Speaker
Description
Accurate prediction of localised beam features such as peak loss and rapid envelope oscillations is essential for reliable operation of the Medium Energy Beam Transport (MEBT) at ISIS. Standard neural network surrogate models often struggle to capture these effects due to spectral bias, favouring smooth variations over sharp changes. In this work, we investigate the use of Fourier feature mapping to improve surrogate model performance on high-frequency beam behaviour. The method projects spatial inputs into a sinusoidal basis, enabling the network to represent high-frequency variations more effectively. We train two surrogate models on MEBT simulation data, one using Fourier feature mapping and the other without, and compare their performance in capturing peak regions. Results show that Fourier feature mapping significantly improves peak prediction accuracy while maintaining comparable performance on global beam trends. This demonstrates a straightforward and computationally efficient approach for enhancing surrogate modelling of beam dynamics in low-energy transport systems.
Funding Agency
This work was supported by the Ada Lovelace Centre.
| Paper status | Proceeding files received |
|---|