Start of funding 01.07.2023

Exploration of deep-learning-based reconstruction algorithms for high-energy gamma-ray astronomy

Prof. Dr. Stefan Funk
Friedrich-Alexander-University of Erlangen-Nuremberg
ECAP-Erlangen Centre for Astroparticle Physics

Dr. Benjamin Nachman
University of California, Berkeley
Berkeley Institute for Data Science



The Cherenkov Telescope Array (CTA) is a next-generation flagship project in high-energy astrophysics and will provide astronomic observations of the universe at the highest energies and will provide unprecedented amounts of data.

Recent advancements in artificial intelligence (AI), i.e., deep learning, offer promising prospects for significant improvements in the sensitivity of CTA by analyzing the vast amounts of data the observatory provides, beyond simplistic approximations.

The collaboration between the Erlangen Centre for Astroparticle Physics and the Berkeley Institute for Data Science strives to establish a strong connection between fundamental research in physics and information technology. It focuses on applying state-of-the-art machine learning algorithms to CTA to advance and establish modern data-driven methods in the field of gamma-ray astronomy.

Final report:
The project enabled a research stay of Dr. Jonas Glombitza in the working group of Dr. Benjamin Nachman at the Lawrence Berkeley National Lab. During the stay, connections between the Californian and Bavarian Institute have been established that led to a first inter-experimental collaboration. The emerged joint project investigates the application of diffusion models to Imaging Air Cherenkov Telescopes (IACTs), like the future flagship project CTA.

The stay stimulated an exchange on the latest state of the art methods in computer science. This led to a new ansatz in the generative modeling of IACT simulations by utilizing diffusion models. On the one hand, this will enable to increase the robustness of deep-learning-based algorithms by making current simulation more data-like. On the other hand, this ansatz enables fast and cost-efficient simulation for the next-generation flagship CTA, and the search for anomalies in the data that may indicate new physics. Within the scope of this project improvements to previous published methods have been identified, providing a more accurate modeling of the data distribution for protons events, than generative adversarial networks that have been state-of-the-art.

The results (Elflein et al. 2024) have been accepted as contribution to the NeurIPS 2024, workshop ML4PS, and a detailed manuscript is in preparation.

The emerged successful project involves two postdocs, a PhD student, and a graduate student and motivated new potential collaborations between gamma-ray astronomy and information technology that are planned for the future.

C. Elflein et al., "Generation of Air Shower Images for Imaging Air Cherenkov Telescopes using Diffusion Models", ML4PS Workshop NeurIPS 2024, https://openreview.net/pdf?id=3by2lOSII6