⚛️ high-energy experiments

Reconstruct everything everywhere all at once: An end-to-end model for future colliders

This paper introduces the Reconstruct Everything Model (REM), a pioneering end-to-end framework that directly transforms raw detector hits into reconstructed particles and resonances, achieving superior efficiency and resolution compared to conventional and machine-learned particle-flow methods in high-energy physics simulations.

Original authors: Max Hart, Kwok Yiu Wong, Samuel Van Stroud, Philippa Duckett, Tim Scanlon, Gabriel Facini

Published 2026-11-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Max Hart, Kwok Yiu Wong, Samuel Van Stroud, Philippa Duckett, Tim Scanlon, Gabriel Facini

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand the work of particle physicists, one must first imagine the machine they use: a colossal, ring-shaped laboratory buried deep underground or stretching across a landscape, designed to smash particles together at speeds approaching that of light. When these tiny constituents of matter collide, they do not simply bounce off one another; they transform, creating a shower of new, often fleeting particles that fly out in every direction. To make sense of these collisions, scientists rely on detectors that are essentially giant, three-dimensional cameras built from layers of different materials. As the new particles fly through these layers, they leave behind traces: tiny electrical signals in silicon sensors, flashes of light in crystal blocks, or faint tracks in gas-filled chambers. The challenge is not just to see these traces, but to reconstruct the invisible story they tell. The raw data is a chaotic mess of millions of individual signals, and the traditional method of sorting them involves a long, rigid assembly line of separate software programs. Each program handles one specific task, such as grouping nearby signals into a track or measuring the energy of a flash, passing the result to the next program in the chain. This step-by-step approach, while reliable, often loses subtle connections between the different layers of the detector, limiting how clearly scientists can see the fundamental processes at work.

A team of researchers at University College London has proposed a different way to look at this problem, one that abandons the assembly line in favor of a single, unified system. They have developed a model called the Reconstruct Everything Model, or REM, which acts as a complete, end-to-end learner for particle physics. Instead of breaking the job into separate stages, this system takes the raw, unprocessed signals from every part of the detector—the silicon sensors, the energy-absorbing calorimeters, and the outer muon detectors—and processes them all at once. It learns to recognize patterns directly from the noise, assigning each tiny signal to the specific particle it came from, identifying what that particle is, and calculating its speed and energy, all in a single pass. To test this idea, the researchers simulated a future particle collider known as the Future Circular Collider, specifically using a proposed detector design called CLD. This simulated machine is immense, containing more than ten billion readout channels, far more than any current detector. In their simulations, the REM model successfully reconstructed the entire event, turning millions of raw hits into a clear picture of the particles produced, and even identifying the heavier parent particles that decayed to create them.

The results of this simulation show a significant improvement over the standard methods currently used in the field. When the researchers compared REM to the conventional software pipeline, they found that the new model was far more efficient at finding particles and much better at correctly identifying them. For example, in events where particles decay into jets of other particles, the new model reduced the uncertainty in measuring the energy of those jets by more than half compared to the traditional approach. It also proved particularly adept at distinguishing between different types of particles that are notoriously difficult to tell apart, such as electrons and photons, or neutral particles that leave no track in the inner sensors. Perhaps most notably, the model was able to reconstruct the mass of the original heavy particles, like the W and Z bosons, with greater precision. In the simulated collisions, the new system could separate the signals of these two different particles much more clearly than the old methods, a crucial capability for spotting rare or new physics.

What makes this achievement distinct is not just the accuracy, but the architecture of the solution itself. The model is built on a type of artificial intelligence known as a transformer, which is the same technology behind many modern language tools, but here it is applied to the spatial and energetic data of a particle detector. The system learns to treat the detector as a single, interconnected sensor array rather than a collection of separate instruments. It does not require pre-processed tracks or clusters as inputs; it starts with the raw hits and builds the entire picture from scratch. This allows it to use information from one part of the detector to help interpret signals in another, a capability that the traditional, separated software cannot easily replicate. In the simulation, the model processed a full collision event in about thirty milliseconds on a single graphics card, a speed that suggests it could handle the high data rates of future experiments without the need for massive computing farms.

