⚛️ nuclear experiments

Track finding and particle identification performance with particle physics foundation models

This paper demonstrates that foundation models trained on multi-modal data significantly outperform traditional methods in reconstructing charged particles and identifying them within Time Projection Chamber datasets, while also showing adaptability to diverse downstream physics tasks.

Original authors: Shuhang Li, Haiwang Yu, Gabor Galgoczi, Yeonju Go, Yi Huang, David Park, Michael Peters, Christof Roland, Jin Huang, Yihui Ren, Joseph D. Osborn

Published 2026-10-23

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuhang Li, Haiwang Yu, Gabor Galgoczi, Yeonju Go, Yi Huang, David Park, Michael Peters, Christof Roland, Jin Huang, Yihui Ren, Joseph D. Osborn

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

In the heart of modern physics, scientists build massive machines to smash particles together at incredible speeds, hoping to reveal the fundamental building blocks of the universe. These collisions happen billions of times a second, creating a chaotic storm of new particles that fly out in every direction. To understand what happened, detectors surrounding the collision point must record the path of every single charged particle, like a giant, high-speed camera taking a picture of a hurricane. However, the sheer volume of data these machines produce is overwhelming. The detectors generate so much information that it must be drastically reduced before physicists can even begin to study the results. The challenge is to find the true paths of the particles hidden within this digital noise, a task that has become increasingly difficult as the machines become more powerful and the collisions more frequent.

To solve this, researchers are turning to a new kind of artificial intelligence known as a foundation model. Unlike older computer programs designed to do just one specific job, these models are trained on vast amounts of raw data to learn general patterns on their own. Once they have learned these patterns, they can be quickly adapted to solve different problems without needing to be retrained from scratch. A team of physicists recently tested whether this approach could improve the way they reconstruct particle tracks and identify what kind of particle they are looking at. They focused on data from the sPHENIX experiment at the Relativistic Heavy Ion Collider, using a detector called a Time Projection Chamber, which records the positions of particles as they drift through gas.

The researchers took a foundation model that had already been trained on nearly 200 million parameters using unlabeled data from the detector. They did not teach this model to identify specific particles or calculate energy loss; instead, it learned to understand the spatial relationships between the points where particles hit the detector. They then attached a small, specialized layer to this model to see if it could find the paths of particles and tell them apart. They compared this new method against the standard algorithms currently used in the field. The results showed that the foundation model was significantly better at finding the paths of slow-moving particles. In the magnetic field of the detector, slow particles curve sharply, making them hard to track, and the new model managed to find these difficult paths with much higher success than the traditional methods. It also avoided making mistakes, such as creating duplicate tracks or missing parts of a path, maintaining a high level of accuracy even in the most challenging conditions.

Beyond just finding the paths, the model also proved capable of identifying the type of particle, distinguishing between pions, kaons, and protons. This is a difficult task because these particles leave very similar traces in the detector, and their differences are often subtle. The foundation model achieved performance levels comparable to, and in some cases slightly better than, the established methods used by physicists for decades. Remarkably, it accomplished this without ever being explicitly taught about the energy loss that particles experience as they move through the gas. The model simply learned the patterns of the data so well that it could infer the identity of the particles on its own. This suggests that a single, powerful AI backbone could eventually handle multiple reconstruction tasks at once, simplifying the complex data analysis required for future experiments.

The study was conducted using simulated data that mimics the conditions of the sPHENIX detector, meaning the results represent what the model can do in a controlled digital environment. The researchers found that while the model excelled at finding tracks for slow particles, it performed just as well as traditional methods for faster particles. They also noted that the model's ability to identify particles was impressive given that its initial training focused only on finding patterns in space, not on measuring energy. This indicates that the model learned deep, transferable features about how particles behave. The work points toward a future where one adaptable AI system could replace many specialized tools, helping scientists extract more information from the vast amounts of data generated by the next generation of particle colliders.

Technical Summary: Track Finding and Particle Identification Performance with Particle Physics Foundation Models

Problem Statement
Modern high-energy nuclear and particle physics experiments, such as the sPHENIX experiment at the Relativistic Heavy Ion Collider (RHIC) and future facilities like the High-Luminosity Large Hadron Collider (HL-LHC), generate vast amounts of sparse data that require significant reduction before physics analysis can occur. Traditional reconstruction algorithms face increasing challenges as detector occupancies rise and luminosity increases. Specifically, conventional track-finding algorithms, such as those based on Cellular Automata (CA), struggle with low transverse momentum ( $p_T$ ) tracks. In these regimes, multiple scattering and high track curvature cause complex patterns, including "looper" tracks that traverse the detector multiple times, leading to reduced efficiency and purity. Furthermore, most existing machine learning (ML) approaches for reconstruction are optimized for specific tasks using labeled data, limiting their adaptability to new objectives without retraining from scratch.

