Machine-learned particle flow as a foundation model for collider physics

This paper establishes machine-learned particle flow (MLPF) as a foundation model for collider physics by demonstrating that its learned latent representations serve as a shared, information-rich bridge between low-level detector data and diverse high-level analysis tasks, significantly improving performance and efficiency compared to traditional modular approaches.

The Gist

Imagine a massive, high-speed collision happening inside a particle accelerator. When particles smash together, they shatter into a chaotic spray of smaller fragments. To understand what happened, physicists need to rebuild the story from the debris.

Traditionally, this reconstruction process is like a factory assembly line with disconnected stations.

1. Station A looks at the raw, messy signals from the detectors and creates a basic list of "what particles are here."
2. Station B takes that list and tries to answer specific questions, like "Was this a heavy particle?" or "How much energy did it have?"

The problem is that once Station A finishes its job and hands over the list, it throws away all the subtle, messy details it saw in the raw data. Station B has to start from scratch, often having to manually invent new tools (called "features") to guess what it missed.

The Big Idea: The "Foundation Model"
This paper proposes a new way to run the factory. Instead of just handing over a simple list, the first station (a machine learning model called MLPF) keeps a "secret notebook" of high-level insights it learned while doing its job.

Think of this notebook as a universal translator or a rich internal memory. Even though the machine wasn't explicitly taught to answer the specific questions in Station B, its internal memory contains the raw physics of the event in a compressed, intelligent format.

The researchers took this "secret notebook" (called latent representations) and handed it to three different experts (the downstream tasks) to see if it helped them do their jobs better.

The Three Tests

The team tested this idea on three very different jobs:

1. Identifying the "Flavor" of a Jet (The Detective)

• The Job: Particles often clump together into "jets." Physicists need to know if a jet came from a heavy "beauty" quark, a "charm" quark, or a lighter particle. This is like a detective trying to identify a suspect's nationality based on their clothing.
• The Old Way: The detective only had a photo of the suspect's outfit (standard data).
• The New Way: The detective was given the photo plus the secret notebook from the first station.
• The Result: The detective became much better at spotting the heavy "beauty" quarks, even when they looked very similar to the others. The secret notebook contained clues about the suspect's history that the photo alone didn't show.

2. Measuring Jet Energy (The Accountant)

• The Job: Calculating exactly how much energy a jet carries.
• The Old Way: The accountant used standard math on the photo.
• The New Way: The accountant used the photo plus the secret notebook.
• The Result: The accountant's numbers were much more precise, especially for very high-energy jets. The notebook helped correct small errors that the standard math missed.

3. Finding "Missing" Momentum (The Balance Sheet)

• The Job: Sometimes particles (like neutrinos) escape the detector unseen. Physicists have to calculate where they went by seeing what is "missing" from the total balance.
• The Old Way: The balance sheet was often off because the individual numbers were slightly fuzzy.
• The New Way: The balance sheet was updated using the secret notebook, which understood the reliability of every single piece of data.
• The Result: This was the biggest win. The new method found the missing momentum with 35 times fewer parameters (a much simpler, lighter model) than the previous best method, and it was significantly more accurate.

The "Linear Probe" Surprise

The most surprising part of the paper is a test they called the "Linear Probe."

Imagine you have a super-complex, 2048-page secret notebook. Usually, you'd need a huge team of analysts to read it and find the answer. But the researchers asked: "Can a single, simple line of math read this notebook and still get a good answer?"

Yes.
Even with just a single, simple line of math (a linear layer), the model could extract useful physics information from the notebook.

• For the "Missing Momentum" test, this simple line of math beat the complex, industry-standard models.
• For the "Flavor" test, it did surprisingly well, even though the notebook was never explicitly trained to look for flavors. This proves the notebook naturally organizes the physics information in a way that is easy to read.

The Takeaway

The paper concludes that reconstruction and analysis don't need to be separate steps.

By using a machine learning model that learns a "shared language" (the latent representations) during the reconstruction phase, we can feed that language directly into analysis tasks. It's like if the factory worker didn't just hand you a box of parts, but also handed you a manual that explained exactly how those parts fit together, making the assembly process faster, cheaper, and more accurate.

This establishes the reconstruction model as a "Foundation Model" for particle physics: a powerful, pre-trained brain that can be easily adapted to solve many different problems without needing to be retrained from scratch.

