Source-linked AI summary
"Transforming" LHCb: self-supervised maps of heavy-flavour decays
Marko Stamenkovic, Greg Landsberg
TL;DR
The paper addresses limited heavy-flavour tagging information and incomplete-decay reconstruction in LHCb. It introduces a self-supervised transformer that learns decay maps without flavour or exclusive-decay labels, outperforming random-weight models and approaching supervised performance across five tasks, while showing systematic responses in simulated and 2017 collision data.
Problem
Current LHCb flavour tagging underuses full decay-environment information and relies on less powerful architectures, limiting sensitivity to incomplete heavy-flavour decays.
Method
A self-supervised transformer learns heavy-flavour decay maps by masking particle-identification information and completing jets after constituents are removed, without flavour or exclusive-decay labels.
Results
Across five classification tasks, the learned representation recovers 97–100% of fully supervised performance, while removal tests show systematic missingness responses in simulation and 2017 collision data.
Takeaways & Limitations
The study provides a proof of principle for learning heavy-flavour hadronization and decay maps that could support searches for incomplete and unusual decays.
Takeaways & Limitations
The anomaly study is limited by reduced jet records and requires larger models, more collision data, stability tests, and complete event information to establish the population’s origin.
Abstract
from arXiv · showhide
Decays of beauty and charm hadrons provide sensitive probes of physics beyond the standard model, including decays with invisible particles, in which part of the final state leaves no reconstructed detector signature. The large heavy-flavour data samples recorded by the LHCb experiment at the CERN LHC, together with its precise tracking, displaced vertex reconstruction, and particle identification, make it particularly well suited to learning a map of reconstructed heavy-hadron decay environments directly from data. We propose to bring recent advances in jet flavour tagging at ATLAS and CMS to significantly improve on the performance of the current LHCb taggers and extend them to the reconstruction of heavy-flavour decays with several invisible particles in the final state. To achieve this, we introduce a self-supervised transformer architecture that learns the decay maps without flavour or exclusive-decay labels by inferring masked particle identification information and completing jets from which constituents have been removed. Across five classification tasks in simulated LHCb Open Data, the self-supervised model outperforms an otherwise identical transformer with random weights, and performs comparably to a fully supervised transformer. We achieve a tagging power of about 10\%. In addition, removing constituents from reconstructed exclusive decays also systematically increases the model anomaly score relative to random removals from the same heavy hadrons. We confirm this behaviour directly in 2017 LHCb proton-proton collision Open Data: the score increases for all eight studied heavy-flavour channels, and the signal region response exceeds that in the adjacent sidebands. These studies provide a proof of principle that mapping heavy-flavour decay environments through jets can transform flavour tagging in LHCb and extend the discovery reach for incomplete or otherwise unusual decays.
1 Introduction
Heavy-flavour decays offer probes of physics beyond the Standard Model, but current LHCb tagging underuses full decay environments. The paper introduces self-supervised transformer mapping of reconstructed heavy-flavour decays and validates it in simulation and collision data.
- Heavy-flavour decays probe new flavour and CP violation and partially reconstructed final states involving invisible particles.
- Current LHCb taggers focus on secondary vertices and underuse fragmentation, underlying-event, and pileup information.
- The proposed approach maps heavy-flavour decay environments with self-supervised particle-set completion using reconstructed visible constituents.
- Masked PID inputs and completely removed constituents are reconstructed to learn a continuous representation of hadronization and decay environments.
- The study combines controlled simulated tests with label-free training on 2017 collision data and validation on eight reconstructed charm and beauty modes.
2 Simulated LHCb Open Data sample
The study uses simulated 13 TeV LHCb dijet events reconstructed with detector, tracking, vertex, and PID information. Event-level partitions and reserved analysis data support independent training, testing, and intervention studies.
- The simulated sample contains 13 TeV proton-proton dijets reclustered as anti-kT jets with R = 0.5.
- The sample includes inclusive b¯b, c¯c, and light-parton dijets plus Z →b¯b events across jet-momentum intervals and magnet polarities.
- Each event contains two reconstructed secondary-vertex jets with kinematic, displacement, and probabilistic RICH PID information.
- Jets contain at most 64 unordered reconstructed constituents, with RICH and calorimeter PID responses used as input features while reconstructed PID classes provide targets.
- Event-level splitting assigns disjoint events to training, validation, test, and analysis sets, with analysis data reserved for reconstructed-decay interventions.
3 Self-supervised mapping of hadronization and decay environments
A self-supervised transformer learns heavy-flavour decay representations by reconstructing masked PID information and constituents removed from artificial jet views. Its missingness score compares inferred incompleteness against the training baseline.
- The transformer learns heavy-flavour representations by inferring masked PID information and reconstructing particles removed from reconstructed decay environments.
- Training pairs complete jets with views containing zero to four removed constituents, using multiple artificial removal mechanisms.
- Removed constituents are deleted without placeholders, and jet quantities are rebuilt from the surviving particles to avoid positional and reconstruction biases.
- The encoder forms a global completeness representation from surviving particles, while decoder heads predict removal counts, mechanisms, set properties, particle properties, and masked PID.
- The missingness score compares the model’s inferred incompleteness with the unchanged-jet baseline π0 = 0.25.
4 Demonstration in simulation
In simulation, the self-supervised decay map detects incomplete heavy-flavour environments, retains flavour and charge information, and approaches fully supervised tagging performance. Controlled removals from reconstructed decay candidates also produce topology-specific anomaly responses, supporting validation across several decay modes.
- Controlled constituent deletion: 0.72 inclusive AUC separates paired complete and incomplete jet views, with response increasing as more constituents are removed.The AUC rises from 0.65 at P = 1 to 0.87 at P = 4.
