⚛️ high-energy experiments

ResoSeg: Resonance Tagger using Transformer and Segment Model

The paper introduces ResoSeg, a novel deep learning model that combines Transformer and segmentation techniques to perform joint particle-level segmentation and event-level classification for resonance tagging, demonstrating significantly improved efficiency and generalizability in reconstructing $\eta_c$ decays at the BESIII experiment compared to conventional methods.

Original authors: Chunkai Li, Junhao Yin, Ke Li, Jingde Chen

Published 2026-09-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chunkai Li, Junhao Yin, Ke Li, Jingde Chen

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 subatomic world, matter is built from a zoo of short-lived particles that appear and vanish in the blink of an eye. Physicists call these fleeting entities resonances. They are not stable building blocks like protons or electrons; instead, they are intermediate states that decay almost instantly into other particles. To understand the fundamental forces of nature, scientists must catch these particles in the act, identifying exactly which debris came from which parent. This task is like trying to reconstruct a specific type of shattered vase from a pile of mixed pottery shards found in a room full of other broken objects. For decades, researchers have relied on a method of sorting these shards one category at a time, checking for specific patterns that match known decay paths. While this approach has worked, it is slow and misses many opportunities because it cannot look at the whole picture at once.

A team of researchers at Nankai University and the Institute of High Energy Physics has now introduced a new way to solve this puzzle, one that changes how scientists analyze these collisions. They developed a computer model called ResoSeg, which acts like a highly skilled observer capable of looking at an entire event and instantly deciding two things: whether the event contains the particle they are hunting, and exactly which of the dozens of resulting particles came from that specific parent. Instead of checking for one specific pattern after another, the model learns to recognize the unique "fingerprint" of the parent particle across all its possible decay routes simultaneously. This allows them to reconstruct the full story of a particle's life in a single pass, capturing data that previous methods would have discarded or missed entirely.

The researchers tested this system using data from the BESIII experiment, a massive detector in China that records collisions between electrons and positrons. They focused on a particularly difficult particle called the eta-c, a type of charmonium that is known for having hundreds of different ways to decay, most of which are extremely rare. Because there is no single dominant way for this particle to break apart, traditional methods had to focus on only the sixteen most common paths, ignoring the rest. The new model, however, was trained to look at every track left by a particle in the detector and assign it a label: did it come from the eta-c, did it come from something else in the event, or was it just background noise? By doing this, the model could reconstruct the eta-c even when it decayed in ways the old methods were not designed to see.

The results were striking. When the researchers compared the new model against the standard approach across a range of energy levels, the new system was more than twice as efficient at finding the signal. In simpler terms, it found more than double the number of eta-c particles that the old method could identify, while keeping the background noise just as low. This improvement is not just about finding more numbers; it is about gaining a complete picture of the particle's properties, such as its mass and energy, with much greater precision. The model also proved to be flexible. It was able to work on energy points it had never seen before and could be adapted to study different production chains of the same particle with only a small amount of additional training.

To ensure the model was not just memorizing the data, the team tested its robustness by slightly altering the simulated properties of the particle, such as its mass or width. The model remained stable, continuing to perform accurately even when the conditions changed. This suggests that the system has learned the underlying physics of how these particles behave rather than just recognizing specific patterns in the training data. The researchers also demonstrated that the model could separate the signal from the background in a complex, mixed dataset, a crucial step for real-world physics analysis where the signal is often buried deep within a sea of other events.

This work represents a shift in how high-energy physics experiments are conducted. By moving from a channel-by-channel approach to a unified, all-at-once analysis, the researchers have opened the door to studying rare decay modes that were previously too difficult to isolate. The model is general enough to be applied to other particles and other experiments, particularly those at lepton colliders where background noise is low but the need for statistical precision is high. The source code for the model has been made available to the scientific community, inviting others to build upon this foundation. As physicists continue to search for exotic forms of matter and test the limits of the Standard Model, tools like this provide a clearer, more efficient lens through which to view the invisible world of subatomic resonances.

Technical Summary: ResoSeg – Resonance Tagger using Transformer and Segment Model

Problem Statement
In experimental high-energy physics, particularly at lepton colliders like BESIII, the identification and reconstruction of intermediate resonant states (e.g., $\eta_c$ , $J/\psi$ ) are critical for mapping the hadron spectrum and searching for exotic matter. While deep learning models such as Particle Transformer (ParT), SPANet, and GN3 have achieved state-of-the-art performance in event-level classification and object tagging, they possess a fundamental limitation: they do not provide per-track information regarding the origin of particles relative to intermediate resonances. Consequently, analyses relying on these models still require dedicated, decay-channel-specific reconstruction algorithms to determine the four-momentum of target resonances. This channel-by-channel approach is inefficient, especially for particles like the $\eta_c$ meson, which has hundreds of possible decay channels with tiny branching fractions and no single dominant mode, making traditional reconstruction impractical for "resonance-to-anything" decays.

