PyFlowCL

Compressible and incompressible CFD on structured 3D meshes. Mostly Python, with MPI through mpi4py, PyTorch arrays on CPU or GPU, and a few C++ kernels.

• Contact: Jon MacArt (jmacart@nd.edu)
• Code: code/, a snapshot of the master branch. There is no public repository.
• Guides: data/workstation-setup-pyflow.pdf (environment), data/running-pyflowcl.pdf (running), data/using-paraview.pdf (viewing output). Some links in them point to the group’s internal wiki.

Getting started

• master supports nonreacting compressible flows (for example supersonic flow over a wedge or cone) and some reacting flows. Other branches add features such as incompressible flow. Ask Jon if you want to work with them.
• You need MPI with mpi4py and HDF5 with h5py.
• Example cases are in code/verification/. shear_layer_2D is a good first one and needs no input data.
• Keep the CFL number printed at run time below 1, ideally below 0.5, by reducing dt or refining the grid.

Minimal run

Build a CPU-only environment once, from this directory on a CRC front end:

scripts/setup-env.sh

It installs h5py from PyPI, which has no MPI-IO, so only single-rank runs can write output. Build HDF5 and h5py from source (see the setup guide) before running on multiple ranks.

Then submit the smoke test from crcfe01 or crcfe02:

mkdir -p results
qsub scripts/run-smoke.sge

scripts/driver_smoke.py is the 2D shear layer shrunk to 256×256 and 100 steps. Change the size with NX1, NX2, and NSTEPS. Output goes to results/smoke/ and opens in ParaView. The initial condition uses unseeded random perturbations, so runs are not bitwise identical.