Von Karman Vortex Street

A 2D decaying turbulence pseudo-spectral solver, simulating Navier-Stokes flow around a cylindrical rod held fixed at the domain centre, with a uniform horizontal free-stream flowing past it — i.e. the rod's own reference frame as it translates horizontally (a von Karman-style obstacle / vortex street).

Source code: https://github.com/mannetroll/2DVonKarman

Screenshots

The dark PySide6 GUI:

The pure-ASCII terminal front-end:

Install

pip install mannetroll-vonkarman      # from PyPI

This installs two console scripts, simulation and simulation_ascii.

Run

uv run simulation          # the dark PySide6 GUI
uv run simulation_ascii    # pure-ASCII vortex street in the terminal (no Qt)

# or, once installed:
simulation                 # the dark PySide6 GUI
simulation_ascii           # pure-ASCII vortex street in the terminal (no Qt)

Run straight from PyPI without installing anything:

uv run --python 3.13 --with "mannetroll-vonkarman==0.1.0" simulation
uv run --python 3.13 --with "mannetroll-vonkarman==0.1.0" simulation_ascii

Run the GPU build (NVIDIA + CUDA only):

uv run --python 3.13 --with "mannetroll-vonkarman[cuda]==0.1.0" simulation
uv run --python 3.13 --with "mannetroll-vonkarman[cuda]==0.1.0" simulation_ascii

The ASCII front-end runs the same N=512 solver and draws the vorticity field as a live 128×32 character grid (" .:-=+*#%@" brightness ramp, 256-color, the rod marked o), with a small diagnostics header. Flags: --mono, --no-diff, --cmap {Inferno,Ocean,Gray}, --backend {auto,cpu,gpu}, --nsteps, --frames, and the usual --re/--vr/--nr/--cfl/--k0/--n.

Numerics

• Doubly-periodic domain [0, 2*pi)^2, vorticity-streamfunction formulation.
• Pure decaying Navier-Stokes (Visc = 1/Re) — no forcing, no extra terms.
• Pseudo-spectral nonlinear term with 3/2 zero-padding de-aliasing.
• LS-IMEX-RK3 time stepping: viscous term and uniform free-stream advection implicit (Crank-Nicolson per substage), fluctuation advection explicit, low storage.
• Centred rod via volume penalization (exact relaxation, operator-split); solved in the rod frame with a uniform horizontal free-stream (VR, 0).
• All state is float32 / complex64; spectral operators (i*k, the velocity projector i*k/|k|^2, the viscous symbol nu*|k|^2) are precomputed once.
• CPU / GPU backend (technique from cuda/turbo_simulator.py): the same code runs on the CPU via scipy.fft (multithreaded, workers=-1) or on an NVIDIA GPU via CuPy + cupyx.scipy.fft.

GPU (NVIDIA, e.g. RTX 3090)

The solver auto-detects CuPy and a CUDA device. On a GPU box install the extra:

uv sync --extra cuda         # CUDA 13.x (cupy-cuda13x)

Then Auto uses the GPU (the Compute dropdown also offers explicit CPU / GPU; the window title shows the active device). With no CUDA device present the GPU option is greyed out and everything runs on the CPU.

Fast rendering

Each frame the selected field is normalized to a contiguous uint8 2D array, that memory is wrapped directly in a QImage (Format_Indexed8), a 256-entry color table is applied, and a QPixmap is built from it.

Controls

CFL, VR, field, colors and frame skip update live; N, Re, K0, NR take effect on Restart.