Laser Wakefield Accelerator modelling with Variational Neural Networks

Abstract

A machine learning model was created to predict the electron spectrum generated by a GeVclass laser wakefield accelerator. The model was constructed from variational convolutional neural networks which mapped the results of secondary laser and plasma diagnostics to the generated electron spectrum. An ensemble of trained networks was used to predict the electron spectrum and to provide an estimation of the uncertainty on that prediction. It is anticipated that this approach will be useful for inferring the electron spectrum prior undergoing any process which can alter or destroy the beam. In addition, the model provides insight into the scaling of electron beam properties due to stochastic fluctuations in the laser energy and plasma electron density.

Data-driven science and machine learning methods in laser–plasma physics

“…To facilitate some hands-on experimentation, we conclude with a short guide on how to get started in implementing the techniques we have discussed in this paper. [95] Forward model Transfer learning Inertial confinement fusion Gonoskov et al, 2019 [106] Forward model Neural network High-harmonic generation Maier et al, 2020 [26] Forward model Linear regression ✗ Laser wakefield acceleration Kluth et al, 2020 [97] Forward model Autoencoder & DJINN Inertial confinement fusion Kirchen et al, 2021 [29] Forward model Neural network ✗ Laser wakefield acceleration Rodimkov et al, 2021 [107] Forward model Neural network ✗ Noise robustness in PIC codes Djordjević et al, 2021 [108] Forward model Neural network ✗ Laser-ion acceleration Watt, 2021 [109] Forward model Neural network ✗ Strong-field QED McClarren et al, 2021 [110] Forward model Neural network ✗ Inertial confinement fusion Simpson et al, 2021 [111] Forward model Neural network ✗ Laser-solid interaction Streeter et al, 2023 [112] Forward model Neural network ✗ Laser wakefield acceleration Krumbügel et al, 1996 [186] Inverse problem Neural network ✗ Spectral phase retrieval Sidky et al, 2005 [274] Inverse problem EM algorithm ✗ X-ray spectrum reconstruction Döpp et al, 2018 [162] Inverse problem Statistical iterative reconstruction ✗ X-ray tomography with betatron radiation Huang et al, 2014 [171] Inverse problem Compressed sensing ✗ ICF radiation analysis Zahavy et al, 2018 [187] Inverse problem Neural network ✗ Spectral phase retrieval Hu et al, 2020 [188] Inverse problem Neural network ✗ Wavefront measurement Ma et al, 2020 [173] Inverse problem Compressed sensing ✗ Compton X-ray tomography Li et al, 2021 [275] Inverse problem Compressed sensing ✗ ICF radiation analysis Howard et al, 2023 [192] Inverse problem Compressed sensing/deep unrolling ✗ Hyperspectral phase imaging Bartels et al, 2000 [208] Optimization Genetic algorithm ✗ High-harmonic generation Yoshitomi et al, 2004 [209] Optimization Genetic algorithm ✗ High-harmonic generation Zamith et al, 2004 [211] Optimization Genetic algorithm ✗ Cluster dynamics Yoshitomi et al, 2004 [209] Optimization Genetic algorithm ✗ Cluster dynamics Nayuki et al, 2005 [212] Optimization Geneti...…”

“…In the work of Simpson et al [ 111 ] , a fully connected neural network with three hidden layers is constructed to assist the analysis of an X-ray spectrometer, which measures the X-rays driven by MeV electrons produced from high-power laser–solid interaction. Finally, Streeter et al [ 112 ] used convolutional neural networks (CNNs) to predict the electron spectrum produced by a laser wakefield accelerator, taking measurements from secondary laser and plasma diagnostics as the inputs.…”

Divergence angle consideration in energy spread measurement for high-quality relativistic electron beam in laser wakefield acceleration

“…In general, the energy spectrum can be fitted to a linear combination of multiple Gaussian distributions by using the algorithm given in Ref. [26]. The central energy of the electron beam is 150 MeV, the simulation time scale is set to 1.4 ns in time steps of 1 ps and the coordinates of all electrons can output at each time step.…”

“…To address this issue, we can draw inspiration from the variational autoencoder (VAE) in the field of machine learning, which decomposes the input waveform into a superposition of multiple Gaussian distributions. [26] The VAE is a neural network model that can map input data to a low-dimensional latent space, and the decoder maps the latent space vectors back to the original input space. By decoding the energy spectrum through using the Gaussian mixture model (GMM), [32] we can generate the corresponding means and variances of the Gaussian distributions and adjust the weights to achieve the best fitting effect.…”

Control systems and data management for high-power laser facilities

“…Since being introduced, DAaaS has become popular among the CLF user community and has been used to build ML models of experimental data [ 135 ] , highlighting its suitability for data management and analysis.…”