Machine Learning in Particle Physics: UCL’s REM Lifts Efficiency to 92%

Machine learning in particle physics has mostly been used to tune individual steps of the long chain of algorithms that turns raw detector signals into particles. A team at University College London (UCL) now says a single model can do the whole job. Its Reconstruct Everything Model, or REM, rebuilds particles from every major detector system at once, and it raised reconstruction efficiency from 61% to 92% in a simulated collider test.

UCL announced the work on 2 November 2026. It was developed at UCL’s Centre for Data Intensive Science and Industry and is described in a paper titled “Reconstruct everything everywhere all at once.” The project was led by Max Hart and Kwok Yiu (Sunny) Wong, working with Samuel Van Stroud, Philippa Duckett, Tim Scanlon and Gabriel Facini.

Why particle reconstruction needs machine learning

When particles collide, a detector records millions of tiny signals called hits. Physicists then have to work backward from those hits to figure out which particles were produced. Traditionally, this is done with a chain of specialized algorithms, each handling one detector system, such as the silicon tracker, the calorimeters that measure energy, or the muon system.

Each step in that chain makes its own assumptions, and errors can carry from one step to the next. Future colliders will make the problem harder. The team tested REM on a simulation of the CLD detector proposed for CERN’s Future Circular Collider (FCC-ee), which would have more than 10 billion readout channels.

How the Reconstruct Everything Model works

REM is a transformer model based on the MaskFormer architecture, a design originally built for image segmentation. It uses 512 “object queries,” which act like slots the model fills with candidate particles. For each one, it predicts which hits belong to the particle and what kind of particle it is: an electron, a muon, a photon, a charged hadron or a neutral hadron.

Because it looks at the tracker, the calorimeters and the muon system together, REM does not need the traditional chain of algorithms. The team says it goes further still and reconstructs W, Z and Higgs bosons directly from their decay products, an approach it calls going from “hits to Higgs.”

What the machine learning results show

In the simulated FCC-ee tests, UCL reports these gains over conventional reconstruction:

• Particle reconstruction efficiency rose from 61% to 92% (91.7% in the paper).
• Jet energy resolution improved by 52%.
• Separation of W and Z bosons improved by about 50%.
• Processing one collision took about 30 milliseconds on a single GPU.

Better jet energy resolution matters because many important particles, including the Higgs boson, often decay into sprays of particles called jets. Telling W and Z bosons apart is also a long-standing challenge, since their masses are close.

What machine learning in particle physics could change next

The results so far come from simulation, not from a running experiment. The FCC-ee is still a proposal, so the model has not yet been tested on real collisions from that detector. Even so, a unified model could make it easier to design detectors and analysis software together, rather than building and tuning dozens of separate algorithms.

In my view, the most interesting part is the speed. A full reconstruction in about 30 milliseconds on one GPU suggests this kind of model could be practical at scale, not just accurate on paper. It is another sign of how quickly transformer models are spreading beyond language, a trend we also saw when an AI weather model topped an ECMWF contest. For a closer look at how these models learn, see our explainer on how diffusion models move from copying to creating.