Methodology
The authors propose ResoSeg (Resonance Segmenter), a unified deep learning model that jointly performs particle-level segmentation and event-level classification. The architecture integrates two primary components:

Encoder (ParT-inspired): The model utilizes a Particle Embedding layer and Pair Embedding layer to process input tracks and their interactions. It employs Particle Attention Blocks (modified from ParT with SwiGLU activation) to encode tracks into a high-dimensional latent space. Crucially, the interaction features are optimized via Bayesian optimization, selecting a subset of four features ( $M^2$ , $RM^2$ , $Q$ , $k_t$ ) that capture both local track-pair kinematics and global event-level center-of-mass context.
Decoder (SAM-inspired): Drawing from the Segment Anything Model (SAM), the decoder uses a Two-Way Attention mechanism. It takes the encoded particle features and interacts them with learnable "Event Tokens" and "Mask Tokens." The Mask Token generates semantic labels for each track, while the Event Token produces a score for the entire event. This allows the model to simultaneously determine if an event contains a signal process and assign a semantic category to every reconstructed track (e.g., fake/background, real tracks from the resonance, or real tracks not from the resonance).

The model is trained on a BESIII- $\eta_c$ dataset constructed for the process $e^+e^- \to \pi^+\pi^-h_c, h_c \to \gamma\eta_c, \eta_c \to \text{anything}$ . True labels for supervised learning are generated using a Truth-Matching Algorithm that minimizes the difference between reconstructed and true invariant masses. The total loss function is a linear combination of cross-entropy losses for event classification and particle segmentation.

Key Contributions

First Application of Segmentation: This work introduces the first application of semantic segmentation to resonance tagging in high-energy physics, enabling the direct identification of which tracks originate from a specific intermediate resonance.
Unified One-Pass Analysis: ResoSeg enables a single inference pass to reconstruct "resonance-to-anything" decay chains, eliminating the need for separate reconstruction algorithms for each decay channel.
Optimized Interaction Features: The authors demonstrate that specific interaction features, particularly those incorporating global kinematic context ( $RM^2$ ), significantly outperform standard jet-tagging features for quarkonium reconstruction.
Generalizability: The model is designed to be general, capable of adapting to other resonances and production modes via transfer learning without architectural changes.

Experimental Results
The model was evaluated on the BESIII- $\eta_c$ dataset and an inclusive Monte Carlo (MC) test set:

Performance vs. Conventional Methods: On energy points ranging from 4.19 to 4.60 GeV, ResoSeg achieved an average combined reconstruction efficiency more than 2.6 times higher than the conventional method, which is limited to 16 specific decay channels.
Event-Level Classification: ResoSeg achieved 97.31% accuracy and 98.84% recall in distinguishing signal from background events, comparable to a ParT baseline.
Particle-Level Segmentation: The model achieved a high Perfect Match Rate (PMR) of 87.91% and average Intersection over Union (IoU) scores of 96.12% (for tracks from $\eta_c$ ) and 95.20% (for tracks not from $\eta_c$ ).
Background Rejection: On the complex inclusive MC test set, ResoSeg demonstrated superior background rejection capabilities compared to ParT, maintaining high signal efficiency while significantly reducing background mistag rates.
Robustness and Generalization: The model remained robust against variations in the $\eta_c$ mass, width, and branching fractions. It also successfully generalized to unseen energy points (extrapolation) and adapted to a different production mode ( $\psi(2S) \to \gamma\eta_c(2S)$ ) via transfer learning with as few as 300,000 training events.

Significance and Claims
The authors claim that ResoSeg represents a paradigm shift in resonance analysis by transforming the workflow from channel-specific reconstruction to a general, resonance-aware approach. By providing per-track semantic labels, the model allows for the precise reconstruction of full four-momentum information for intermediate resonances without relying on decay-channel-specific algorithms. This capability substantially increases available signal statistics for "resonance-to-anything" decays, which is particularly vital for studying particles with complex decay topologies like the $\eta_c$ . The authors posit that this approach can be extended to other resonances (e.g., $J/\psi$ , $B$ mesons) and experiments (e.g., Belle II, LHCb), potentially enhancing the sensitivity of measurements for exotic hadron candidates and enabling new analysis strategies such as foundation models for particle physics.

Technical Summary: ResoSeg – Resonance Tagger using Transformer and Segment Model

More like this