⚛️ high-energy experiments

Full Event Interpretation at Belle II: Current Status and Recent Developments

This paper presents the design, implementation, and recent advancements of the Full Event Interpretation (FEI) algorithm at Belle II, detailing its hierarchical reconstruction strategy, multivariate training improvements, calibration techniques, and future developments aimed at enhancing the efficiency and accuracy of exclusive B-meson decay reconstruction.

Original authors: R. Tiwary, M. Marfoli, V. Vobbilisetti, K. Trabelsi

Published 2026-10-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: R. Tiwary, M. Marfoli, V. Vobbilisetti, K. Trabelsi

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 quiet hum of a laboratory in Tsukuba, Japan, physicists are engaged in a delicate game of cosmic reconstruction. They are studying the B meson, a fleeting particle that appears when electrons and positrons smash together at nearly the speed of light. These particles are unstable and decay almost instantly into a shower of other, lighter particles. To understand the fundamental laws of nature, scientists need to see exactly how these B mesons fall apart, but there is a catch. Sometimes, the decay produces invisible particles, like neutrinos, that slip right through the detectors without leaving a trace. When this happens, the event looks incomplete, like a puzzle with missing pieces. To solve this, researchers use a clever trick: they try to reconstruct the other B meson that was created in the same collision. If they can figure out exactly what happened to that first particle, the laws of physics allow them to deduce the properties of the invisible one. This process of identifying and reconstructing the companion particle is known as "tagging," and it is the key to unlocking secrets about the universe that would otherwise remain hidden.

The paper at hand details the latest evolution of a sophisticated computer program called the Full Event Interpretation, or FEI, which serves as the primary tool for this tagging process at the Belle II experiment. For years, this algorithm has been the workhorse of the collaboration, sorting through millions of collision events to find the rare instances where a B meson decays in a specific way. However, the program relies heavily on simulations—computer models that predict how particles should behave. If these models do not perfectly match reality, the program makes mistakes, missing good events or misidentifying bad ones. The authors of this study describe a major overhaul of the FEI system, driven by a realization that the underlying models of how particles decay were slightly off. By fixing these models and refining how the computer sees the data, they have significantly improved the program's ability to find and correctly identify these elusive particle decays.

The core of the improvement lies in how the researchers modeled the decay of hadronic B mesons, which are particles that break apart into other particles made of quarks. The team discovered that their previous computer simulations were based on outdated measurements and simplified assumptions about how these particles behave. In one specific case involving a decay into a D meson and pions, the old simulation assumed a certain pattern of decay that was actually too broad and inaccurate. This error stemmed from relying on very old data that had large statistical uncertainties. By incorporating new, more precise measurements from recent experiments, the team updated the simulation to reflect the true, more complex structure of these decays. They also corrected a subtle error where the simulation had accidentally counted the same decay path twice, a mistake that had skewed the results. These changes were not minor tweaks; they fundamentally altered how the computer understood the particle's behavior, bringing the digital model into much closer alignment with the physical world.

Alongside these modeling updates, the researchers refined the way the algorithm handles neutral particles, such as photons and neutral pions, which are notoriously difficult to detect. In the past, the selection rules for these particles were overly strict, causing the system to discard many valid signals. The team relaxed these criteria to match the successful methods used by the experiment's predecessor, allowing more data to flow through the system. They also introduced a mathematical constraint that forces the reconstructed mass of a neutral pion to match its known physical value, which sharpens the precision of the measurements. These adjustments allowed the FEI to recover a significant number of events that were previously lost, effectively increasing the sensitivity of the entire experiment.

To ensure these changes actually worked, the team developed a new method for testing the system's accuracy. Instead of relying solely on the old methods, they paired the FEI with a specific, well-understood decay channel involving a D meson and a pion. This new control channel acted as a rigorous benchmark, allowing them to measure exactly how well the simulation matched the real data. The results were striking. The new system showed a dramatic improvement in agreement between the computer models and the actual collision data. The calibration factor, which measures how well the simulation predicts reality, moved from a value of roughly 0.65 to 0.83, indicating a much more reliable system. Furthermore, the purity of the selected events—the fraction of correct identifications among all candidates—rose from 0.56 to 0.63.

The most tangible outcome of this work is a massive boost in the experiment's efficiency. The updated algorithm, now trained on these improved models and refined data, can find and correctly tag B mesons far more often than before. The authors report that the efficiency for identifying positively charged B mesons has increased by approximately 50 percent, while the efficiency for neutral B mesons has risen by about 30 percent. This means that for every hundred events the old system could successfully analyze, the new system can now handle significantly more, providing physicists with a much larger dataset to study rare and forbidden processes. The paper concludes by looking ahead, noting that while these gains are substantial, further improvements are possible. The team suggests that future work could involve even more detailed decay models and the adoption of modern machine-learning techniques, such as graph neural networks, to capture complex relationships between particles that current methods might miss. For now, however, the Full Event Interpretation stands as a more powerful and precise tool, ready to help the Belle II experiment probe the deepest mysteries of matter and antimatter.

