Recurrent neural network encoder-decoder surrogate models for replacing computational beam simulations in beamline optimization.

The Compact Laser Plasma Accelerator II at Peking University provides high-gradient proton acceleration with potential applications in medical treatment. However, the laser-generated beam exhibits shot-to-shot fluctuation and the beamline transport system is highly complex, making beam simulation and tuning challenging. The results from beam simulation software may deviate from experimental observations and the long simulation time limits their applicability in online diagnostics and beam tuning. In this work, we propose a recurrent neural network encoder-decoder surrogate model for accelerator beam prediction. This model aligns well with the sequential characteristics of magnet components and beam diagnostic outputs in accelerator systems. Our results show that the proposed model outperforms a multilayer perceptron baseline, and we further leverage it to enable fast genetic algorithm optimization and backpropagation-based optimization of detector outputs.