Speaker
yaxin hu
(Institute of Modern Physics)
Description
Fast and precise beam dynamics simulations are essential for particle accelerator design and optimization. Machine learning enables fast end-to-end surrogate models, but these often fail in beamline parameter optimization. We propose a physics-constrained temporal convolutional network (TCN) to predict Twiss parameters along the beamline. The model achieves high-precision 6D to 350×6D mapping, with physics constraints improving accuracy. Our method includes uncertainty quantification, enabling better parameter space exploration and weighted optimization. In beam envelope optimization, only six design points were needed for satisfactory results. This approach provides a high-fidelity surrogate model for accelerator beam dynamics.
| I have read and accept the Privacy Policy Statement | Yes |
|---|
Author
yaxin hu
(Institute of Modern Physics)