The researchers also tested the model in more crowded and complex environments, simulating collisions that produce a dense spray of particles, similar to what might be seen in heavy quark interactions. Even in these chaotic conditions, the model maintained high accuracy, correctly identifying the vast majority of particles and keeping the number of false alarms low. It successfully reconstructed the paths of charged particles and the energy deposits of neutral ones, demonstrating that a single learned framework can handle the diverse and sparse data coming from different detector technologies. The study also showed that the model could go a step further than just identifying individual particles; it could group them back into the parent particles they came from, effectively reversing the decay process to reveal the original collision event. This "hit-to-boson" capability suggests a future where the entire chain of reconstruction, from the first signal to the final physics result, is learned and optimized as one continuous process.

While these results are based on computer simulations and not yet on real data from a physical machine, they offer a compelling vision for the future of high-energy physics. The work demonstrates that it is possible to replace a complex, multi-stage software pipeline with a single, flexible model that learns to see the whole picture at once. By retaining the intermediate steps of its own reasoning, the model also provides a way to inspect and validate its decisions, ensuring that the results remain transparent and trustworthy. As the field moves toward more powerful colliders that will generate even more data, the ability to process this information efficiently and accurately will be essential. The Reconstruct Everything Model suggests that the key to unlocking the secrets of these future machines may lie in teaching the computer to see the entire event, everywhere, all at once.

Technical Summary: Reconstruct Everything Model (REM)

Problem Statement
High-energy physics collider experiments rely on complex, multi-stage reconstruction pipelines to convert raw detector signals (hits) into physical objects (particles, jets, resonances). Traditional workflows, such as those used at the Large Hadron Collider (LHC) and proposed for future colliders like the Future Circular Collider (FCC), operate in a modular fashion. They typically separate tasks into distinct subsystems: silicon tracking, calorimeter clustering, and particle flow. These stages are often optimized independently, requiring external reconstruction of tracks or clusters before subsequent steps can begin. This separation limits the exploitation of complementary information across detector subsystems, complicates adaptation to new detector conditions, and creates computational bottlenecks due to the combinatorial complexity of traditional algorithms (e.g., Kalman filtering) at high particle densities. While machine learning has been applied to individual stages (e.g., graph neural networks for tracking or deep learning for calorimeter clustering), no prior framework has demonstrated a single, unified model performing end-to-end reconstruction directly from raw hit-level data across all detector subsystems.

Methodology
The paper introduces the Reconstruct Everything Model (REM), an end-to-end, multitask deep learning framework designed to reconstruct multiple particle types directly from raw detector hits in the silicon tracker, electromagnetic/hadronic calorimeters (ECAL/HCAL), and muon system.

Architecture: REM adapts the MaskFormer architecture (inspired by object detection and segmentation) to sparse, heterogeneous, multimodal sensor data. It employs a two-stage encoder-decoder transformer pipeline:
1. Particle Reconstruction Stage: Takes raw hits from six subsystems (Vertex, Inner/Outer Tracker, ECAL, HCAL, Muon) as input. It uses a self-attention encoder with a sliding window mechanism to handle the quadratic complexity of full attention on millions of hits. The decoder utilizes a fixed set of 512 learned object queries to simultaneously perform hit assignment (tracking and clustering), particle classification, and kinematic regression.
2. Resonance Reconstruction Stage: A secondary encoder-decoder transformer takes the particle query embeddings from the first stage and assigns them to candidate resonance bosons (W, Z, H), predicting both the boson identity and the constituent particle mask.
Input Representation: The model ingests digitized detector hits without pre-reconstructed tracks or clusters. Features include Cartesian coordinates, radial/angular features, and logarithmic energy deposits.
Training Strategy: Training proceeds in stages with earlier networks frozen to stabilize learning. The loss function combines pattern recognition (cross-entropy, Dice, Focal, IoU), kinematic regression (Barron loss for robustness), and resonance assignment losses. The model is trained on simulated $e^+e^-$ collision events at $\sqrt{s} = 240$ GeV (diboson and ZH), 91 GeV (Z), and 365 GeV ( $t\bar{t}$ ), using the CLD (CLIC-like Detector) concept.
Baselines: REM is compared against:
- PANDORAPFA: The standard conventional particle-flow algorithm.
- FLOWFORMER: A learned model using the same transformer architecture but taking externally reconstructed conformal tracks as input (similar to HITPF).
- HITPF: A specialized learned particle-flow model.

