Accelerating diagnostic analysis using non-surrogate machine learning for improved understanding of Thomson scattering

Over the past 30 years Thomson scattering has become one of the premier diagnostics for temperature and density measurements at large multiple-beam laser facilities. With improved diagnostic expertise and modeling capabilities it is now possible to extract more information Thomson scattering such as the electron distribution function [1], heat transport [2], and local current fluctuations [3]. Speeding up diagnostic analysis allows deeper investigation of the data and more scientists to utilize the data in their experiments.



Thomson scattering is typically analyzed by matching a model spectral density function to lineouts of the measured data often using gradient-descent-based optimization. However, this approach suffers from the curse of dimensionality, making it increasingly difficult to extract more information. A new algorithm that leverages the use of GPUs and automatic differentiation allows for efficient gradient calculation and significantly increased computation speed. These techniques enable 10-100x faster parameter estimation from Thomson scattering in HED plasmas. This improved algorithm also allows estimation of additional parameters at minimal additional cost, resulting in more information from a single spectrum in less time.

References

[1] A. L. Milder et al., Phys. Rev. Lett. 127, 015001 (2021)

[2] R. J. Henchen et al., Phys. Rev. Lett. 121, 125001 (2018)

[3] C. Bruulsema et al., Phys. Plasmas 27, 052104 (2020)