Supervisor: Dr. Jonathan R Pritchard
I am currently affiliated with the Radio Cosmology group of Max Planck Institute for Radioastronomy (MPIfR) as a guest scientist.
Constraining Cosmic Dawn with Neural Inference
My research explores Cosmic Dawn and the Epoch of Reionization, the era when the first stars and galaxies emerged and began shaping the large-scale structure of the Universe. Using simulation-based inference with neural posterior estimation (NPE) applied to mock Square Kilometre Array (SKA) 21-cm observations, I investigate both the astrophysical processes governing the formation of the first luminous sources and the cosmological parameters that influence structure formation on large scales. This includes constraining the properties of early star-forming galaxies, the thermal and ionization history of the intergalactic medium, and cosmological signatures such as primordial non-Gaussianity, which encodes information about the physics of inflation. By combining forward simulations, Bayesian inference, and machine learning, this work aims to develop robust analysis pipelines capable of extracting both astrophysical and cosmological information from next-generation 21-cm observations, offering new insights into the evolution of the early Universe.
Scientific Software : 21cmjaxSBI
21cmjaxSBI is a high-performance, simulation-based inference framework engineered for 21-cm cosmology and neural posterior estimation (NPE). Built in Python using JAX and Flax, the library bridges forward cosmological simulations across Cosmic Dawn and the Epoch of Reionization with rapid, amortized parameter recovery.
To support and accelerate this research, I authored and maintain `21cmjaxSBI`.
Realistic Instrumental Noise: Incorporates dynamic Square Kilometre Array (SKA-1 Low) sensitivity curves directly into the forward training loop.
Simulator-Agnostic Design: Modular interface designed to pair seamlessly with varied cosmological simulators (e.g., `21cmFAST`) and custom lightcone mock datasets.
Citable Release: Officially published under `v0.1.0` and permanently archived via Figshare (`DOI: 10.6084/m9.figshare.34515834`).
Future extensions (`v0.2.0+`) will integrate higher-order topological summary statistics (Minkowski Functionals and Betti numbers) and cross-correlations with kinetic Sunyaev-Zel'dovich (kSZ) maps.
For more information, visit https://github.com/Subham0306/21cmjaxsbi