Lightweight AI Boosts Gamma-Ray Telescope Performance Without Needing GPUs

Astronomers hunting the universe’s most violent particle accelerators may have just gained a powerful new tool that runs on ordinary office computers. Researchers at the Bhabha Atomic Research Centre in Mumbai have developed a compact deep learning model that sorts cosmic signals from background noise in imaging atmospheric Cherenkov telescopes more effectively than the field’s standard machine learning methods, while consuming a fraction of the computational resources that modern artificial intelligence typically demands. When applied to real observations of the Crab Nebula made with India’s MACE telescope, the new approach delivered detection significances up to 25 percent higher than the conventional pipeline, a substantial gain in a field where every percentage point of sensitivity can determine whether a faint gamma-ray source is seen at all.

The challenge the team tackled lies at the heart of very-high-energy gamma-ray astronomy. When a gamma ray from a distant cosmic source slams into the upper atmosphere, it triggers a cascade of charged particles that race downward faster than light travels through air, producing a fleeting flash of blue Cherenkov radiation. Imaging atmospheric Cherenkov telescopes capture these flashes as faint images lasting only a few nanoseconds. But gamma rays are vastly outnumbered by cosmic rays, charged particles that produce strikingly similar atmospheric showers and images. Separating the two, a task known as gamma-hadron separation, is the critical filtering step that determines the quality of nearly every scientific result a Cherenkov telescope produces.

For decades, the standard approach has relied on ensemble learning methods, most commonly random forests, trained on so-called Hillas parameters. These parameters, first described by A.M. Hillas in 1985, compress each telescope image into a handful of moment-based quantities such as its width, length, orientation and asymmetry. The strategy is elegant and computationally cheap, but it deliberately discards much of the detailed structure contained in the raw camera images. In recent years, deep convolutional neural networks have shown they can learn directly from full images and often outperform hand-crafted features. Yet most deep learning efforts in this field have come with steep costs: elaborate image preprocessing pipelines, custom convolution kernels designed for hexagonal camera geometries, and heavy reliance on graphics processing units for both training and inference.

The new model, described in the journal Experimental Astronomy by Chinmay Borwankar, Praduman Pandey, Mradul Sharma and Kuldeep Yadav, takes a deliberately minimalist path. It consists of a simple two-layer convolutional neural network followed by an attention pooling mechanism. The architecture ingests a 25 by 25 pixel matrix extracted around the cleaned image, fed to the network without any of the spatial transformations that other approaches require, such as Gaussian kernel smoothing, oversampling onto square grids, or specialized hexagonal convolution kernels. This means the raw telescope image, once basic cleaning has removed isolated noise pixels, goes into the network essentially as the camera recorded it.

The most innovative element is the attention pooling step, which bridges the old world of Hillas parameters and the new world of deep learning. Rather than discarding the classical image moments, the model uses them to generate query vectors that provide global event-level context. These queries guide how features from the convolutional maps are aggregated, allowing the network to weigh the importance of different spatial regions of the image in light of the overall geometry of the shower. In effect, the model lets decades of accumulated domain knowledge about Cherenkov image shapes inform a neural network that simultaneously learns richer patterns directly from the pixel data, achieving a synthesis that neither approach could accomplish alone.

To benchmark the method, the researchers trained and compared their model against a standard random forest classifier on simulated imaging atmospheric Cherenkov telescope data, generated with established Monte Carlo tools for extensive air showers. The comparison showed enhanced background suppression for the convolutional network at all levels of signal acceptance, meaning that at every threshold a scientist might choose for accepting candidate gamma rays, the deep learning model retained more genuine signal while rejecting more cosmic-ray background. This kind of uniform improvement across the full operating range is particularly valuable because different analyses, from bright flaring sources to faint steady emitters, demand different trade-offs between signal purity and detection efficiency.

The decisive test came with real data. The team applied the model to observations of the Crab Nebula collected with the Major Atmospheric Cherenkov Experiment, or MACE, a 21-meter telescope standing at Hanle in the Indian Himalayas, one of the highest observatory sites in the world. The Crab Nebula, the remnant of a supernova witnessed by Chinese astronomers in 1054, serves as the standard candle of gamma-ray astronomy, observed by every major Cherenkov instrument from HESS to MAGIC to VERITAS because its gamma-ray output is steady and well characterized. Against this benchmark, the lightweight network achieved up to a 25 percent increase in detection significance relative to the standard analysis method, a striking result for an architecture deliberately kept as small as possible.

Equally remarkable is what the model does not require. The complete network can be trained on a 16-core CPU machine in roughly 90 minutes, and once trained it achieves an inference rate of about 2.96 kilohertz on a CPU, meaning it can classify nearly three thousand events every second without any graphics processing unit. For observatories operating in remote, high-altitude locations where power and cooling are limited, and for collaborations that cannot afford large GPU clusters, this efficiency transforms deep learning from a luxury into a practical daily tool. It also opens the door to real-time processing at the telescope, enabling rapid alerts when a transient gamma-ray source flares into view, precisely the kind of event that multi-messenger astronomy campaigns are eager to catch.

The work arrives at a moment when the field is preparing for the Cherenkov Telescope Array Observatory, a next-generation facility that will generate data volumes far beyond what current instruments produce. Recent deep learning efforts across the community, including frameworks such as CTLearn and GammaLearn applied to prototype telescopes, have demonstrated the promise of neural networks but often at significant computational expense. The Indian team’s results suggest a complementary philosophy: that lightweight architectures with minimal preprocessing can match or exceed heavier approaches while remaining simple enough for large-scale analyses. The authors note that their study is restricted to the single-telescope case, and extending the method to stereoscopic systems that combine images from multiple telescopes remains future work. Even so, the demonstration that a two-layer network with attention pooling, trained on commodity hardware, can beat a mature random forest pipeline on real data is likely to resonate across the gamma-ray community, offering a reminder that in the race to decode the high-energy sky, clever design can matter as much as raw computing power.

Subject of Research: A lightweight deep learning model for gamma-hadron separation in imaging atmospheric Cherenkov telescope data

Article Title: Efficient deep learning for single imaging cherenkov telescope event classification: improving over standard ML with minimal compute cost

Article References: Borwankar, C., Pandey, P., Sharma, M., & Yadav, K. (2026). Efficient deep learning for single imaging cherenkov telescope event classification: improving over standard ML with minimal compute cost. Experimental Astronomy, 62(2), Article 20. https://doi.org/10.1007/s10686-026-10080-w

Image Credits: AI Generated

DOI: 10.1007/s10686-026-10080-w

Keywords: gamma-ray astronomy, Cherenkov telescope, deep learning, convolutional neural network, attention pooling, gamma-hadron separation, MACE telescope, Crab Nebula, random forest, Hillas parameters, machine learning, high-energy astrophysics