Abstract
Real-time, bunch-by-bunch monitoring of transverse position, longitudinal phase, and bunch length is crucial for beam control in diffraction-limited storage rings, where complex collective dynamics pose unprecedented diagnostic challenges. This study presents a neural network framework that simultaneously predicts these parameters directly from beam position monitor waveforms. The hybrid architecture integrates specialized Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory with Attention (LSTM-Attention) sub-networks, overcoming key limitations of traditional methods such as serial processing chains and batch-mode operation. Validated on experimental data from the Shanghai Synchrotron Radiation Facility and Hefei Light Source, the model achieves high prediction accuracy with a sub-millisecond latency of 0.042 ms per bunch. This performance demonstrates its potential as a core tool for real-time, multi-parameter diagnostics and active feedback in next-generation light sources.
The Challenge of Real-Time Beam Diagnostics
Next-generation synchrotron radiation sources, known as diffraction-limited storage rings (DLSRs), operate with extremely low emittance and high beam currents. In these environments, narrow vacuum chambers amplify complex collective effects and inter-bunch coupling, making beam stability a significant challenge.
To control the beam effectively, physicists must monitor "macro-particle" dynamics—transverse position, longitudinal phase, and bunch length—for every single bunch in real-time. Traditional diagnostic tools often rely on sequential processing (e.g., calculating phase before position), which allows errors to accumulate. Furthermore, traditional batch-mode operations suffer from latencies on the order of seconds, which is far too slow for the real-time active feedback required to stabilize DLSRs.
Methods: A Hybrid Neural Network Framework
The proposed solution is a unified, end-to-end neural network architecture that takes raw, four-channel Beam Position Monitor (BPM) waveforms as input.
Instead of a monolithic model, the architecture branches into three specialized sub-networks tailored to the physics of each parameter:
- Transverse Position (MLP): Uses a Multilayer Perceptron. Since position is physically derived from the difference-over-sum of electrode amplitudes, it does not require complex time-series modeling. The MLP is computationally lightweight and highly effective for this task.
- Bunch Length (1D-CNN): Employs a 1D Convolutional Neural Network to extract local fluctuations and global spatiotemporal features from the waveforms.
- Longitudinal Phase (CNN + BiLSTM + Attention): To account for systematic phase drift caused by discrepancies between ADC sampling rates and RF frequencies, the authors introduce a Dynamic Feature Fusion Module. This module uses a learnable weighting mechanism to fuse a physical time-shift (
Tshift) compensation term with temporal features extracted by the LSTM.
Experimental Results and Performance
The models were trained and evaluated on real-world experimental data from the Shanghai Synchrotron Radiation Facility (SSRF) and the Hefei Light Source II (HLS-II).
Key performance highlights include:
- Sub-Millisecond Latency: Total inference time is just 0.042 ms per bunch, meeting the requirements for real-time active feedback.
- High Accuracy: On the HLS-II dataset, the model achieved near-perfect bunch length prediction ($R^2 = 0.997$) and excellent transverse position tracking ($R^2 > 0.98$).
- Robustness: By splitting tasks into parallel branches rather than a sequential pipeline, the model avoids the cascading error accumulation typical of traditional algorithmic approaches.
Conclusion and Future Work
This hybrid neural network framework demonstrates the feasibility of real-time, multi-parameter beam diagnostics using deep learning. While the model currently processes single-turn data independently, future work will incorporate multi-turn sequences and multi-bunch interaction modeling to capture complex dependencies. The ultimate goal is to embed this framework directly into feedback hardware to actively suppress beam instabilities in next-generation facilities.
Summary updated