Speaker
Description
Understanding the Epoch of Reionisation remains one of the pivotal tasks for modern cosmology, and next-generation telescopes such as EUCLID and JWST are opening up the path to the first precision constraints on reionisation derived from the Lyman-alpha damping wing signature imprinted by the foreground neutral intergalactic medium (IGM) on the spectra of high-redshift quasars.
We developed a new simulation-based inference framework – coded fully differentiably in the machine learning framework JAX – to disentangle the IGM damping wing from a quasar's unknown intrinsic spectrum and infer its lifetime as well as two physical measures of the local ionisation topology in front of the quasar: the HI column density and its distance to the first neutral patch. Our pipeline accounts for all relevant modelling uncertainties, caused by IGM transmission fluctuations, quasar continuum reconstruction, and spectral noise. Enabled by a normalising flow model as neural likelihood estimator, our framework is the first that harnesses the full-resolution spectral information, including the highly non-Gaussian pixels blueward of the Lyman-alpha line. Utilising the fully generative nature of normalising flows and analysing higher-order statistics of synthetic flow-generated spectra, we demonstrate that our model has truly learned non-Gaussian information, significantly tightening the resulting parameter constraints.
Based on realistic mock spectra resembling the distribution of our ongoing 94-hour Cycle 4 JWST program and upcoming quasars found by Euclid, we show that we will soon be able to robustly constrain the evolution of the IGM neutral fraction at the <5% level between 6 ≲ z ≲ 10. We present the first such constraints for 41 ground- and/or JWST-based quasar spectra at 5.75 < z < 7.6, constraining both global timing and the local ionisation topology near these objects, and simultaneously conducting an unprecedented census of the lifetimes of these quasars, holding crucial information about supermassive black hole growth.
| Talk category | Plenary |
|---|---|
| Second preference | NOVA Network 1, Splinter 6: Data science |
