ERC grant recipient Sebastien Rettie: “We’re going to reinvent track finding”

What is the focus of this ERC proposal?

“From a physics perspective, it’s about the nature of the Higgs particle. It was discovered in 2012, but what exactly does it do, and is it exactly what we expect in the Standard Model, or is it slightly different?”

Why is that important?

“The Higgs particle is what we can measure of the Higgs field—the field that gives particles mass in the Standard Model. The exact shape of that field determines what happened around the time of the Big Bang. That’s where the answers lie to the question of why the universe contains matter and not antimatter. About how the expansion suddenly accelerated incredibly—cosmic inflation. About how stable the universe is.”

Big questions.

“Certainly, and there’s really only one way to study that: through processes in which two Higgs particles are created simultaneously. In such di-Higgs events, we can study the purest form of Higgs interaction.”

When are two Higgs particles produced simultaneously?

“In the LHC accelerator, protons collide at high energy, occasionally producing Higgs particles and, very rarely, two Higgs particles at once. We think we can find the first hints of di-Higgs events in all the LHC data combined. But they won’t be truly measurable until the accelerator’s upcoming high-luminosity phase.”

Provided their tracks can be found in the crowded detector.

“Our UNMASK project aims to reinvent track reconstruction using modern pattern recognition techniques. These are rapidly gaining ground in language models and medical imaging. We want to introduce them into particle physics.”

How does that work, roughly?

“In classical track reconstruction, you have a detector with sensors that light up here and there, and the question is which signals belong together. Traditionally, you start at the center and look for nearby points that connect to it, but that only works if the image isn’t too crowded. A modern technique like MaskFormer—which is used, for example, by META for Facebook—views the entire illuminated detector as a single image and searches for coherent structures within it. Tracks are then recognized as image elements.”

The model has to learn that, though.

“Absolutely. But we’ve already shown with a sample model that it’s possible. Before I joined Nikhef, we had already published promising results on this topic.”

So the technique already exists.

“We’re going to further develop that in UNMASK for ATLAS and, more broadly, for particle physics, and ensure that it works and is available when HiLumi goes online in a few years. In the meantime, we’re working on it using data from LHC Run 3, searching for evidence of the di-Higgs. That dataset is far from having been thoroughly analyzed.”

So there’s a lot coming up. A lot of work, too.

“Definitely. And technology isn’t the only thing here. For example, it will also be sociologically interesting to implement this. Track reconstruction is the central component of experimental particle physics. A shift to AI is a paradigm shift, and that certainly won’t happen on its own. That, too, is just hard work.”

In the Netherlands, far away from Canada or CERN.

“It’s great here. The only real downside is that there are no mountains here. For someone like me, who loves the mountains, there isn’t much skiing or climbing to do.”