Leveraging Artificial Intelligence to Enhance Gamma-Ray Detection Techniques

Written by Daniel Nieto , Universidad Complutense de Madrid (UCM) , and matthieu heller , Université de Genève

Although the concept of Artificial Intelligence (AI) dates back centuries, with ancient Greek and Egyptian mythology featuring artificial beings with human-like intelligence, recent advances in one of its primary pathways, machine learning, have brought AI back into the spotlight. Popular applications, including generative tools, search engines, and similar technologies, have captured public interest. The impact these AI tools are having on our daily lives is exciting, but they also have a groundbreaking impact on science. You only have to look to the recent 2024 Nobel Prize winners in Physics and Chemistry to understand why.

Originating from condensed matter physics and statistical physics, John Hopfield and Geoffrey Hinton developed powerful artificial neural networks that pushed the boundaries of machine learning and artificial intelligence, earning them the 2024 Nobel Prize in Physics for these advancements. Similarly, Demis Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for their AlphaFold2 AI model, which significantly advanced protein structure prediction.

But what exactly is machine learning? Simply put, it is a method that enables computers to learn independently. By supplying them with data (such as images or patterns), computers gradually learn to recognise and identify these data on their own. When we are young, for example, we learn that a dog is an animal with certain features, like a snout, fur, and a tail. We don’t need to see every dog in the world to distinguish it from other animals. This is the goal with computers, as well. There are different types of machine learning methods. Most of these applications run on artificial neural networks: computational models inspired by the brain's own neural wiring.

The machine learning revolution is transforming and enhancing science across nearly every research field, including gamma-ray astronomy. Since the early 2000s, machine learning has played a crucial role in imaging atmospheric Cherenkov telescopes (IACTs), the type of telescopes that the CTAO uses to study the extreme Universe.?

When gamma rays enter the Earth’s atmosphere, they interact with it, generating cascades of subatomic particles. A by-product of these cascades is Cherenkov light, which IACTs’ mirrors reflect onto their high-speed cameras capable of capturing a billion frames per second. That Cherenkov light leaves an elliptical-shaped pattern on the cameras. Scientists then extract features from these recorded events, such as the length of the ellipse's major axis, to train machine learning models. With this, computers learn to distinguish gamma-ray events from other irrelevant data and can even predict the path and energy of the original gamma ray based on the event’s shape.??

But the process can be improved. Today, we are applying advanced, neural-network based machine learning strategies that process entire events as inputs, rather than relying on individual features, to enhance the capabilities of the next generation of IACTs, specifically for the CTAO. By shifting to general AI models instead of domain-specific analyses, we can leverage rapid advancements from diverse fields and applications, enabling us to incorporate state-of-the-art techniques that are continually evolving across the AI landscape. This approach is anticipated for use in the so-called offline data analysis, that is, once the telescope data arrives at the data centres.

Additionally, progress in hardware accelerators, such as FPGAs and GPUs — designed for executing AI-related operations — opens up new possibilities for applications in this field.?For example, the rapid growth of Edge ML, a practice that focuses on running machine learning algorithms on devices with limited computing resources (like smartphones), offers new tools and methods to simplify and optimise complex models. This enables real-time data processing and decision-making at the data source, reducing latency to almost no delay and bandwidth costs by minimising data transmission, as the processing happens directly on the device.

In this context, our goal is to adapt AI algorithms developed for offline data analysis to support real-time processes to capture gamma rays within the future CTAO cameras and data acquisition hardware. If successfully implemented, this approach could improve sensitivity to lower energy levels and enhance data reduction at the earliest stages of acquisition.

Several teams within the CTAO are actively working on optimising these machine learning models for both offline and real-time data analysis, alongside developing hardware capable of running them effectively. And as these AI tools continue to rapidly grow and advance, we can only anticipate that the impacts will become even more profound on our mission to unveil the mysteries of the Universe.

Stay Informed! Subscribe to Flashpoint, CTAO's LinkedIn newsletter, and visit the News Section on our website for the latest insights.