Bunch-by-Bunch Prediction of Beam Transverse Position, Phase, and Length in a Storage Ring Using Neural Networks

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.

💡 Research Summary

The paper addresses a critical need in diffraction‑limited storage rings (DLSRs) for real‑time, bunch‑by‑bunch diagnostics of three key beam parameters: transverse position (x, y), longitudinal phase (ϕ), and bunch length (σ). Conventional systems such as HOTCAP rely on sequential processing, batch‑mode operation, and global calibration assumptions that limit both speed (seconds of latency) and accuracy (especially under multi‑bunch, unstable conditions). To overcome these limitations, the authors propose a unified neural‑network framework that directly maps raw broadband BPM waveforms (four electrode channels) together with a Tshift compensation term to the three target parameters.

The architecture is a hybrid of three specialized sub‑networks:

1. MLP branch for transverse position. Because position is essentially a difference‑over‑sum of electrode amplitudes, a shallow multilayer perceptron (input → two hidden layers → output) efficiently learns the global correlation without needing temporal modeling. This eliminates dependence on pre‑calibrated sensitivity constants (kx, ky) used in traditional formulas.

2. 1‑D CNN branch for bunch length. Multi‑scale dilated convolutions capture local time‑frequency patterns in the waveform that encode the Gaussian‑like spectral roll‑off associated with σ. The convolutional stack extracts hierarchical features, enabling sub‑picosecond length estimation.

3. Bi‑LSTM‑Attention branch for longitudinal phase. Phase prediction benefits from long‑range temporal dependencies; a bidirectional LSTM processes the waveform forward and backward, while an attention mechanism emphasizes time steps most informative for phase. The Tshift term—computed as frac(N) · n · k, where frac(N) is the fractional part of the sampling‑to‑RF ratio, n is the turn number, and k is the bunch index—is fed into the attention module, allowing the network to correct systematic phase drift caused by non‑integer sampling.

Data were collected from two facilities with markedly different operating regimes: SSRF (9 bunches, 7,736 turns, quiet, high‑stability) and HLS‑II (35 bunches, 56,673 turns, exhibiting coupled‑bunch oscillations). Both used a 16 GHz ADC, resulting in 32‑point (SSRF) and 78‑point (HLS‑II) waveform windows per bunch. Pre‑processing involved baseline subtraction, dynamic threshold detection of bunch centers, amplitude normalization, and the insertion of the Tshift feature. The final input shape is 4 × N + 1 (N = 32 or 78). Datasets were split 95 %/5 % for training and testing, with a uniform sampling interval of one out of every 20 turns to preserve temporal diversity.

Training employed the Adam optimizer (learning rate 1e‑3) for 200 epochs, with loss functions tailored to each branch (MSE for position, MAE for phase, and a custom spectral loss for length). Evaluation on the held‑out test sets yielded:

• Position RMSE ≈ 1.8 µm (sub‑micrometer accuracy)
• Phase RMSE ≈ 0.018 rad
• Length RMSE ≈ 0.42 ps

Inference latency measured on an NVIDIA RTX 3090 GPU was 0.042 ms per bunch, a four‑ to five‑order‑of‑magnitude improvement over batch‑mode systems. Importantly, the model trained on HLS‑II data generalized well to the quieter SSRF data, and vice‑versa, demonstrating robustness to both stable and unstable beam conditions. Transfer‑learning experiments suggest that only a small fine‑tuning dataset is needed to adapt the framework to new machines such as the upcoming HALF or HEPS facilities.

The authors discuss future directions including model compression (pruning, quantization) and deployment on FPGA/ASIC platforms to achieve microsecond‑scale hardware latency, enabling closed‑loop feedback directly within the accelerator control system. They also note the potential to extend the framework to other diagnostics (e.g., beam loss, energy spread) by adding appropriate branches.

In summary, this work presents a novel, end‑to‑end neural‑network solution that simultaneously predicts transverse position, longitudinal phase, and bunch length with high precision and sub‑millisecond latency. By eliminating the need for sequential processing and extensive calibration, the proposed system offers a practical pathway to real‑time, multi‑parameter beam diagnostics and active feedback in next‑generation synchrotron light sources.