End-to-end event reconstruction for precision physics at future colliders

This paper presents an end-to-end global event reconstruction framework combining geometric algebra transformers and object condensation clustering that significantly outperforms state-of-the-art rule-based algorithms in efficiency, fake-particle suppression, and mass resolution for future collider experiments like FCC-ee, thereby enabling rapid detector design iteration by decoupling performance from detector-specific tuning.

The Gist

Imagine you are trying to solve a massive, chaotic jigsaw puzzle, but instead of picture pieces, you have millions of tiny, glowing sparks flying in every direction. This is what happens inside a particle collider like the future FCC-ee. When particles smash together, they explode into a shower of new particles. To understand the physics of the universe, scientists need to figure out exactly what those original particles were, how heavy they were, and where they were going.

This paper introduces a new, smarter way to solve that puzzle, called HitPf.

Here is the breakdown of how it works, using simple analogies:

1. The Old Way: The "Assembly Line" with Strict Rules

For decades, scientists have used a method called Particle Flow. Think of this like an old-school factory assembly line.

• Step 1: A worker looks at the sparks in the "calorimeter" (a giant energy-measuring box) and groups them into piles based on strict rules (e.g., "If sparks are close together, they must be one object").
• Step 2: Another worker looks at the "tracks" (paths left by charged particles) and tries to match them to the piles.
• Step 3: A third worker decides what the object is (a photon? a proton?) based on a checklist.

The Problem: This process is rigid. If the factory layout changes (a new detector design), you have to fire the workers, rewrite the rulebook, and train everyone from scratch. It's slow, and if the sparks are messy (overlapping), the workers often get confused, merging two different objects into one giant, wrong pile.

2. The New Way: The "Super-Intelligent Detective" (HitPf)

The authors propose HitPf, which skips the assembly line entirely. Instead of following a rulebook, it uses a neural network (a type of AI) that acts like a super-intelligent detective.

• Direct Observation: Instead of grouping sparks first, the AI looks at every single spark and every single track simultaneously. It sees the whole picture at once.
• Geometric Algebra (The "3D Brain"): The AI doesn't just see numbers; it understands geometry. Imagine the AI has a 3D brain that understands how shapes, angles, and volumes relate to each other naturally. It knows that a "track" leading into a "spark pile" usually means they belong together, without needing a manual rule to tell it so.
• Object Condensation (The "Magnet" Effect): The AI uses a technique called "object condensation." Imagine every spark has a tiny magnet. Sparks belonging to the same particle are attracted to each other and clump together naturally. Sparks from different particles repel each other. The AI finds the "center of gravity" for each clump to identify the particle.

3. Why is this a Big Deal?

The paper tested this new AI against the current gold-standard method (called PandoraPfa) using simulated collisions. The results were impressive:

• Better at Untangling Knots: In crowded areas where particle showers overlap (like two fireworks exploding at the same time), the old method often merged them into one giant blob. HitPf successfully separated them, like a detective untangling two knotted necklaces.
• Fewer "Fake" Particles: The old method sometimes invented particles that didn't exist (false alarms). HitPf reduced these fake alarms by a huge margin (up to 100 times fewer!).
• Sharper Vision: Because it separates particles better, the measurement of their energy and mass is much more precise. The paper says the "blur" in the image was reduced by 22%.
• The "Plug-and-Play" Advantage: This is the most exciting part for the future. If engineers change the shape of the detector (like remodeling a house), the old method requires months of re-tuning the rules. HitPf just needs to be "re-trained" for about two days on powerful computers. It learns the new layout from scratch automatically.

The Bottom Line

Think of the old method as a human librarian trying to sort books by following a very specific, rigid card catalog system. If the library changes its layout, the librarian is lost.

HitPf is like a genius librarian who walks into the room, looks at every book and every shelf instantly, understands the relationships between them, and organizes them perfectly, no matter how the room is arranged.

By using this "end-to-end" AI approach, future particle colliders will be able to see the universe with much sharper eyes, allowing scientists to discover new physics that was previously hidden in the blur.

Technical Summary

Here is a detailed technical summary of the paper "End-to-end event reconstruction for precision physics at future colliders."

1. Problem Statement

Future collider experiments, specifically the FCC-ee (Future Circular Collider), aim to measure Higgs couplings, electroweak parameters, and flavor observables with unprecedented precision. The sensitivity of these measurements is directly limited by the resolution with which visible final-state particles and their invariant masses can be reconstructed.

Current state-of-the-art reconstruction relies on Particle Flow (PF) algorithms (e.g., PandoraPfa), which use a sequential, rule-based approach:

1. Iterative clustering of calorimeter hits.
2. Association of these clusters with charged particle tracks.
3. Identification of neutral particles.

