testing on simulations, we empirically evaluate the accuracy of this approach to generate calibrated posterior samples.

Figures (15)

• Figure 1: The Forward SDE and the Reverse SDE for a score model trained on galaxy images (SKIRT dataset). The starting point of each process is colored in gray. Prior samples are generated by starting from Gaussian noise and solving the reverse SDE while approximating the score function by a SBM.
• Figure 2: Diagram of the methodology used in this work to sample from the posterior using a score-based model as a prior.
• Figure 3: Statistical tests results of posterior samples obtained via the Euler sampler and various combinations of corrector steps and SNR for the PC sampler with the proposed approach using a score model trained under VP SDE. Each curve is color-coded consistently across both figures. Left: TARP coverage test. The shaded regions show a 99.7% confidence interval over multiple TARP tests computed with bootstrapping and the plain curve correspond to the mean. The dashed line represents the ideal case where the posterior estimator is calibrated. Right: Histograms of the samples obtained by computing the $\chi^2$ of our posterior samples.
• Figure 4: Same statistical tests as in Figure \ref{['fig:tarp_chi']}, but for a score model trained under VE SDE.
• Figure 5: The proposed approach applied on the protoplanetary disks from the DSHARP survey using the VE PROBES score-based prior. From left to right: posterior sample, percentile range (pixel-wise) and residuals (real and imaginary parts) for each DSHARP protoplanetary disk imaged with IRIS using the VE PROBES prior. For each disk, we show on the posterior sample figure a scale bar for $0.1"$ on the bottom right corner.