Differentiable Software for Plasma Physics


adept is a differentiable two-fluid code that was used to show how to incorporate kinetic effects into fluid codes using neural networks. It uses JAX and equinox to solve the plasma two-fluid equations in 1D as a proof of concept.

The manuscript that introduces and uses adept is

[arXiv] [IOP - Machine Learning: Science and Technology]

The code is [here]


tsadar performs parameter estimation of Thomson scattering spectra measured at the [Omega Laser Facility] via AD-powered gradient descent using JAX.

The code is [here]

Electrostatic Plasma Dispersion

plasmadisp calculates the complex roots to the electrostatic dispersion relation given by

$$ 1 + \frac{\omega_p^2}{k^2} \int du \frac{d_u g(u)}{\omega/k - u} = 0 $$

where g(u) is the 1D distribution function, here assumed to be a Maxwell-Boltzmann.

Our github can be found here