Astronomers use spectral energy distributions (SEDs) — measurements of a galaxy’s light at many wavelengths — to infer its stars, chemical makeup, dust, and history. New data from the James Webb Space Telescope (JWST) and upcoming surveys such as Euclid, Rubin’s LSST, and Roman are producing far richer SEDs and many more galaxies. The paper presents CERIDWEN, a software package designed to make full Bayesian SED inference much faster and flexible enough to match these richer datasets.

CERIDWEN is written in JAX and runs natively on graphics processing units (GPUs). The authors build an “end-to-end differentiable forward model,” which means the code simulates how a galaxy’s physical properties produce the observed light, and those calculations are differentiable so the software can compute exact gradients. The forward model includes stellar populations, nebular emission (light from ionised gas), dust attenuation and emission, and conversion into the observer’s frame. Its computations are vectorised and compiled so many samples can be evaluated in parallel on the GPU.

That architecture lets CERIDWEN run Bayesian samplers efficiently. The code supports nested sampling — a method that explores the full probability distribution of parameters — implemented so a batch of live points can be replaced in parallel on the GPU. Automatic differentiation supplies exact gradients for gradient-based samplers, speeding convergence. The package also allows much more flexible physical models than many older tools. For example, the authors infer time-dependent chemical-enrichment histories instead of assuming a single metallicity, and they show non-parametric star-formation histories with about 120 age bins.

CERIDWEN can also use alpha-enhanced stellar libraries from FSPS (Flexible Stellar Population Synthesis). That lets it sample stellar [α/Fe] (the abundance of alpha elements relative to iron) jointly with [Fe/H] (iron abundance), stellar mass, and the star-formation history. Sampling these together means the resulting probability distribution explicitly captures the degeneracy between element ratios and iron abundance. In tests on controlled mock data, CERIDWEN recovered parameters with well-calibrated posterior uncertainties. Fits to real JWST observations reproduced posteriors from the established Prospector framework. On a single GPU, the median sampling time per fit was about 4 minutes, roughly 134 times faster per fit than equivalent CPU-based Prospector runs reported by the authors.