Abstract
The task of identifying B meson flavor at the primary interaction point in the LHCb detector is crucial for measurements of mixing and time-dependent CP violation. Flavour tagging is usually done with a small number of expert systems that find important tracks to infer the B meson flavour from. Recent advances show that replacing all of those expert systems with one ML algorithm that considers all tracks in an event yields an increase in tagging power. However, training the current classifier takes a long time and is not suitable for use in real-time triggers. In this work we present a new classifier, based on the DeepSet architecture. With the right inductive bias of permutation invariance, we achieve great speedups in training (multiple hours vs 10 minutes), a factor of 4-5 speed-up in inference for use in real time environments like the trigger and less tagging asymmetry. For the first time we investigate and compare performances of these Inclusive Flavor Taggers on simulation of the upgraded LHCb detector for the third run of the LHC.
Introduction
In particle physics, specifically at the LHCb experiment, scientists study $B$ mesons to understand CP violation—the subtle difference between matter and antimatter. A critical step in these studies is "flavour tagging": determining whether a $B$ meson was a $B^0$ or a $\bar{B}^0$ at the moment of its creation.
Traditionally, this has been achieved using "classical taggers"—a suite of expert algorithms designed to identify specific signatures, such as a particular kaon or muon produced alongside the $B$ meson. However, these systems suffer from low efficiency; if the specific tagging particle is not detected, the algorithm cannot provide a decision. Furthermore, maintaining and calibrating numerous independent expert systems is both computationally and labor-intensive.
The DeepSet Architecture for Inclusive Tagging
To overcome these limitations, researchers have developed an "inclusive" tagger that analyzes all particles in an event simultaneously. The core of this approach is the DeepSet Neural Network architecture.
The DeepSet architecture is uniquely suited for particle physics data because it treats the tracks from charged particles as a set rather than a sequence. Key features include:
- Permutation Invariance: The model's output remains unchanged regardless of the order in which tracks are processed. This is achieved by passing each track $x_i$ through a network $\phi$, summing the outputs, and then passing the aggregate through a second network $\rho$.
- Variable Input Length: The model naturally handles events with varying numbers of tracks.
- Efficiency: By baking permutation invariance into the architecture as an inductive bias, the model learns significantly faster than generic neural networks, reducing training time from days to under an hour.
Experimental Setup and Datasets
The performance of the DeepSet tagger was evaluated using simulated data for two distinct LHC eras:
- Run 2 (2015-2018): Low "pileup" conditions (approx. 1 interaction per collision).
- Run 3 (2022-2025): High "pileup" conditions (approx. 6 interactions per collision), presenting a noisier environment for the upgraded LHCb detector.
The models were tested across several $B$ meson decay channels, including $B^0 \to J/\psi K^{*0}$, $B^+ \to J/\psi K^+$, and $B_s^0 \to D_s^+ \pi^-$.
Results and Performance
The DeepSet NN demonstrates a significant improvement over classical methods. Most notably, it achieves 100% tagging efficiency, meaning it provides a flavor estimate for every single event, whereas classical combinations often fail for 10-15% of events.
In terms of "tagging power" ($\epsilon_{eff}$), which measures the statistical sensitivity retained, the DeepSet NN provides a 20% to 25% boost over the best classical combinations for Run 2 data. In the challenging Run 3 environment, the DeepSet tagger maintained its superiority, achieving an $\epsilon_{eff}$ of 6.36% for $B^+ \to J/\psi K^+$ decays compared to 3.75% for classical methods.
Furthermore, the inference speed is approximately $7\mu s$ per event, making it fast enough to be integrated directly into the LHCb software trigger for real-time data processing.
Conclusion and Future Work
The transition from "sniper-like" expert systems to a "wide-angle" inclusive approach using DeepSet NNs represents a major step forward for flavour tagging at LHCb. While these results on simulated data are promising, future work will focus on validating these gains with real Run 3 data and fully integrating the system into the LHCb trigger. Refining input selection to better handle high-pileup backgrounds remains a key area for further optimization.
Summary updated