⚛️ high-energy experiments

Learning transferable event representations for charmed baryon physics at BESIII

This paper presents a Particle Transformer-based framework that leverages large-scale pre-training on Monte Carlo simulations to learn transferable event representations for charmed baryon physics at BESIII, demonstrating significant improvements in event classification and momentum-direction regression across various decay channels, particularly in low-statistics regimes.

Original authors: Kaixuan Huang, Yangu Li, Junpeng Zhao, Peilian Li, Peirong Li, Xiaorui Lyu, Yunxuan Song, Shengsen Sun, Yangheng Zheng

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kaixuan Huang, Yangu Li, Junpeng Zhao, Peilian Li, Peirong Li, Xiaorui Lyu, Yunxuan Song, Shengsen Sun, Yangheng Zheng

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

Imagine a massive, high-speed cosmic racetrack where tiny particles zoom around at nearly the speed of light, smashing into each other to create a shower of new, exotic fragments. This is the world of high-energy physics, a field dedicated to understanding the fundamental building blocks of our universe. To make sense of these chaotic collisions, scientists use giant detectors that act like super-fast cameras, capturing billions of snapshots of particle debris. But here's the catch: the data is overwhelming. It's like trying to find a specific, rare type of seashell on a beach that is constantly being buried by millions of ordinary rocks. Traditionally, scientists have built a unique, custom-made "shell-finder" for every single type of rare shell they want to study. This works, but it's slow, repetitive, and often fails when there aren't enough shells to learn from.

Enter the idea of "transfer learning," a concept borrowed from how humans learn. Just as a master chef who knows how to cook a perfect steak can quickly learn to grill a perfect burger without starting from zero, scientists are now teaching computer programs to learn general "cooking skills" for particle physics first. By training a model on a huge, diverse mix of data, it learns the universal patterns of how particles behave. Then, instead of building a new chef from scratch for every new dish, scientists can simply "fine-tune" that expert chef for the specific task. This paper explores whether this smart shortcut works for a specific type of particle called a "charmed baryon" in a famous experiment called BESIII.

The Paper's Story: Teaching a Computer to Spot Rare Particles

In this study, a team of researchers at the University of Chinese Academy of Sciences and other institutions asked a big question: Can we teach a computer to become a master of particle physics by letting it study a huge library of simulated events first, and then just tweaking it for specific jobs later? They focused on the BESIII experiment, a giant detector in Beijing that studies collisions of electrons and positrons (the antimatter twin of electrons). Specifically, they looked at "charmed baryons," which are heavy particles made of a charm quark and two lighter quarks. These particles are like the "rare seashells" of the physics world; they are fascinating but hard to spot because they are often buried under a mountain of boring background noise.

The Old Way vs. The New Way
Previously, if scientists wanted to study a specific decay of a charmed baryon (like a particle breaking apart in a specific way), they had to train a brand-new computer model from scratch. They would feed it thousands of examples of that specific decay and thousands of examples of background noise, teaching it to tell the difference. The problem? This is like hiring a new intern for every single task. It takes a lot of time, and if you don't have enough examples of the specific decay (a "low-statistics" situation), the intern never learns well enough to do the job.

The authors proposed a smarter approach using a framework called the "Particle Transformer." Think of this as a super-smart student who first spends years studying a massive encyclopedia of all possible particle interactions (pre-training). This student learns the general rules of the game: how particles move, how they cluster, and how energy is conserved. Once this student is an expert, the researchers can quickly "fine-tune" them for a specific job, like spotting a specific type of decay, by showing them just a few examples.

The Experiment: A Training Camp for AI
To test this, the team created a massive training camp using computer simulations (Monte Carlo samples). They didn't just look at one type of particle; they fed their model about 71 million simulated events covering three main categories: the charmed baryons they care about, other charm particles, and general background noise. The model learned to recognize the "shape" and "feel" of these events without being told exactly what to look for.

Once the model was pre-trained, they tested it on 12 different ways the charmed baryon could decay. They compared three strategies:

Direct Application: Using the pre-trained model immediately, without any extra training.
Fine-Tuning: Taking the pre-trained model and giving it a little extra training on the specific decay channel.
Training from Scratch: Ignoring the pre-trained model and building a new one from the ground up for each channel.