Technical Summary: Full Event Interpretation at Belle II

Problem Statement
The Belle II experiment at the SuperKEKB collider aims to study $B$ -meson physics, including rare decays and processes involving missing energy (e.g., neutrinos or invisible particles). A primary challenge in these analyses is that reconstructing only the signal-side $B$ meson ( $B_{sig}$ ) is often insufficient to fully constrain the event kinematics. To address this, the companion $B$ meson ( $B_{tag}$ ) must be reconstructed to infer the missing kinematic quantities and suppress background. While inclusive tagging offers high efficiency, it suffers from low purity. Exclusive tagging provides higher purity but traditionally suffers from low efficiency due to the limited number of reconstructed final states. The Full Event Interpretation (FEI) algorithm was developed to maximize the efficiency of exclusive tagging by reconstructing hundreds of decay channels hierarchically. However, the performance of the FEI relies heavily on the accuracy of simulated data (MC) used for training, and discrepancies between simulation and real data can degrade reconstruction efficiency and purity.

Methodology
The FEI employs a hierarchical reconstruction strategy where final-state particles (charged tracks and neutral clusters) are combined to form secondary particles (e.g., $K^0_S$ , $\Lambda$ , $\pi^0$ ), which are then used to reconstruct intermediate particles (e.g., $D$ , $D^*$ , $\Lambda_c$ ), and finally $B$ -meson candidates. At each stage, multivariate classifiers based on Gradient-Boosted Decision Trees (BDTs) distinguish signal candidates from background.

The paper details a comprehensive program to improve the FEI workflow, focusing on three main areas:

Calibration Strategy: The authors introduced a new calibration method using the hadronic decay channel $B \to D^{(*)}\pi$ . In this approach, a charged pion is paired with an FEI-tagged $B$ candidate, and the recoiling system is examined for $D$ or $D^*$ mesons. This channel is statistically potent and orthogonal to previous semileptonic calibration channels. Calibration factors are derived from the ratio of yields in data and MC ( $C_{FEI} = N_{Data}/N_{MC}$ ), utilizing the $sPlot$ technique for background subtraction.
Simulation Model Improvements: The authors identified deficiencies in the MC modeling of hadronic $B$ decays, specifically regarding branching fractions and resonance structures. They updated the MC to reflect recent measurements (e.g., from Belle II and LHCb) for modes like $B^- \to D^0 \pi^- \pi^0$ and $B^- \to D^0 \pi^- \pi^+ \pi^-$ . This involved correcting the overestimation of $\rho$ components and incorporating intermediate resonant contributions (e.g., $a_1$ ) rather than relying solely on non-resonant phase-space models.
Neutral Particle Reconstruction: The selection criteria for photons were relaxed to match those of the predecessor Belle experiment to recover efficiency. Additionally, a mass-constrained fit was applied to reconstructed $\pi^0$ candidates to improve energy and momentum resolution.

Key Contributions and Results
The implementation of these improvements led to significant performance gains:

Data-MC Agreement: The updated simulation models and reconstruction techniques resulted in a dramatic improvement in the agreement between data and simulation. The overall FEI calibration factor increased from approximately 0.65 to 0.83.
Purity and Efficiency: The tagging purity improved from 0.56 to 0.63. Following a retraining campaign using the updated MC and neutral reconstruction algorithms, the new FEI version demonstrated substantial efficiency gains relative to previous implementations: approximately 50% for $B^+$ tagging and 30% for $B^0$ tagging.
Calibration Factors: Using the $B \to D^{(*)}\pi$ control channel, the retrained FEI achieved calibration factors close to unity ( $C(B^0) = 1.06 \pm 0.02 \pm 0.04$ and $C(B^+) = 1.22 \pm 0.04 \pm 0.05$ ), indicating a robust alignment between data and simulation.

Significance
The paper asserts that the Full Event Interpretation is a cornerstone of the Belle II physics program, enabling precision measurements and searches for rare decays involving missing energy. The reported developments—specifically the refined calibration strategy, improved decay modeling, and optimized neutral reconstruction—are presented as critical steps in maximizing the experiment's physics reach. The authors emphasize that these updates provide a more reliable foundation for physics analyses by reducing discrepancies between simulation and data.

Future Directions
The authors outline several avenues for further improvement without making speculative claims:

Decay Modeling: Continued efforts to better measure branching fractions and characterize resonance structures for poorly measured hadronic modes (e.g., $B^+ \to \Lambda_c p \pi^- \pi^+$ ) to further align MC with data.
Input Features: Potential removal of the $\Delta E$ variable from training to allow its use as an independent fit observable in physics analyses, and the exploration of improved particle identification variables.
Machine Learning: Investigation of modern techniques, specifically the GraFEI framework based on Graph Neural Networks (GNNs), which represents decay chains as graphs to potentially capture correlations better than traditional hierarchical BDTs.

More like this