Research on Accelerator Beam Orbit Prediction Based on Deep Learning

In high-energy physics and particle application scenarios, the stable and efficient operation of linear accelerators is crucial, and the debugging of beam orbits in the mediumenergy beam transport (MEBT) section has a profound impact on the overall performance. Beam transport is susceptible to complex factors such as electromagnetic interference, equipment aging, and environmental fluctuations, resulting in intricate beam orbit variations and posing a great challenge for accurate prediction. In this study, we integrated the spatio-temporal convolutional network (TCN), Transformer and cross-attention mechanisms to develop a TCN-Transformer-attention model for beam orbit prediction for linear gas pedals. The model utilizes upstream BPM2 and downstream BPM5 data from the MEBT section. Spatial features are extracted through the TCN layer and temporal features are captured through the Transformer encoder layer. The cross-attention mechanism dynamically fuses the spatio-temporal features by calculating the attention weights so that the model can prioritize the key features to enhance the fusion effect. Finally, the fully connected layer realizes highprecision prediction. Based on 44,000 samples collected from the CAFE-II platform, the optimal model parameters are determined through normalization and extensive experiments. The test results show that the model has an MSE of 0.158576, an RMSE of 0.398216, an MAE of 0.311590, and an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}$</tex> of 0.967515, which is better than the traditional method and other hybrid models. The model provides support for real-time gas pedal control, MEBT cross-section optimization and beam orbit calibration, which is expected to promote the development of gas pedal technology.