⚛️ high-energy experiments

Low-Latency Beam-Background Suppression for the BESIII Trigger System Using Machine Learning Trained on Experimental Data

This paper presents a data-driven machine learning approach using compact neural networks trained on BESIII experimental data to significantly reduce beam-background false-positive rates in trigger systems while maintaining hadron-event efficiency, demonstrating suitability for low-latency FPGA implementation and future high-luminosity collider upgrades.

Original authors: Hanlin Zhang, Minggang Zhao, Junhao Yin, Boping Chen, Jingzhou Zhao, Wenxuan Gong, Ke Li

Published 2026-10-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hanlin Zhang, Minggang Zhao, Junhao Yin, Boping Chen, Jingzhou Zhao, Wenxuan Gong, Ke Li

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 world of high-energy physics, scientists smash particles together at incredible speeds to uncover the fundamental building blocks of the universe. To do this, they use massive machines called colliders, which accelerate beams of particles until they collide. However, these machines are not perfect. As the beams race around the ring, they occasionally interact with stray gas molecules or the machine's own components, creating a chaotic shower of false signals. These are known as beam-induced backgrounds. Imagine trying to listen to a single violin in a concert hall while a construction crew is drilling nearby; the noise is not part of the music, but it is loud enough to drown it out. In a particle detector, this noise creates a flood of data that overwhelms the computers, forcing them to discard potentially interesting events just to keep up. The challenge for physicists is to build a filter that is fast enough to reject this noise in real-time while keeping the precious, rare signals that reveal new physics.

At the Beijing Spectrometer III, or BESIII, a team of researchers has tackled this problem by teaching a computer to recognize the difference between a genuine particle collision and the machine's own noise. They focused on the trigger system, the first line of defense that decides which events to save and which to ignore. Traditionally, this system uses fixed rules based on what the detectors see, but these rules can be too blunt, letting too much noise through or accidentally cutting out good data. The researchers wanted to see if a more flexible approach, using a type of artificial intelligence called a neural network, could do a better job. They trained this digital brain not on computer simulations, but on real data recorded by the BESIII detector itself. This allowed the system to learn the specific, messy patterns of the machine's background noise as it actually occurs in the real world.

The team tested two different versions of this new system. One version looked only at information from the tracking chambers and timing detectors, while the other looked only at the electromagnetic calorimeter, a device that measures energy. In both cases, the neural network was fed a set of forty-eight different conditions that the detector checks for every single event. The goal was simple: teach the network to spot the difference between a real hadron event, which is a collision of interest, and the background noise. When they compared the new system to the old, established rules, the results were clear. The neural network was significantly better at filtering out the noise. For the version using tracking and timing data, it reduced the amount of background noise slipping through by about twenty percent. For the version using the energy detector, the improvement was even more dramatic, cutting the noise by roughly thirty-one percent.

The researchers did not stop there. They wanted to know if this smart filter would still work when the machine was running at higher energies, where the noise levels are expected to be much worse. They tested the system on data collected at a higher energy level and found that it held up remarkably well. In fact, at this higher energy, the energy-based detector saw its noise reduction jump to nearly sixty percent. This suggests that as the machine becomes more powerful and the noise gets louder, this new method becomes even more valuable. To ensure the system was robust, they also tested it on data collected with a random trigger, which captures events without any specific selection criteria. The network successfully identified and rejected the background noise across different energy levels, proving that it had learned the underlying patterns of the noise rather than just memorizing specific examples.

Finally, the team had to answer a practical question: could this complex computer program actually run on the hardware inside the detector? The trigger system operates with extreme speed, making decisions in mere nanoseconds, and it runs on specialized chips called field-programmable gate arrays. The researchers simulated the neural network on one of these chips and found that it could make a decision in thirty-five nanoseconds. This is fast enough to fit within the tight time limits of the trigger system, and it uses less than half of the available computing resources on the chip. This means the system is not just a theoretical idea but a practical tool that could be installed in the near future.

The implications of this work extend beyond the current experiment. As scientists plan for even more powerful colliders in the future, the amount of background noise will only increase. The ability to use real experimental data to train a system that can distinguish signal from noise offers a new path forward. By improving how these machines filter data, scientists can preserve more of the rare, interesting events that lead to discoveries, ensuring that the next generation of particle physics experiments can see the universe more clearly than ever before.

Technical Summary: Low-Latency Beam-Background Suppression for the BESIII Trigger System Using Machine Learning

Problem Statement
At high-luminosity electron–positron ( $e^+e^-$ ) colliders, beam-induced backgrounds arising from interactions with residual gas, beam losses, and accelerator components constitute a significant fraction of the low-latency event stream. Unlike hadron colliders, where triggers must select rare signatures from a vast background, $e^+e^-$ triggers aim to retain a broad physics program while rejecting dominant machine-related backgrounds. At the Beijing Electron-Positron Collider II (BEPCII), specifically at the $J/\psi$ peak ( $\sqrt{s} = 3.097$ GeV), the rate of beam-background events passing loose Level-1 (L1) requirements is comparable to that of hadronic events. As BEPCII plans future upgrades to increase peak luminosity (e.g., via a crab-waist scheme), the demand on trigger systems to suppress these backgrounds without sacrificing physics efficiency will intensify. Traditional trigger logic relies on validated detector-level information but may lack the flexibility to adapt to time-dependent machine conditions as effectively as data-driven approaches.