- Controlled constituent deletion: 0.82 AUC is obtained for both displaced-vertex tracks and cascade-decay removals, exceeding the 0.71 random-charged-particle baseline.These values use realized mixtures of removal mechanisms and multiplicities, so they do not isolate topology at fixed multiplicity.
- Flavour and quark charge sign tagging with limited labels: The pretrained representation outperforms random features for every flavour or charge task and every labelled-data fraction.For b versus c, the AUC is 0.83 with 0.1% of labelled events versus 0.67 for random features.
- Flavour and quark charge sign tagging with limited labels: 97–100% of compact fully supervised performance is recovered by affine probes across the five tasks when all labelled events are available.The nominally better pretrained-probe performance on light-jet tasks should not be overinterpreted because the simulated light-jet sample is small.
- Dilution and tagging power: 10.6% tagging power is obtained for the simulated beauty sample, compared with 31.1% for the simulated charm sample.The corresponding effective dilutions are 32.6% and 55.8%, respectively.
- Dilution and tagging power: The quoted 100% tagging efficiency is conditional on an accepted opposite-side b-jet, and charged heavy-hadron samples make quark-charge identification easier than for neutral B mesons.A direct candidate-level comparison with LHCb would require selection-efficiency and mistag calibration in data.
- Removing reconstructed decay candidates: Removing reconstructed candidates produces larger anomaly-score shifts than matched control removals, including a 4.05 versus 1.33 shift for the Kπ pair.The excess is positive for 88.5% of matched pairs, indicating that the response is not explained by charge and kinematic matching alone.
- Removing reconstructed decay candidates: Candidate-specific excess responses are observed across D0, D+, and other reconstructed topologies, but the D∗ and ϕ controls limit or do not support a universal resonance-specific interpretation.The study demonstrates controlled artificial anomalies across several topologies while retaining explicit interpretation limits.
5 Demonstration in LHCb Open Data proton-proton collision samples
The method is trained and tested on inclusive 2017 LHCb collision data using reconstructed charm and beauty decays. Removing decay candidates raises the missingness response, while simulation-trained probes transfer useful flavour information to collision data.
- 1.03 fb^-1 of 2017 collision data enabled training and validation entirely in collision data.
- Eight reconstructed charm and beauty control modes span different particle multiplicities and decay chains.
- 3.99, 3.76, 3.78 and 4.99 were the mean signal-window scores for the four high-statistics charm modes, each exceeding both adjacent-sideband results.
- 54.5% for B+ →J/ψK+, 38.1% for prompt-decay-enriched D0 →Kπ and 81.6% for prompt-enriched D+ →Kππ were the estimated tagging powers.
- Simulation-trained affine probes separated charged beauty and charm flavour categories in collision events that were not used for probe training.
- The exploratory high-score sample contained a charged-rich, neutral-poor group, but the authors require larger data, stability tests and the complete event record before interpreting its origin.
6 Discussion
The study establishes a proof of principle for self-supervised heavy-flavour mapping while identifying detector, reconstruction and event-representation boundaries for future analyses.
- 97–100% of fully supervised performance was recovered across five flavour and quark charge sign classification tasks.
- LHCb’s precise vertexing, particle identification and abundant heavy-flavour samples support self-supervised mapping of decay environments.
- Anomalous populations may arise from detector effects, reconstruction failures or acceptance boundaries, requiring dedicated detector-condition checks.
- The present study treats jets independently, while future representations could use larger-radius jets or complete events centered on decay candidates or the primary vertex.
- Cosine-similarity retrieval can organize high-scoring events, but determining whether they reflect known decays or detector effects requires larger models, more data and calibration.
7 Conclusions
The self-supervised transformer learns a heavy-flavour decay map without flavour or exclusive-decay labels and transfers its learned responses to collision-data decays. The results support a proof of principle for mapping incomplete or unusual heavy-flavour decays.
- 97–100% of fully supervised performance was recovered across five flavour and quark charge sign tasks in simulation.
- Simulation-trained encoders and affine probes separated B+ and charm flavour categories in 2017 collision data excluded from training.
- The independently collision-trained method responded systematically to eight reconstructed charm and beauty decays, with high-statistics charm signals exceeding both adjacent sidebands.
- A map extended to complete events and calibrated across magnet polarities could support searches for incomplete and unusual decays.
A Training diagnostics
Training diagnostics show separate completion objectives for simulation and collision data, while PID prediction is evaluated with balanced class accuracy. The two datasets are explicitly not intended as a direct completion benchmark.
- Epochs 60 and 97 were selected checkpoints, with training and validation objectives of 2.90 and 2.90 in simulation and 5.75 and 5.73 in collision data.
- The simulated and collision-data samples contain different reconstructed particle mixtures and are not intended as a direct benchmark against one another.
- 0.63 in simulation and 0.49 in collision data were the equal class accuracies for masked visible-particle PID prediction.
B Reconstructed candidate definitions
Reconstructed candidates are defined independently of the missingness score using charge, particle identification, displacement, and track geometry. Frozen mass windows establish signal regions and adjacent sidebands for simulation removal tests.
- Candidates use charge, PID, displacement, and track geometry, independently of the missingness score.
- One candidate per jet is selected without using the model response, with all charge-conjugate modes included.
- The primary D0 study in b-jets uses |m(Kπ)−mD0| < 20 MeV for its signal window and 40 < |m(Kπ)−mD0| < 80 MeV for its sideband.
- Beauty-enriched categories use the sideband of the reconstructed charm decay, with frozen signal and sideband definitions recorded in Table 8.