Key Limitations of Current Methods:

• Detector Specificity: These algorithms require extensive, manual tuning for specific detector geometries. Changing a detector design necessitates re-optimizing the entire reconstruction chain, hindering rapid iteration during the design phase.
• Intermediate Steps: They rely on pre-formed clusters. In dense hadronic environments (common at FCC-ee), this leads to shower merging, where overlapping electromagnetic and hadronic showers are incorrectly combined. This degrades energy resolution and creates "fake" particles.
• Scalability: Handling high particle multiplicities and geometrical overlaps with rule-based heuristics is increasingly difficult.

2. Methodology: The HitPf Algorithm

The authors propose HitPf, a unified, end-to-end global event reconstruction framework. Instead of sequential clustering, HitPf maps raw detector inputs (calorimeter hits and charged tracks) directly to physics-level particle objects in a single trainable model.

The architecture consists of two distinct stages:

A. Stage 1: Hit Clustering (Candidate Determination)

This stage treats particle reconstruction as an instance segmentation problem on a point cloud of detector hits.

• Geometric Algebra Transformer (GATr): The network encodes raw hits (position $\vec{x}$, energy $E$, sub-detector type $s$) into a projective geometric algebra ($G_{3,0,1}$) representation. This allows the network to natively handle geometric operations (projections, angles, volumes) rather than learning them implicitly from scalar features.
• Object Condensation: The network is trained using an object condensation loss. It learns to cluster hits belonging to the same particle in a latent space by defining an "attractive potential" toward a condensation point (the particle center) and repelling points from other particles.
• Density Peak Clustering (DPC): At inference, particle candidates are extracted from the latent space. DPC identifies cluster centers as points with both high local density ($\rho$) and high distance ($\delta$) from higher-density points. This dual-criteria approach effectively suppresses "fake" clusters (which typically have either high density/low distance or low density/high distance).
• Consistency Check: A post-processing step removes spurious track assignments if the track momentum deviates significantly from the cluster energy.

B. Stage 2: Property Regression

Once candidates are formed, dedicated networks determine their physical properties:

• Input: The hits and associated tracks within a candidate cluster are processed by a smaller GATr network to create an embedding.
• Feature Concatenation: This embedding is combined with 11 aggregate features (e.g., energy fractions in ECAL/HCAL, track momentum, hit variances).
• Output: Two separate networks predict:
  1. Particle Identification (PID): Classification into one of five categories: charged hadron, photon, neutral hadron, electron, or muon.
  2. Energy Regression: Prediction of the particle's energy.
• Special Handling: For charged particles, direction and momentum are primarily derived from the track. For neutral particles, direction is computed from energy-weighted hit positions.

3. Key Contributions

• End-to-End Learning: HitPf bypasses intermediate clustering stages, learning a global event representation directly from low-level hits.
• Geometric Inductive Bias: The use of Geometric Algebra Transformers provides a strong inductive bias for spatial detector data, improving the handling of complex shower geometries.
• Detector Agnosticism: The framework decouples reconstruction performance from detector-specific tuning. Adapting to a new geometry requires only retraining (~48 hours on 4 H100 GPUs) rather than manual heuristic re-engineering.
• Superior Handling of Overlaps: The method effectively disentangles overlapping showers in dense environments, a critical capability for FCC-ee physics.

4. Results

The algorithm was benchmarked on fully simulated $e^+e^- \to q\bar{q}$ events at $\sqrt{s} = 91$ GeV using the CLD detector concept, comparing HitPf against the baseline PandoraPfa.

• Reconstruction Efficiency:
  • HitPf achieves 10–20% higher relative efficiency for charged hadrons (especially in the 1–10 GeV range).
  • Neutral hadron efficiency improves by up to 20% around 10 GeV.
  • Photon efficiency improves by ~5% at high energies and up to a factor of 2 at low energies ($E < 1$ GeV).
• Fake Particle Rates:
  • A reduction of up to two orders of magnitude in fake rates for charged hadrons.
  • Significant reductions in fake rates for photons and neutral hadrons across all energy ranges.
• Resolution:
  • Visible Energy and Invariant Mass Resolution: HitPf improves both by 22% compared to PandoraPfa.
  • Neutral Particles: HitPf significantly reduces energy overestimation caused by merging nearby showers, a common failure mode in rule-based algorithms.
• Confusion Matrices: HitPf shows fewer misidentifications (e.g., charged hadrons misidentified as neutral hadrons) and lower "missed" particle rates across all energy bins.

5. Significance

• Precision Physics: The 22% improvement in invariant mass resolution directly translates to a reduction in the integrated luminosity required to achieve a specific statistical precision for Higgs coupling measurements. This is vital for rare decays like $H \to c\bar{c}$ and $H \to s\bar{s}$.
• Accelerated Detector Design: By removing the need for detector-specific tuning, HitPf enables rapid iteration cycles. Physicists can evaluate different detector geometries and technologies with consistent, high-performance reconstruction, optimizing the detector design for physics goals rather than reconstruction constraints.
• Scalability: The approach demonstrates that machine learning can handle the high-multiplicity, overlapping shower environments of future colliders more effectively than traditional heuristic algorithms, paving the way for next-generation event reconstruction in both lepton and hadron colliders.