Methodology
The study develops a data-driven machine-learning method to suppress beam backgrounds at the trigger level using experimental data recorded by the BESIII detector. The approach utilizes the 48 binary Level-1 trigger conditions generated by the Global Trigger Logic (GTL), which integrates signals from the Main Drift Chamber (MDC), Time-of-Flight (TOF), Electromagnetic Calorimeter (EMC), and Muon Counter (MUC).

Data Samples: The models were trained and validated on real experimental data collected at $\sqrt{s} = 3.097$ GeV and $\sqrt{s} = 4.680$ GeV. Two specific single-channel configurations were analyzed to minimize correlations between subdetector responses: "MDC&TOF only" (Channel 03) and "EMC only" (Channel 12). Random-trigger data collected at various energies served as an unbiased sample to test generalization.
Model Architecture: To satisfy the strict latency and resource constraints of the BESIII trigger system (implemented in FPGAs), compact feedforward neural networks (NNs) with two hidden layers were employed.
Training Strategy:
- Binary Classification: Initial models distinguished signal events (Hadron, Bhabha, Dimuon) from beam-background events.
- Four-Class Classification: Extended models distinguished four classes: Hadron, Bhabha, Dimuon, and Beam Background. The primary objective was to separate Hadron events from beam backgrounds while retaining Bhabha and Dimuon events if they were not classified as background.
- Thresholding: Decisions were made by comparing class scores against operating thresholds selected on validation samples to match the hadron-event efficiency of the established trigger selection while minimizing the false-positive rate (FPR).
Validation: Performance was evaluated against the established traditional trigger selection using metrics such as True-Positive Rate (TPR) and False-Positive Rate (FPR). The robustness of the models was tested on random-trigger data collected at energies not present in the training set.
Hardware Feasibility: The compact network was synthesized using hls4ml targeting an AMD Xilinx Virtex UltraScale+ FPGA to estimate inference latency and resource utilization.

Key Contributions

Data-Driven Training: The study demonstrates the use of experimental BESIII data, rather than Monte Carlo simulations, to train trigger-level classifiers, capturing the actual dependence of beam backgrounds on machine conditions.
Compact Neural Networks: The development of a low-latency, two-hidden-layer NN architecture specifically designed for FPGA deployment within the BESIII trigger system.
Complementary Trigger Logic: The proposal of a background-veto logic that operates alongside existing trigger configurations, offering a complementary optimization path without targeting specific signal signatures.
Cross-Energy Validation: Validation of the model's ability to suppress beam backgrounds across different center-of-mass energies ( $\sqrt{s} = 3.097$ GeV and $4.680$ GeV) and on unbiased random-trigger data.

Results

Background Suppression: Relative to the established trigger selection, the NN models achieved significant reductions in the beam-background FPR while maintaining matched hadron-event TPR.
- At $\sqrt{s} = 3.097$ GeV: The four-class model reduced the FPR by 20.38% for the MDC&TOF-only sample and 31.16% for the EMC-only sample.
- At $\sqrt{s} = 4.680$ GeV: The reductions were 19.28% (MDC&TOF-only) and 59.60% (EMC-only).
Generalization: Tests on random-trigger data at energies ranging from 3.097 GeV to 4.740 GeV confirmed that the classifier, trained at 3.097 GeV, retains high hadron efficiency and substantial beam-background rejection capability across the energy spectrum.
Hardware Feasibility: Hardware-oriented synthesis indicated that the compact network is suitable for low-latency implementation. The estimated inference latency is 35 ns at a 200 MHz clock frequency, with total resource utilization below 50% across all reported FPGA categories (DSP, LUT, FF, BRAM).
Comparison with BDT: A comparison with a Boosted Decision Tree (BDT) model showed that the NN implementation requires fewer storage resources and computational operations, making it more favorable for FPGA deployment.

Significance and Claims
The paper claims that compact neural-network classifiers can serve as a complementary option for future BESIII trigger upgrades. By leveraging experimental data, the method provides a data-driven framework for optimizing trigger systems in high-luminosity environments. The authors position this work as a reference for trigger optimization at future high-luminosity $e^+e^-$ colliders, specifically citing the Super Tau-Charm Facility (STCF), the Circular Electron Positron Collider (CEPC), and the Future Circular Collider (FCC-ee).

The study modestly notes that the current hardware synthesis results are initial estimates and do not yet include the timing associated with the full trigger-selection logic or integration into the complete trigger chain. Consequently, the authors state that further hardware studies incorporating realistic operating scenarios are necessary to validate the latency and resource utilization estimates. The primary contribution is establishing the feasibility and performance potential of data-driven, low-latency background suppression using real experimental data.

Technical Summary: Low-Latency Beam-Background Suppression for the BESIII Trigger System Using Machine Learning

More like this