Machine learning of plasma fluid closures

In this Hackweek project, we first build a simple machine learning model for plasma fluid closures. We enhance the model's performance and extend the its capability, during which we learn about basic machine learning techniques and best practices.

Tentative project tasks

Note: Please suggest any topics in machine learning or python GPU that you find potentially useful for space science (for example, for your own work).

• Set up a handy Python environment and clone this repository following the introduction.
• We will first build a simple model to learn the closure given by equation (1).
  • The input/feature data T consists sinusoidal modes of random phases and amplitudes.
  • The output/label data q will be computed from the closure equation (1).
  • The model has only three fully-connected (dense) layers, i.e., only one hidden layer.
  • The model is implemented using TensorFlow and is demonstrated in the Jupyter notebook
• We will enhance the simple model. Tentative tasks are
  • Learn how to apply a model learned from data of lower-resolution to data of higher-resolution.
  • Learn and try Convolutional Neural Networks(CNN).
  • Learn how to train data batch-by-batch.
  • Learn how to find optimal hyper-parameters, i.e., those related to machine learning algorithms.
  • Learn how to accelerate the code using GPU.
  • Learn the saved model weights etc. and possibly how to use them outside of TensorFlow.
  • Learn to use PyTorch.
  • Extend the code to 2d using CNN and GPU.
• Other tasks
  • Accelerate non-ML codes using Numba and cupy, and possibly Dask. A base example can be found here.
  • Make interactive plots using bokeh or plotly.
  • Learn common open-source software licenses.
  • Build a pip or anaconda distributable package.
  • Create DOI for citing the project.
  • Build an affiliated website using Github Pages.

Team

• David Hall (Nvidia; Support)
• Sheng Huang (Boston University)
• Colin Komar (NASA/GSFC)
• Jian Li (NASA/GSFC; Support)
• Ryan McGranaghan (ASTRA)
• Saumitra Mukherjee (Jawaharlal Nehru University)
• Lutz Rastaetter (NASA/GSFC)
• Liang Wang (Princeton University; Team lead)
• Yihua Zheng (NASA/GSFC)

Project folder structure

• notebooks: Demo Jupyter Notebooks.
• contirbutors: Folder to store stuff for individual contributors.
• docs: Background introduction and technical notes to get started.
• scripts: Miscellaneous scripts.