The Results: The Smart Shortcut Wins
The results were impressive, especially when data was scarce.

The Generalist: Even without any extra training, the pre-trained model was already quite good at spotting the charmed baryons. It could reject 97.0% of the background noise while keeping 90.0% of the real signals. That's like a security guard who can spot 97 out of 100 imposters without needing to memorize the face of every single criminal.
The Specialist: When the researchers "fine-tuned" the model for specific decay channels, it became even better. In 11 out of the 12 channels they tested, the fine-tuned model performed as well as or better than a model trained from scratch.
The Low-Data Hero: The biggest win came in the "low-statistics" scenarios—when there were very few examples to learn from. Here, the fine-tuned model crushed the "from scratch" models. For example, in a channel with only about 92,000 training events, the fine-tuned model was clearly superior, while the model trained from scratch struggled. This proves that the pre-training gave the model a head start, allowing it to learn effectively even with limited data.

Predicting the Invisible
The team also tested the model on a different task: predicting the direction a particle was moving, even when parts of it were missing. In some decays, a neutrino (a ghost-like particle that doesn't leave a trace) flies away, making it impossible to calculate the direction using standard methods. The pre-trained model, however, learned the patterns well enough to guess the direction of the remaining particles. When they fine-tuned this model for a specific decay involving a neutrino, it improved the prediction accuracy by about 28% compared to training from scratch. It's like being able to guess the direction a car was driving just by looking at the tire tracks, even if the car itself vanished.

What This Means
The paper concludes that this "pre-train then fine-tune" strategy is a powerful, scalable tool for the BESIII experiment and potentially for other high-energy physics labs around the world. It solves the problem of having to build a new model for every single experiment and makes it possible to get high-quality results even when there isn't a mountain of data available. While the results are based on simulations (which are highly reliable but not yet real-world data), they suggest that this method could revolutionize how physicists analyze data, making their searches for new physics faster and more efficient. The authors are essentially saying: "Don't reinvent the wheel for every new road; build a smart car that can learn any road quickly."

Technical Summary: Learning Transferable Event Representations for Charmed Baryon Physics at BESIII

Problem Statement
In the BESIII experiment, deep learning has proven effective for event classification and reconstruction, particularly using Particle Transformer architectures. However, current applications predominantly rely on a "training-from-scratch" paradigm, where dedicated models are developed for individual decay channels. This approach presents two primary limitations:

Redundancy: Significant development effort is duplicated across decay channels that share similar topologies and detector conditions.
Data Scarcity: Effective training of deep learning models typically requires tens of thousands of signal and background events. Many specific physics processes, particularly rare decay channels or those with limited statistics, lack sufficient data to train robust models from scratch, leading to suboptimal performance.

While transfer learning via pre-training has shown success in other high-energy physics (HEP) experiments (e.g., CMS and ATLAS) and other fields, a generalizable strategy tailored to the specific detector responses and event topologies of BESIII had not been established.

Methodology
The authors propose a transfer learning framework based on the Particle Transformer architecture, designed to learn transferable event representations through large-scale pre-training followed by fine-tuning.