Key Contributions

First End-to-End Demonstration: REM is presented as the first single learned framework to simultaneously reconstruct multiple particle types directly from hit-level measurements across the tracker, calorimeters, and muon system in a full collider simulation, without relying on externally reconstructed tracks or clusters.
Unified Reconstruction Chain: The model integrates hit assignment, particle identification, kinematic regression, and resonance reconstruction into a single pipeline, retaining intermediate outputs (hit assignments, tracks, clusters) for validation.
Resonance-Level Learning: The framework extends beyond particle reconstruction to learn the assignment of particles to resonance bosons (W, Z, H) and predict their identities, effectively creating a "hits-to-Higgs" chain.
Performance Gains: The paper demonstrates that learning from raw hits improves performance across a wide range of observables compared to both conventional methods and other machine-learning approaches that rely on intermediate reconstructed objects.

Results
Evaluated on the CLD detector concept at $\sqrt{s} = 240$ GeV:

Particle Reconstruction: REM achieves an integrated reconstruction efficiency of 91.7% and purity of 92.2% after particle identification. This significantly outperforms PANDORAPFA (61.1% efficiency, 73.0% purity) and FLOWFORMER (89.0% efficiency, 91.7% purity). The largest gains are observed in electron identification (92–94% for REM vs. 75–78% for FLOWFORMER) and neutral hadron reconstruction.
Kinematic Resolution: REM provides superior energy resolution. The jet energy resolution (JER) is 2.19%, compared to 2.88% for FLOWFORMER and 4.61% for PANDORAPFA. The visible energy response interquartile range (IQR) is reduced by approximately 36% relative to PANDORAPFA.
Resonance Reconstruction: In diboson events, the learned resonance assignment improves the separation of W and Z mass distributions. The separation metric $S$ reaches 3.03 for REM (using resonance masks), compared to 2.87 for REM using standard jet clustering and 2.02 for PANDORAPFA.
High-Occupancy Performance: In high-density $t\bar{t}$ events at 365 GeV, REM maintains high efficiency (91.2%) and purity (91.7%), significantly outperforming PANDORAPFA (61.7% efficiency).
Timing: REM performs the full reconstruction pass in approximately 30 ms on a single GPU, comparable to or faster than conventional CPU-based pipelines when accounting for the removal of external tracking steps.

Significance and Claims
The paper claims that REM demonstrates the feasibility of a complete, learned reconstruction chain from detector hits to composite resonance objects. Its primary significance lies in:

Simplification: It offers a route to simpler reconstruction workflows by unifying stages that are traditionally separate and optimized independently.
Performance: It achieves state-of-the-art performance in particle efficiency, purity, and resolution, particularly in challenging environments with overlapping showers and high particle densities.
Interpretability: Unlike "black box" approaches, REM retains intermediate outputs (hit assignments, tracks, clusters), allowing for the inspection and validation of each stage of the reconstruction.
Generalizability: The authors suggest the approach of treating the detector as a sparse, heterogeneous sensor array could be applicable to other large-scale, multimodal sensor systems beyond high-energy physics.

The paper remains modest regarding future deployment, noting that robustness to beam-induced backgrounds, detector defects, and misalignment must still be established, and that validation with alternative event generators is required before experimental use.

Technical Summary: Reconstruct Everything Model (REM)

More like this