Psydac-Struphy integration

[spossann]

In order to make an impact in plasma simulations, or scientific computing more broadly with Python, we need to bundle our efforts at NMPP. Specifically, we should aim to merge the Psydac (https://github.com/pyccel/psydac) and Struphy groups and coordinate the development of those packages.

This can be the space for general discussions about this integration.

[spossann]

Overview, current status and future vision:

• Struphy provides a stable, common Python API for plasma physics codes and simulations developed at NMPP. Examples of this API can be found here:

  • https://struphy-hub.github.io/struphy/sections/userguide.html
  • https://struphy-hub.github.io/struphy/_collections/tutorials/tutorial_poisson.html
  • https://struphy-hub.github.io/struphy/_collections/tutorials/tutorial_02_parameter_files.html
  • https://struphy-hub.github.io/struphy/sections/api_guide.html
  • The core of this API is a Propagators - Model - Simulation framework

• Psydac is the FE library used by Struphy, alongside modules for particles and other numerical methods such as SPH, Hermite polynomials and maybe FV in the future. Psydac can and will be used by other codes/institutions, and is therefore a standalone project and a dependency of Struphy.

• Students/researchers at NMPP should implement new Physics codes within the Struphy API, using Psydac's FEEC capabilities within the models Propagators (see MaxwellWeakAmpere.allocate() for an example). This will have the following long term benefits for the codes:

  • Visibility through the Struphy platform (documentation, community, etc.), hence codes will not just vanish
  • Maintenance by IPP employees
  • Unified usage through the common API, establishing an NMPP standard
  • Automatic code improvement via improvements to Psydac and Struphy (e.g. GPU acceleration)

• New FE features (solvers, assembly, etc.) should still be implemented and tested in Psydac, but Physics models should use the Struphy API. This means using the common Propagators - Model - Simulation framework.

• The goal is to have a single "Python" group that uses the same tools and development framework. This will enable us to bring the newest methods to real Physics applications quickly - and to build upon the works from each other.

[spossann]

• Currently, Struphy is using a fork of Psydac called feectools: https://github.com/struphy-hub/feectools. The first goal must thus be to get rid of this fork and use the newest Psydac release directly. This has now been enabled since Psydac is available on PyPI. There are several diffs in the fork that have to be worked out (feectools is 216 commits ahead and 45 commits behind Psydac devel at the time of writing). Most pressing are probably:

  • Use of precompiled kernels for matrix assembly and for examples here
  • remove sympde, gelato and igakit as dependencies
  • Use of different mapping framework and thus assembly of matrices
  • Expose inter-/histopolation matrices in GlobalProjector

We should come up with a strategy to merge these approaches with what is currently available in Psydac devel. I am sure some things are obsolete or should not be used anymore.