Finding Unexpected Non-Helical Tracks

This paper presents a model-agnostic tracking algorithm capable of reconstructing a broad class of smooth, non-helical particle trajectories without prior specification, thereby enabling the discovery of unexpected physics signatures that standard helix-based algorithms miss.

The Gist

The Big Idea: Looking for the "Unpredictable"

Imagine you are a detective trying to find a specific type of suspect in a crowded city. For decades, your police force has only been trained to look for people walking in perfect, predictable circles. They have a special algorithm that says, "If it's not walking in a circle, it's not a suspect; ignore it."

This works great for finding standard criminals (Standard Model particles). But what if a new, weird criminal type exists who walks in squiggly lines, spirals, or wobbly zig-zags? Your current algorithm would completely miss them because they don't fit the "circle" rule.

This paper introduces a new kind of detective (a Machine Learning algorithm) that doesn't care about the shape of the path. It just looks for smooth, continuous movement, no matter how weird the shape is. This allows scientists to find "new physics" that has been hiding in plain sight all along.

---

The Problem: The "Helix" Trap

In particle colliders like the Large Hadron Collider (LHC), particles smash together and fly out in all directions. Most of these particles are electrically charged, so when they fly through a magnetic field, they naturally twist into a helix (like a spring or a slinky).

• The Old Way: Traditional computers try to find these tracks by assuming every particle must be a spring. If a particle makes a weird path, the computer gets confused and throws the data away.
• The Risk: Theories suggest that some new, undiscovered particles (like "quirks" or magnetic monopoles) might not act like springs. They might wiggle, loop, or dance in ways we haven't predicted. Because our computers are obsessed with springs, we might be missing the biggest discoveries of the century.

The Solution: The "Smoothness" Detective

The authors built a new tool using Graph Neural Networks (GNNs). Think of this not as a rule-follower, but as a pattern-recognizer.

Instead of telling the computer, "Look for a spring," they taught it a simpler rule: "Look for a path that is smooth and connected."

How they trained it (The "Smoothie" Analogy)

To teach the computer what a "smooth path" looks like, they didn't use real physics equations. Instead, they used Fourier Series.

Imagine you are making a smoothie.

1. The Ingredients: You have a blender with many different frequencies (speeds) of blades.
2. The Rule: To make a smooth drink (a smooth track), you can't just dump in random chunks. You have to follow a rule where the "speed" of the blades decreases as you go higher. This is called a Schwartz function.
3. The Result: If you follow this rule, the resulting path is always smooth. It might look like a spring, a figure-eight, or a crazy squiggle, but it will never have sharp corners or sudden jumps.

The researchers generated thousands of these "smoothie paths" (some looking like springs, some looking like crazy scribbles) and fed them to the AI. They told the AI: "Learn what these paths look like. If you see a group of dots that form a smooth line, connect them."

The Magic: Generalization

The real test was: Can the AI find a path it has never seen before?

• The Test: They trained the AI on "Scribble Type A" and then tested it on "Scribble Type B" (a completely different shape).
• The Result: The AI didn't just memorize "Type A." It learned the concept of smoothness. It successfully found the new, weird shapes even though it had never seen that specific shape before.

This is like teaching a child to recognize "dogs" by showing them Golden Retrievers. If you then show them a Poodle, a Chihuahua, and a Great Dane, they still say, "That's a dog!" They learned the essence of a dog, not just the look of one specific dog.

The Results: Catching the "Quirks"

The team tested their new AI on a specific theoretical particle called a "Quirk."

• The Scenario: Quirks are particles that are stuck together by a new kind of force, making them bounce back and forth like a yo-yo as they move. They leave a very strange, non-springy trail.
• The Outcome: The old "spring-hunting" computers missed them. The new "smoothness-hunting" AI found them with high accuracy and very few false alarms.

Why This Matters

Currently, our data from particle colliders is full of these weird tracks, but we are blind to them because our software is too rigid.

This paper is a proof-of-concept. It shows that we can build a "model-agnostic" tracker—one that doesn't need to know the laws of physics in advance. It just needs to know what "smooth" looks like.

The Takeaway:
We might be sitting on a goldmine of new physics right now, buried in data we've already collected. We just need to stop looking for springs and start looking for smooth, continuous lines, no matter how strange they are. This new AI is the flashlight that finally lets us see them.

Technical Summary

1. Problem Statement

Particle colliders like the Large Hadron Collider (LHC) produce vast amounts of data, but standard track reconstruction algorithms are fundamentally limited by their reliance on simplifying assumptions.

• The Limitation: Traditional algorithms (e.g., Kalman filters) assume that charged particles follow helical trajectories in a uniform magnetic field. This assumption is valid for Standard Model (SM) particles but renders current detectors "blind" to particles predicted by Beyond the Standard Model (BSM) theories that exhibit non-helical trajectories (e.g., magnetic monopoles, quirks).
• The Gap: While specific algorithms exist for known non-helical trajectories, they cannot discover unanticipated physics. Current model-agnostic approaches are computationally intractable due to the combinatorial explosion of possible hit subsets.
• The Goal: Develop a model-agnostic tracking algorithm capable of reconstructing a broad class of smooth, non-helical tracks without requiring explicit parametric definitions of the particle's trajectory, thereby enabling the discovery of unexpected physics in existing datasets.

