Starts 11 Jun 2026 16:00
Ends 11 Jun 2026 17:00
Central European Time
Search
Search in Conferences:
Seminar by Suman Majumdar - 'Emulator is all you need' for the Reionization Inference With SKA-Low
Hybrid
Leonardo Building - Euler Lecture Hall and via Zoom
Abstract
The cosmologically redshifted 21-cm signal originating from the neutral hydrogen distribution during the Epoch of Reionization (EoR) provides a direct probe to the state of the intergalactic medium in this formative phase of our Universe. The upcoming Square Kilometre Array (SKA-Low) is expected to observe this signal with unprecedented sensitivity and resolution. However, interpretation of such observations is complex due to the strong inherent and evolving non-Gaussian nature of the signal. In this talk, I will discuss our efforts towards the development of machine learning based 21-cm signal emulators designed to accelerate parameter inference by circumventing expensive simulations of this signal. Starting with the development of emulators based on Neural Networks for different summary statistics of the signal e.g. power spectrum and bispectrum, conditioned on reionization model parameters, we have recently developed CosmoUiT, a hybrid Vision Transformer-UNet architecture that emulates the entire time evolving 21-cm field as expected to be observed by the upcoming SKA-Low. CosmoUiT is orders of magnitude faster than simulations—while maintaining high fidelity to emulated maps--enabling field-level inference for future SKA observations.
The cosmologically redshifted 21-cm signal originating from the neutral hydrogen distribution during the Epoch of Reionization (EoR) provides a direct probe to the state of the intergalactic medium in this formative phase of our Universe. The upcoming Square Kilometre Array (SKA-Low) is expected to observe this signal with unprecedented sensitivity and resolution. However, interpretation of such observations is complex due to the strong inherent and evolving non-Gaussian nature of the signal. In this talk, I will discuss our efforts towards the development of machine learning based 21-cm signal emulators designed to accelerate parameter inference by circumventing expensive simulations of this signal. Starting with the development of emulators based on Neural Networks for different summary statistics of the signal e.g. power spectrum and bispectrum, conditioned on reionization model parameters, we have recently developed CosmoUiT, a hybrid Vision Transformer-UNet architecture that emulates the entire time evolving 21-cm field as expected to be observed by the upcoming SKA-Low. CosmoUiT is orders of magnitude faster than simulations—while maintaining high fidelity to emulated maps--enabling field-level inference for future SKA observations.