Magnetic hysteresis, eddy currents, and imperfections during magnet manufacturing significantly impair beam operations in multi-cycling synchrotrons.
In particular, the magnetic fields’ dependency on the cycling history cannot be dynamically addressed with currently available tools in the control room with existing classical magnetic models.
This paper summarizes the advances and first operational tests using data-driven methods, in particular artificial neural networks, to model the static and dynamic effects in the CERN SPS main dipoles, and to implement cycle-by-cycle feed-forward corrections using the existing CERN accelerator control infrastructure.
Specifically, LSTMs with physics-bound loss functions, transformers, and other neural forecasting methods have been employed to model hysteresis and eddy currents using measured magnetic field and current data from the B-Train, the real-time measuring system of the SPS main dipoles.
The models have then been interfaced with the settings management system, LSA, for the CERN accelerators, which propagate the computed corrections in magnetic field to the corresponding correction in current to the magnets’ power converters.
Lastly, results from initial tests in the SPS ring are summarized.