ResoSeg: Resonance Tagger using Transformer and Segment Model

I. INTRODUCTION

• Existing event classifiers and object taggers lack per-track resonance-origin information, so analyses still require decay-channel-specific algorithms to reconstruct resonance four-momenta.This leaves the overall analysis paradigm unchanged despite strong model performance in jet tagging and event classification.Source: §I · p. 1
• Track-level segmentation identifies resonance decay products and supports one-pass reconstruction of resonance decay chains and their full kinematics without channel-specific algorithms.ResoSeg combines event-level classification with particle-level segmentation, assigning semantic categories to reconstructed tracks.Source: §I · pp. 1–2
• Resonance reconstruction precision directly determines measurement sensitivity across hadron-spectroscopy and exotic-candidate studies at BESIII, Belle II, and LHCb.The introduction names ηc and J/ψ at BESIII, B and D mesons at LHCb, and B mesons at Belle II as examples.Source: §I · p. 1
• ηc reconstruction is challenging because it has hundreds of possible decay channels, most with tiny branching fractions and no single dominant mode.Traditional methods focus on 16 relatively large-branching-fraction channels with cleaner backgrounds.Source: §I · p. 2

II. BESIII DETECTOR

• BESIII has observed more than 30 new particles, and several final states involving ηc contribute to exotic-state studies.The detector records symmetric e+e− collisions and its cylindrical core covers 93% of the full solid angle.Source: §II · p. 2

III. BESIII-ηc DATASET

• The dataset defines signal as e+e−→π+π−hc, hc→γηc, ηc→anything, and background as hc decays not containing γηc.The signal and background share the e+e−→π+π−hc production process.Source: §III · p. 2
• Each event has a signal label, while reconstructed tracks are labeled as fake or background, ηc-origin, or not ηc-origin.The latter category includes the production π+π− and the photon from hc decay.Source: §III · p. 2
• The matched ηc invariant-mass spectrum agrees with conventional reconstruction, supporting the Truth-Matching Algorithm’s labels.The algorithm selects the matching result with minimum average invariant-mass resolution; its average efficiency is 5.68% at 4.22, 4.26, and 4.39 GeV.Source: §III · p. 2, Figure 1 · p. 3
• Track inputs include MDC trajectory information, EMC shower properties, and particle-identification information for charged and neutral tracks.The listed inputs include helix parameters, shower position and shape, deposited energy, and PID classifications.Source: §III · p. 3
• The simulated dataset contains 4.8 million events at 4.22, 4.26, and 4.39 GeV, split into training, validation, and test sets in an 8:1:1 ratio.At each energy point, 0.8 million events are generated for each of the signal and background channels.Source: §III · p. 3
• A preselected BESIII inclusive-MC sample at 4.26 GeV provides a test set spanning known hadronic e+e− decay processes, with 222,125 events remaining.Selections require candidate recoil-mass windows and events that pass a kinematic fit to themselves.Source: §III · p. 3
• Classification uses accuracy and recall; segmentation uses category-average IoU and perfect match rate, while reconstruction efficiency and background rejection assess combined performance.The relative signal efficiency is the number of correctly reconstructed signal events divided by the number passing truth matching; inclusive-MC rejection is the inverse background mistag efficiency.Source: §III · p. 3

IV. MODEL ARCHITECTURE

• The model uses permutation-invariant track inputs without positional encoding, and its Particle Attention Blocks replace GELU with SwiGLU.The total loss linearly combines cross-entropy losses for event classification and particle segmentation, with equal fixed weights in the baseline implementation.Source: §IV · pp. 3–5
• Pairwise interaction inputs include track-pair features, with RM² incorporating center-of-mass energy and momentum as global kinematic context.The model embeds the track and interaction inputs before feeding them into Particle Attention Blocks; the interaction features are optimized by Bayesian optimization.Source: §IV · p. 4
• ResoSeg uses a ParT-inspired encoder and SAM-inspired decoder to produce a score for the event and a category score for each track.The decoder takes the encoded particle representations as its sole input and performs multi-class rather than binary classification.Source: §IV · pp. 3–4
• The Mask Token encodes track-level mask information and interacts with particle features to generate the final track scores.A point-wise product between the Mask Token and particle representations produces the track-score output.Source: §IV · pp. 4–5

V. EXPERIMENTS

• Experiments use the BESIII-ηc dataset designed for the process described in Sec. III.The model trained on this dataset serves as a pre-trained model.Source: §V · p. 5
• Transfer learning changes the target decay channel to ψ(2S) →γηc(2S), ηc(2S) → γhc, hc →γηc to evaluate generalization to other ηc decay channels.Source: §V · p. 5
• Robustness tests vary the ηc mass, width, and branching fractions.Source: §V · p. 5

1. Setup

• The full BESIII-ηc dataset is used for training, and the checkpoint with the lowest validation loss is evaluated on the test set.Training uses Adam for 100 epochs with batch size 1024, linear warmup, cosine annealing, and a stable learning rate of 0.0001.Source: §1 · p. 5