Methodology
The authors evaluate the Foundation Model for Nuclear and Particle Physics (FM4NPP), a pre-trained model designed to learn transferable representations from low-level detector data via self-supervised learning.

Data Source: The study utilizes simulated data from the sPHENIX Time Projection Chamber (TPC), consisting of approximately 10 million proton-proton collisions at $\sqrt{s} = 200$ GeV. The input consists of TPC "spacepoints" (clusters of hits) rather than high-level reconstructed objects.
Model Architecture: FM4NPP is based on the Mamba sequence model. During pretraining, TPC spacepoints are voxelized in $(\eta, \phi, r)$ space and hierarchically serialized. The model was trained on nearly 200 million parameters using a self-supervised $k$ -Next-Nearest-Neighbor prediction task on unlabeled data.
Downstream Adaptation: To evaluate performance, a frozen pretrained backbone is paired with lightweight, task-specific supervised adapters:
- Track Finding: Uses transformer decoder layers to aggregate spacepoint embeddings, followed by two multi-layer perceptron (MLP) heads for classification.
- Particle Identification (PID): Uses a transformer decoder and a single MLP classifier to distinguish between pions, kaons, and protons.
Baselines:
- Track Finding: Compared against a CA-based seeder followed by Kalman Filter propagation.
- PID: Compared against a traditional Gaussian Process Regression (GPR) method fitting the mean ionization energy loss ($dE/dx$) as a function of momentum.
Evaluation Metrics: Track matching uses the "double-majority" method (requiring $\ge$ 50% overlap between truth and reconstructed spacepoints). PID performance is measured via Receiver Operating Characteristic (ROC) curves in the momentum range $p \in [0.8, 1.2)$ GeV/c, where species separation is most difficult.

Key Contributions and Results

Superior Low- $p_T$ Track Finding: The FM-based reconstruction significantly outperforms the traditional CA-based seeder at low transverse momentum ( $p_T \lesssim 1.0$ GeV/c). While CA algorithms suffer from efficiency losses due to multiple scattering and track looping, the FM maintains extremely high efficiency across the entire $p_T$ spectrum.
High Purity for Looping Tracks: The FM achieves a purity of nearly unity even at low $p_T$ ( $p_T \lesssim 0.5$ GeV/c). Unlike the CA seeder, which fails to join consecutive loops for particles looping within the magnetic field, the FM correctly identifies these tracks without duplicating them.
Effective Particle Identification without Explicit Training: Despite the pretraining objective focusing solely on spatial pattern recognition (and not energy loss), the FM-based PID adapter performs comparably to, and slightly better than, the traditional GPR $dE/dx$ method.
- Protons: Both methods achieve excellent separation (AUC $\ge$ 0.998).
- Pions: The FM slightly outperforms GPR (AUC 0.792 vs. 0.774).
- Kaons: The methods are comparable, with the FM showing a marginally higher Area Under the Curve (AUC 0.684 vs. 0.676) in this challenging region between pion and proton bands.
Transferability: The results demonstrate that a single frozen backbone can support multiple downstream tasks (tracking and PID) through lightweight adapters, even for tasks (like $dE/dx$ estimation) not explicitly present in the pretraining objective.

Significance
The paper claims that Foundation Models offer a powerful alternative to conventional reconstruction chains by learning generalizable representations from low-level detector data. The primary significance lies in the ability of a single architecture to handle the full $p_T$ spectrum with high efficiency and purity, solving the specific "looper" problem that plagues traditional algorithms at low momentum. Furthermore, the ability to perform particle identification with performance comparable to specialized $dE/dx$ methods—without explicit training on energy loss—suggests that FM4NPP can simplify data analysis pipelines. By using a shared pretrained backbone with task-specific adapters, experiments can potentially extend the reach of measurements, particularly in difficult low- $p_T$ regimes, while reducing the computational and engineering overhead of maintaining multiple specialized reconstruction algorithms.

More like this