Technical Summary

Technical Summary: Machine-learned particle flow as a foundation model for collider physics

Problem Statement
In traditional collider physics workflows, event reconstruction and high-level physics analysis are modular and disconnected processes. Standard particle-flow (PF) algorithms translate raw detector signals into a list of stable particle candidates (PF candidates), which then serve as the interface for downstream analysis. However, once this list is produced, the rich low-level correlations encoded in the raw detector signals are lost. Recovering task-relevant information beyond the four-momenta of PF candidates typically requires hand-engineering additional features (e.g., track displacement variables for jet flavor identification). This paper addresses the lack of a shared representation linking low-level detector data to high-level analysis tasks, proposing that casting event reconstruction as a machine learning problem can naturally produce such a representation.

Methodology
The authors utilize a Machine-Learned Particle Flow (MLPF) model, originally designed as a graph neural network and later evolved into a transformer-based architecture, as a "backbone" for event reconstruction. The core methodology involves:

1. Latent Representation Extraction: During standard reconstruction inference, the MLPF model generates high-dimensional (2048-dimensional) per-particle latent representations. These are learned end-to-end to encode detector response and particle interactions, capturing structural information often discarded by conventional algorithms.
2. Unsupervised Compression: To make these representations computationally practical for downstream tasks, the authors apply Principal Component Analysis (PCA) to compress the 2048-dimensional vectors into 128 dimensions. This compression is performed in an entirely unsupervised manner using a dedicated set of events, ensuring no task-specific information leaks into the compression step.
3. Downstream Evaluation: The compressed latent vectors are appended as additional input features to standard kinematic inputs (four-momentum, particle identification) for three distinct downstream tasks. The authors compare three model variants for each task:
   • Baseline: Standard task-specific architecture using only kinematic features (and hand-engineered features where applicable).
   • Latent-augmented: The same architecture as the Baseline, augmented with the 128-dimensional MLPF latent vectors.
   • Linear-probe: A single linear layer trained only on the latent representations to quantify how much task-relevant information is linearly accessible without further non-linear processing.
4. Experimental Setup: The study uses simulated $e^+e^- \to t\bar{t}$ events at 365 GeV from a CLD-like detector (proposed for FCC-ee). The MLPF backbone weights are kept completely frozen, and all downstream experiments use events from the held-out test split of the MLPF fine-tuning procedure to prevent data contamination.

Key Contributions and Results
The paper demonstrates that MLPF latent representations encode essential physics information useful for diverse downstream tasks, establishing MLPF as a foundation model. The results across three distinct tasks are:

• Jet Flavor Identification (Multi-class Classification):

  • The Latent-augmented model (ParticleNet + latents) significantly outperforms the Baseline. At a 1% mis-identification rate, it improves $b$-jet identification efficiency by ~3% against light-flavor jets and ~6% against $c$ jets.
  • The Linear-probe model (387 parameters) achieves an AUC of ~0.922 for $b$-vs-$c$ discrimination, despite the MLPF backbone never being trained on jet-flavor labels. This indicates that flavor-discriminating structure is intrinsically encoded in the latent space.
  • The Latent-augmented model trained on only 100k jets achieves performance comparable to a Baseline model trained on the full 1.83M jet dataset.

• Jet Energy Regression:

  • The Latent-augmented model improves jet energy resolution by approximately 10–15% across the jet $p_T$ range compared to the Baseline.
  • The Linear-probe model trails the Baseline by ~3% in resolution, suggesting that while the latent space contains significant information, the Baseline's ability to learn non-linear aggregations of kinematic features provides an advantage for this specific task.

• Missing Momentum ($\vec{p}_{miss}$) Regression:

  • This task showed the most dramatic improvement. The Latent-augmented model (DeepMET + latents) reduced the validation loss by 26% compared to the Baseline.
  • Crucially, the Linear-probe model (129 parameters) outperformed the DeepMET-based Baseline at every training-set size while using approximately 35 times fewer parameters.
  • The Latent-augmented model improved recoil resolution by 15–20% and longitudinal resolution by ~10% across the full range.

Significance and Claims
The paper claims that these results establish MLPF as a foundation model for collider physics. The significance lies in two dimensions of transferability demonstrated in this work and a companion study [19]:

1. Cross-Detector Transfer: MLPF representations can be fine-tuned to new detector geometries with substantially less data than training from scratch.
2. Cross-Task Transfer: The latent representations learned during reconstruction are generically useful for downstream analysis tasks (classification, regression) without requiring retraining of the backbone or explicit design of a foundation model.

The authors argue that this approach offers a concrete step toward an end-to-end pipeline from detector data to physics analysis. By providing a shared representation that encodes low-level correlations, reconstruction models can reduce the need for hand-engineered features and allow for more efficient training of downstream analysis models. The paper concludes that reconstruction and analysis need not be treated as separate pipeline stages, as the reconstruction model itself serves as a natural foundation for physics analysis.