2. Post-processing

• Post-processing sums four-momenta of tracks identified as originating from ηc and requires at least one neutral and two charged tracks for the other-track candidate search.For those other tracks, the procedure searches for the best γπ+π− candidate.Source: §2 · p. 5

• ResoSeg reproduces the true-label ηc and γηc mass-spectrum shapes and shows strong event-classification and particle-segmentation metrics.Source: §3 · pp. 5–6, Figure 3 · p. 6
• ResoSeg and ParT perform comparably at event level, but on inclusive MC ResoSeg achieves higher Rej at competitive signal efficiency and recovers full ηc and hc four-momenta.Source: Figure 4 · p. 6, §3 · p. 6

Signal extraction on inclusive MC test set

• A 10,000-repeat input/output check yields a pull mean of 0.03 ± 0.01 and σ of 1.12 ± 0.01, indicating stable, accurate signal extraction.Source: §3 · p. 6, Figure 5 · p. 7, Figure 6 · p. 7

Efficiency and generalization to other energy points

• 2.6 times the conventional value: ResoSeg's average combined efficiency; its εrel,sig remains high at unseen energies but degrades as energy increases.The comparison is made with comparable background levels in the inclusive MC samples.Source: §3 · pp. 6–7, Figure 7 · p. 7

B. Transfer learning

• The transfer dataset uses ψ(2S)→γηc(2S), ηc(2S)→γhc, hc→γηc, ηc→anything at 3.686 GeV, with an 8:1:1 train/validation/test split.Source: B. Transfer learning · p. 8
• 300,000 events already yield good performance, and the transfer-learning metrics saturate at 800,000 events.Source: B. Transfer learning · p. 8
• Two low-energy photons among the three non-ηc tracks are difficult to identify, weakening non-ηc track discrimination and slightly degrading PMR and εrel,sig.Source: B. Transfer learning · p. 8

C. Robustness Evaluation

• ResoSeg is reported robust to the tested ηc mass and width shifts, measured by relative change in signal efficiency.The tested shifts range from ±5σ to ±10σ for mass and ±2.5σ to ±5σ for width.Source: C. Robustness Evaluation · p. 8
• 1.98% and 4.92% are the average ∆ϵrel, sig changes for BESIII-default and uniform branching fractions, respectively.The authors interpret these changes as evidence of robustness.Source: C. Robustness Evaluation · p. 9

VI. BAYESIAN OPTIMIZATION OF THE INTERACTION FEATURES

• The Efficient Global Optimization procedure uses a Gaussian-process surrogate and Expected Improvement to select feature subsets while balancing exploration and exploitation.The search limits costly model retraining evaluations.Source: §VI · p. 9
• The optimal interaction-feature subset, [M 2, RM 2, Q, kt], significantly improves ϵrel, sig and Rej90 for quarkonium with ResoSeg, and Rej90 with ParT, versus the default ParT interaction subset.Source: §VI · p. 9
• M 2 and RM 2 produce the largest individual improvements in ϵrel, sig; the other interaction features contribute marginally by comparison.Importance is defined as the average change in ϵrel, sig when a feature is included versus excluded.Source: §VI · p. 9, Figure 10 · p. 9

VII. GATED ATTENTION AND CONDITIONAL LAYER NORMALIZATION

• Without RM 2, CLN improves PMR by 6.61%, ϵrel, sig by 5.87%, and Rej90 by 44.68% over the noRM model.With RM 2 included, the authors say this improvement disappears, likely because RM 2 already carries event-level information similar to CLN's condition.Source: §VII · p. 9
• The final model omits gated attention and CLN because the tested models show no significant performance difference.The tested modifications replace attention with gated attention, layer normalization with CLN, or apply both changes.Source: §VII · p. 9

VIII. CONCLUSION AND DISCUSSION

• ResoSeg jointly performs particle-level segmentation and event-level classification for unified, single-pass reconstruction of intermediate resonances.The demonstration uses a dedicated BESIII-ηc dataset for e+e−→π+π−hc, hc →γηc, ηc →anything.Source: §VIII · p. 10
• At comparable background levels, ResoSeg’s average combined efficiency is 2.6 times that of the conventional method, while its event-level performance is comparable to ParT.ResoSeg also gives stronger background rejection on the more complex inclusive MC sample.Source: §VIII · p. 10
• The model remains effective at extrapolated energy points from 4.19 to 4.60 GeV, transfers to an alternative ηc production chain, and is stable under tested mass, width, and fraction shifts.These are reported generalization and robustness results for the tested settings.Source: §VIII · p. 10
• Future work includes addressing imbalance among ηc decay modes, optimizing the architecture, and developing a systematic treatment of resonance interference.The approach is presented as transferable beyond ηc and BESIII, especially to other lepton-collider experiments.Source: §VIII · p. 10