Cylinder flow with Warp-LBM

Simjecture includes an optional 2D cylinder-flow research tool contributed by Zifei Meng, under Apache-2.0. It uses NVIDIA Warp’s GPU/CPU kernels for D2Q9 MRT lattice-Boltzmann hydrodynamics with a smooth partially saturated solid boundary. NVIDIA Warp and WarpX are separate projects.

Install and run

In Research tools, select Cylinder flow → Install. When the numerical readiness check passes, use Run cylinder wake for the starter simulation or prepare a custom study in chat. The hosted workspace provides its installed runtime; visitors use it without managing global installations.

The terminal equivalent is:

simjecture install warp-lbm
simjecture doctor --profile warp-lbm

The installer keeps the pinned Python packages, contributor source, license, validation records and source/build hashes in .runtime/warp-lbm-cylinder-1.0. Its generated capability descriptor detects CUDA devices and the WSL driver path. CPU execution is available when a CUDA device is absent. Large CPU runs need their own cost estimate; the GPU timing below does not apply to them.

For recorded minimal studies, copy skills/warp-lbm/examples/run_cylinder.py into the research workspace and use lab.run with capability warp-lbm-cylinder-1.0. The agent guide describes controls, arguments and read-only source inspection. Project-owned modified source can be passed with --solver-source and retained as an explicit experiment input.

What the outputs mean

Each run retains its complete configuration, boundary choices, source identity, force histories, a final wake-field crop and a plot. summary.json records units, the completed interval and statistics over complete lift periods. forces.npz and forces.csv retain the whole force history. fields.npz labels the cropped field bounds and time; it is not a full-domain or time-mean field.

Time is tU/D; lengths are in cylinder diameters. Grid size is full domain length and height multiplied by cells per diameter. The contributed reference case has 32 cells across D and an 80D×80D domain: 2560×2560 cells and 160,000 steps. An independent run reproduced St≈0.164717, mean drag≈1.367441 and RMS lift≈0.238978, in about 257 seconds on an RTX 4000 Ada. These are same-case reproduction results, not an error bound for the continuum solution or a universal hardware estimate.

The supplied qualification is for stationary-circle Re=100 studies. New grids, domains, rotation, shapes and Reynolds numbers remain available for investigation and need relevant numerical controls. This is isothermal hydrodynamics; the solver does not evolve magnetic fields or radiation. The distributed version performs forward simulations and does not include the former autodiff optimization workflow.

Research benchmark

Prepare benchmark opens a conversation containing the versioned cylinder-inlet-v1 task. It asks whether moving the inlet from 8D to 40D changes mean drag or shedding frequency by more than 1%, while keeping the outlet, height, Reynolds number and measurement definitions fixed. New simulations, convergence controls, retained evidence and independent review determine success.

This task definition is separate from the published CSV/RZ diagnostic leaderboard. No model ranking is implied until comparable model runs have been performed and reported with hardware, time budget, tokens and inference cost.