Event Input Representation: To ensure generality, the input features are constructed at the "object level" rather than the hit level (too computationally expensive) or event level (too compressed). The input consists of a point cloud of charged tracks and electromagnetic calorimeter (EMC) clusters. Features include kinematic properties (momentum, angles), charge, and trajectory parameters, along with low-level measurements from the Time-of-Flight (TOF) and EMC systems for particle identification. These features are independent of specific physics channels.
Model Architecture: The framework utilizes a Particle Transformer backbone.
- Embedding: Particle features are projected into a latent space. Crucially, pairwise interaction features are constructed for every particle pair and encoded into an interaction matrix ( $U_{BES}$ ). This matrix incorporates the spherical geometry of $e^+e^-$ collisions and energy-momentum conservation, serving as a bias term in the attention mechanism to explicitly model kinematic correlations.
- Attention: The model employs particle attention blocks (multi-head self-attention) to capture global correlations, followed by class attention blocks that aggregate variable-length particle information into a fixed-dimensional event representation using a learnable class token.
Pre-Training Strategy: Two pre-training tasks are defined using large Monte Carlo (MC) simulation samples ( $\sqrt{s} = 4.600\text{--}4.700$ $s = 4.600 - 4.700$ GeV):
1. Event Classification: Distinguishing between three dominant physics categories: $\Lambda_c^+\bar{\Lambda}_c^-$ , inclusive $D\bar{D}^{(*)}h$ (charm production), and $q\bar{q}$ (hadronic background). The dataset contains $\sim 7.1 \times 10^7$ events.
2. Momentum-Direction Regression: Predicting the polar ( $\theta$ ) and azimuthal ( $\phi$ ) angles of the $\Lambda_c^+$ and $\bar{\Lambda}_c^-$ baryons using $\sim 2.4 \times 10^7$ $\Lambda_c^+\bar{\Lambda}_c^-$ events.
Fine-Tuning: For downstream tasks, the pre-trained weights are used to initialize the model. A differential learning-rate strategy is employed: a smaller learning rate preserves the backbone's learned representations, while a larger learning rate allows the output layer to adapt rapidly to specific signal definitions. This is tested across 12 benchmark $\Lambda_c^+$ hadronic decay channels and one semileptonic channel ( $\Lambda_c^+ \to pK^-e^+\nu_e$ ).

Key Contributions

Generalizable Framework: The first application of a pre-training/fine-tuning strategy specifically designed for BESIII, leveraging the shared topological structures of charmed baryon production (e.g., back-to-back production near threshold).
Scalable Solution: Demonstrates that a single pre-trained model can serve as a foundation for diverse decay channels, reducing the need for channel-specific model development from scratch.
Performance in Low-Statistics Regimes: Provides empirical evidence that transfer learning significantly outperforms training from scratch when training data is limited.

Results

Event Classification:
- The pre-trained model achieved an Area Under the Curve (AUC) of 0.988 for the $\Lambda_c^+\bar{\Lambda}_c^-$ category, rejecting 97.0% of background events at 90.0% signal efficiency.
- Across 12 benchmark channels, fine-tuned models consistently matched or exceeded the performance of models trained from scratch.
- Low-Statistics Advantage: The improvement was most pronounced in channels with limited training samples (e.g., $\Lambda_c^+ \to pK_S^0$ with $\sim 9.2 \times 10^4$ events), where the fine-tuned model achieved a background rejection rate of 96.4% compared to 90.5% for the from-scratch model. In high-statistics channels, performance was comparable.
Momentum-Direction Regression:
- The pre-trained model predicted baryon momentum directions with negligible bias ( $\mu \approx 0$ ) and resolutions of $\sigma(\Delta\theta) = 0.411$ rad and $\sigma(\Delta\phi) = 0.858$ rad on inclusive data.
- Across the 12 hadronic benchmark channels, the fine-tuned model improved angular resolution by approximately 31.1% ( $\Delta\theta$ ) and 22.8% ( $\Delta\phi$ ) compared to conventional single-tag reconstruction.
- Semileptonic Decay: For the challenging $\Lambda_c^+ \to pK^-e^+\nu_e$ channel (where the neutrino is undetected), the fine-tuned model reduced the resolution of $\Delta\theta$ and $\Delta\phi$ by 27.9% and 27.6%, respectively, compared to a model trained from scratch, despite having only $\sim 8.5 \times 10^4$ training events.

Significance and Claims
The paper claims that this strategy provides a scalable and extensible solution for deep learning applications at BESIII. By learning transferable event representations from large-scale inclusive samples, the framework effectively mitigates the data scarcity problem inherent in many specific physics analyses. The authors assert that this approach not only improves signal-background discrimination and reconstruction precision but also reduces development redundancy. Furthermore, they suggest this paradigm can be extended to future high-energy experiments, such as the Super Tau-Charm Facility (STCF) and the Circular Electron Positron Collider (CEPC), where similar challenges regarding data volume and topology diversity are expected.

Technical Summary: Learning Transferable Event Representations for Charmed Baryon Physics at BESIII

More like this