Generation of Imaging Air Cherenkov Telescope images using Diffusion Models
and Lark Wang · 2026
Read Generation of Imaging Air Cherenkov Telescope images using Diffusion Models by and Lark Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-06364.txt.
Introduction
Elflein and his team present the first score based diffusion models making monoscopic gamma ray and proton shower images with nearly 2000 pixels. Tests use Wasserstein GANs on HESS simulations. Air shower simulations for instrument response are heavy and repetitive under varying conditions. Such models speed such work and provide a smooth device surrogate. Their use in gamma ray astronomy has been limited, especially for hadronic showers filling the background. Quality is checked with low level parameters and shower shape signs, while GAN methods remake gamma ray showers well.
Catalog record
| CREATOR | and Lark Wang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68024 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.06364
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Generation of Imaging Air Cherenkov Telescope images using Diffusion Models and Lark Wang Abstract Substantial amounts of air-shower simulations are needed to derive the instrument response for analyzing Imaging Air Cherenkov Telescope (IACT) data. This process is both computationally intensive and requires repetition under varying observation conditions, due to detector aging, changes in the atmosphere, or the instrument hardware. Generative models offer an efficient alternative, significantly accelerating simulations while compactly storing extensive simulation libraries, and providing a differentiable surrogate model of the instrument. However, their applicability has so far been limited in gamma-ray astronomy, particularly for modeling hadronic showers that dominate the background and exhibit significant intrinsic fluctuations that are challenging to model. In this study, we present the first application of score-based diffusion models to generate monoscopic γ \gamma -ray and proton shower images with nearly 2,000 pixels and benchmark the performance against Wasserstein GANs using H.E.S.S. simulations. We examine quality using both low-level parameters and well-established shower-shape observables, and assess analysis readiness via state-of-the-art γ \gamma –hadron separation. While GAN-based approaches can reproduce γ \gamma -ray showers with high fidelity, they fail to generate proton events of comparable quality, particularly in capturing intrinsic higher-order correlations, leading to a measurable degradation in analysis performance. In contrast, score-based diffusion models achieve significantly superior quality for γ \gamma -ray and proton showers, accurately reproducing high-level correlations and generating events that are statistically indistinguishable from simulations at the analysis level. These results establish diffusion-based models as the first analysis-ready surrogate model of a single IACT, opening new prospects for fast instrument response generation, detector optimization, and connected downstream tasks. 1 Introduction Over the past two decades, \gamma -ray observatories, and particularly arrays of Imaging Air Cherenkov Telescopes (IACTs), have changed our understanding of the very-high-energy (VHE) \gamma -ray sky. A key facility is the High Energy Stereoscopic System (H.E.S.S.) in Namibia, which studies cosmic \gamma rays with energies in the GeV to TeV range [ 5 ] in the Southern hemisphere. IACTs image Cherenkov radiation emitted from secondary particles in extensive air showers induced by cosmic particles. The recorded IACT images are then analyzed to extract information about the primary particle. Traditionally, the so-called Hillas parameters [ 43 ] , which describe the shape of the shower, are used to estimate the energy and reconstruct the arrival direction of the incoming primary particle, as well as to reject hadronic showers that outnumber \gamma -rays by 1 to 10^{3}-10^{4} . Figure 1 : Various stages of the generation process of an IACT image starting from Gaussian noise using a score-based diffusion model. The particle causing the signal is a proton with an energy of 5.5 TeV. The black pixels denote zero-value pixels and brighter pixels denote higher values. The analysis of \gamma -ray data requires a thorough understanding of the instrument response functions (IRFs), which are used to interpret observations by forward-folding theoretical models, e.g., by using the gammapy software [ 20 ] . Large-scale Monte Carlo (MC) simulations are required to precisely derive the IRFs. Running these simulations, particularly for the hadronic background, is computationally demanding and time-intensive [ 42 , 10 ] , as classifiers reach rejection powers in the order of 10^{-3} and higher. Running millions of events, each requiring in the order of 30–60 seconds on one CPU core, generates substantial data, typically hundreds of gigabytes per observation period. Furthermore, as the instrument ages, new simulations are needed to account for small variations in performance. Recent advances in deep learning [ 56 ] provide novel perspectives on data analysis in physics 1 1 1 For a pedagogical introduction, see Ref. [ 25 ] . and high-energy-physics research [ 29 , 37 ] . Particularly, generative models offer novel opportunities for simulating high-dimensional data [ 39 , 53 ] . In particle physics, Generative Adversarial Networks (GANs) [ 68 , 69 , 28 , 31 , 60 ] , normalizing flows [ 54 ] , and diffusion models [ 62 ] have been used to construct surrogates of the instrument and accelerate simulations by five to six orders of magnitude. Deriving these surrogate models can be used to adapt simulations to slightly different conditions [ 81 , 19 ] and refine simulations to match measured data [ 27 ] . Furthermore, these differentiable surrogates unlock a broad spectrum of downstream tasks, including instrument optimization [ 21 ] , event reconstruction [ 22 ] , anomaly detection [ 52 ] , unfolding [ 46 ] , and the modeling of non-perturbative processes, such as hadronization [ 47 , 33 ] , where first-principles descriptions are currently unavailable. Addressing the computational needs of future observatories like the Cherenkov Telescope Array Observatory (CTAO) [ 8 ] , recent studies have explored generative models for fast and memory-efficient generation of high-level IACT parameters [ 17 ] or complete IACT images using GANs [ 49 , 48 , 45 , 70 ] , to obtain surrogates of the full-scale simulations. Also, in recent years progress has been made in deep-learning-based reconstructions for \gamma -ray observatories [ 74 , 65 , 12 , 50 , 77 , 34 , 16 , 73 , 35 ] and cosmic-ray observatories [ 3 , 26 , 2 , 1 ] . Early studies have already produced convincing results for generating monoscopic images of \gamma -ray showers [ 24 , 23 ] . However, generating detailed images of Imaging Atmospheric Cherenkov Telescopes (IACTs) with comprehensive fidelity remains a challenge, particularly for hadrons, for which large shower-to-shower fluctuations in the shower development exist. These fluctuations are characterized by distinctive subshowers, manifesting themselves as hadronic substructure in the Cherenkov images, making them more irregular and patchy — distinctly different from the relatively homogeneous and elliptical shapes of \gamma -ray images. Consequently, while current generative models can reproduce \gamma -ray image features, their limited ability to capture these hadronic fluctuations prevents the models from being analysis-ready. Inspired by the application in particle physics calorimeters [ 62 , 63 ] , we use for the first time state-of-the-art score-based diffusion models (SBDM) [ 76 ] in astroparticle physics. Using simulations from the H.E.S.S. CT5 telescope with its FlashCam design [ 32 ] , the camera of the CTAO MSTs on the southern site, we study the application of diffusion models to the generation of IACT images. A simple illustration of the generation process using the SBDM is shown in figure 1 using an IACT image from a proton-induced air shower. We demonstrate that diffusion models provide accurate generation of both \gamma -ray and proton images, outperforming previous approaches based on generative adversarial networks (GANs), which show good performance in generating \gamma -ray images. Finally, we assess the readiness for analysis of both approaches by applying state-of-the-art \gamma -hadron separators to the events generated by WGAN and SBDM. Since stereoscopic integration is an ongoing challenge in deep learning [ 74 , 12 , 50 , 77 , 34 , 82 ] , we limit ourselves in this work to monoscopic, i.e., single-telescope images. 2 Generation of IACT images using deep learning In recent years, particularly generative adversarial networks (GANs), have gained significant attention for their impressive results in generating natural images and have been applied