https://github.com/ergodicio/ml-for-lpi
https://github.com/ergodicio/ml-for-lpi
Last synced: 3 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/ergodicio/ml-for-lpi
- Owner: ergodicio
- License: gpl-3.0
- Created: 2024-11-23T02:41:55.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2025-12-05T17:26:31.000Z (7 months ago)
- Last Synced: 2025-12-09T04:59:09.009Z (7 months ago)
- Language: Jupyter Notebook
- Size: 1.79 MB
- Stars: 0
- Watchers: 0
- Forks: 2
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Laser-Plasma Instability Minimization using Differentiable Simulators
This repo contains code used for gradient-based minimization of laser plasma instabilities (LPI) using ADEPT-LPSE
### The repo
This repository shows how to extend `ADEPT` by using one of its existing solvers to perform gradient-based optimization.
The code is of 3 different categories
1. Python scripts that run `ADEPT` in an optimization loop or parameter scan
2. Configuration `yaml` files for `ADEPT`
3. Module files that extend the `ADEPT` functionality by providing parameterized inputs, loss functions, and postprocessing functions
### The physics
We solve the slowly-varying envelope approximation for modeling electron plasma waves driven at a quarter critical surface by a laser beam.
### The optimization problem
We want to minimize the LPI that occurs in a simulation. The free parameters are those that parameterize the bandwidth of the driving laser. Because our simulation is differentiable, we can take a gradient of the simulation with respect to the free parameters.
### Generative Neural Reparameterization
Rather than find just one set of optimal bandwidth parameters, we can choose to learn a generative function that learns the distribution of optimal parameters. This method is described in `Joglekar, A. S. Generative Neural Reparameterization for Differentiable PDE-constrained Optimization. Preprint at http://arxiv.org/abs/2410.12683 (2024).` This repo provides the code for this method.
### ADEPT
`ADEPT` is a differentiable plasma physics simulation tool. It can be found at https://github.com/ergodicio/adept. This particular set of solvers uses a JAX adaptation of the Laser-Plasma Simulation Environment developed at UR-LLE.