Using a dual network with PNN–LSTM nested structure to realize start-to-end accelerator surrogate model

Abstract

Neural network-based surrogate models can provide fast, reliable and realistic models of accelerators. However, training a surrogate model from start-to-end is still a challenge. For accelerators with dozens or even hundreds of elements, training surrogate models can lead to a dimensional disaster. In this paper, we use a dual network with polynomial neural networks (PNNs) and long short-term memory (LSTM) nested structure to realize a start-to-end surrogate model of China Accelerator Facility for Superheavy Element (CAFe2). Results show that the PNN–LSTM model can effectively realize the surrogate model of CAFe2. Compared to the running speed of TraceWin multi-particle simulation, the new surrogate model has improved by about three orders of magnitude while ensuring a sufficiently good level of accuracy. Moreover, our model can efficiently generate a comprehensive set of beam parameters along the entire beamline, significantly enhancing its practicality as a surrogate model.