Optimising Geometric Deep Learning for Varied Reconstruction Tasks in High Granularity Calorimeters

In the heart of modern physics, scientists are building machines to see the universe at its smallest scales. These machines, called particle detectors, act like incredibly fast and detailed cameras, capturing the fleeting traces left behind when subatomic particles smash into each other at near-light speeds. As these machines become more powerful, they are becoming more detailed, too. Imagine a camera that used to take a picture with a few thousand pixels, and now takes one with millions. This increase in detail, known as granularity, allows physicists to see individual particles more clearly, even when they are crowded together. However, this flood of detail creates a massive problem: the sheer amount of data is so large that traditional computer methods struggle to process it quickly enough to be useful. The challenge is not just to look at the data, but to sort through the chaos of millions of tiny signals to figure out exactly what happened in a split second.

To solve this, researchers at the École Polytechnique in France have developed a new way to teach computers how to understand these complex patterns. They turned to a type of artificial intelligence called a Graph Neural Network. Instead of treating the detector data as a simple grid of numbers, like a standard photograph, they treat it as a map of connections. In this map, every tiny signal from the detector is a point, and the lines connecting them represent how close those signals are to one another. This approach is much better suited to the messy, irregular shapes that particles leave behind. However, building these maps is usually very slow and expensive for a computer to do, especially when there are millions of points to connect. The researchers realized that because the detector's physical layout is fixed and known, they could use that knowledge to make the process much faster without losing any accuracy.

The team tested their ideas on a simulated version of a high-tech detector, similar to one planned for the Large Hadron Collider. They created a system that could take raw, messy data and turn it into clear answers about what particles were present. First, they built a tool to identify the type of particle. In a crowded environment, an electron, a photon, a muon, or a pion can look very similar. The new system successfully distinguished between five different types of particles, including tricky cases where two photons overlap. It did this by learning the unique "shape" of the energy deposit each particle leaves behind. For example, muons leave a single, thin track, while electrons and photons create a spreading shower of energy. The system was so effective that it could tell the difference between a single electron and a pair of photons with high precision, a task that is notoriously difficult for older methods.

Next, the researchers tackled the job of measuring the energy of these particles. When a particle hits the detector, it leaves behind a trail of energy deposits. The goal is to add up these deposits to find the total energy of the original particle. The team created a model that could do this with remarkable accuracy, matching the theoretical limits of what the detector could achieve. They also showed that this process could be split up and run on many computer processors at the same time. By dividing the detector into sections and processing them in parallel, they could speed up the calculation significantly without sacrificing the quality of the result. This is a crucial step toward using these systems in real-time, where decisions must be made in microseconds.

Finally, the team addressed the most difficult scenario: when two particles hit the detector at almost the exact same spot, their signals overlap and blur together. This is like trying to read two overlapping handwritten notes. The researchers built a system that could separate these mixed signals, figuring out exactly how much of the energy belonged to the first particle and how much to the second. They achieved this by training the computer to predict where each particle entered the detector and how its energy was distributed. While the system was not perfect, it performed as well as much more complex and slower methods, proving that a simpler, faster approach could work just as well.

All these separate tools were then combined into a single, flexible blueprint called HIBOU. This design shows how different parts of the reconstruction process can work together in a modular way. One part identifies the particle, another measures its energy, and a third separates overlapping events. If the system identifies a particle as a muon, it takes a simple path. If it sees a complex, overlapping event, it routes the data to a more specialized tool. This structure allows the system to adapt to different tasks without needing to be rebuilt from scratch. The researchers found that while this modular approach is powerful, it requires careful handling to ensure that a small mistake in the first step doesn't get magnified in the next. They suggested that training the entire system together, rather than piece by piece, would help prevent these errors.

The work demonstrates that by using the known geometry of the detector to guide the artificial intelligence, scientists can overcome the computational bottlenecks that usually limit high-detail experiments. The methods developed here are not just theoretical; they were tested on simulated data that mimics the conditions of future particle colliders. The results show that it is possible to process vast amounts of data quickly and accurately, opening the door for next-generation experiments to see the universe with unprecedented clarity. The researchers have provided a clear path forward, showing that with the right tools, the flood of data from these massive machines can be turned into a stream of discovery.