2. Methodology

The authors propose a Graph Neural Network (GNN)-based approach that decouples track finding from track fitting, allowing the network to learn trajectory patterns implicitly from training data.

A. Track Generation (The "Model-Agnostic" Space)

To train the network without bias toward specific BSM theories, the authors generated a synthetic dataset of "smooth" tracks using a Fourier basis approach:

• Smoothness Constraint: Tracks are defined as curves $\vec{x}(t)$ where all spatial derivatives exist and are finite. This is mathematically enforced by constraining the Fourier amplitudes $(a_n, b_n, c_n)$ to decay according to a Schwartz function ($f(n)$).
• Fourier Representation: A track is parameterized as $\vec{g}_N(t) = \sum c_{ni} \exp(\frac{2\pi i n t}{\lambda}) \hat{x}_i$.
• Schwartz Functions: The amplitudes are randomly sampled from a uniform distribution bounded by a Gaussian-like Schwartz function (e.g., $f(n) = A \cdot e^{-\alpha(n-c)^2}$). This ensures the generated tracks are smooth and physically plausible (continuous, non-kinked) while covering a vast space of shapes beyond simple helices.
• Dataset: The simulation uses a custom detector geometry (25 coaxial cylindrical layers). Training samples consist of SM background tracks (helical) plus at most one non-helical signal track.

B. The Tracking Pipeline (Exa.TrkX)

The authors utilize the Exa.TrkX pipeline, adapted for this specific task:

1. Graph Construction: Detector hits are nodes; potential connections between hits are edges.
2. Metric Learning: A neural network learns to map hits from physical space to a latent space where hits belonging to the same track are close together, and hits from different tracks are far apart.
3. Edge Classification: The network evaluates the likelihood of an edge connecting two hits belonging to the same track.
4. Graph Segmentation: Edges are pruned based on classification scores to form track candidates.

• Key Adaptation: Unlike traditional methods, the network does not assume a parametric form (like a helix). It learns the "shape" of the track solely from the training examples.

3. Key Contributions

• First Model-Agnostic Tracker for Smooth Tracks: Demonstrates a proof-of-principle that a GNN can learn to reconstruct non-helical tracks without explicit trajectory specifications, relying only on the implicit definition provided by the training sample.
• Generalization Capability: Shows that the network can generalize to track shapes not present in the training set, provided they share the underlying property of "smoothness" defined by the Schwartz constraints.
• Separation of Finding and Fitting: Validates the approach of using ML to find tracks based on hit correlations, leaving the physical parameter fitting (e.g., momentum, charge) to a subsequent step.

4. Results

The study evaluated performance using tracking efficiency (fraction of true tracks reconstructed) and fake rate (fraction of reconstructed tracks that are not real).

• Baseline Performance:
  • On SM (helical) tracks, the pipeline achieved 99.9% efficiency with near-zero fake rates.
  • Standard helical trackers failed to reconstruct non-helical tracks (efficiency < 3%).
• Non-Helical Reconstruction:
  • When trained specifically on non-helical tracks, the pipeline achieved >96% efficiency for tracks similar to the training set.
  • When trained on a mix of SM and non-helical tracks, efficiency remained strong (~50-60%) for non-helical tracks, with a very low fake rate (<1%).
• Generalization Tests (Disjoint Spaces):
  • The network was trained on tracks generated with one Schwartz function (e.g., Set 19) and tested on tracks generated with disjoint Schwartz functions (different parameters).
  • Results showed high efficiency (40–77%) on these disjoint sets. This proves the network learned the general concept of "smoothness" rather than memorizing specific Fourier coefficients or basis representations.
• Specific BSM Application (Quirks):
  • The pipeline was tested on "quirk" tracks (a specific BSM theory). It achieved a 32% overall reconstruction efficiency with 0 fake tracks, demonstrating its ability to recover physically motivated signals.
• Background Rejection:
  • Non-helical tracks can be distinguished from SM tracks by fitting them to a helix and calculating the $\chi^2$. Non-helical tracks typically yield large $\chi^2$ values, providing a powerful discriminant for discovery.

5. Significance and Future Work

• Discovery Potential: This work opens the door to searching for "striking" but unexpected particle trajectories in current LHC data that have previously been invisible to analysis.
• Paradigm Shift: It moves particle physics tracking from a "model-dependent" approach (assuming a trajectory) to a "model-agnostic" approach (learning trajectory patterns).
• Future Directions:
  • Incorporating realistic detector effects (hit noise, misalignment, pile-up).
  • Investigating hybrid approaches where SM tracks are fitted and removed before searching for model-agnostic tracks.
  • Exploring how additional hit information (timing, direction) could further enhance discovery power.

In conclusion, the paper successfully demonstrates that machine learning, specifically GNNs trained on mathematically defined smooth curves, can effectively reconstruct non-helical tracks, offering a powerful new tool for discovering physics beyond the Standard Model.