josiecerrell/can-it-ford-steady-force

Can It Ford: steady-force simulation records

Run records from the Can It Ford steady-force campaign: mostly a 2010 Toyota Yaris hull held fixed, broadside, in a periodic open channel of water 0.30 m deep, simulated with the WARPMPM material point method engine on TACC supercomputers. The index also lists three car-free channel checks, two box calibration runs and one 0.45 m still-water run, and Stages 1 and 1b used channels 6 and 12 m wide as well as 20 m. Each run carries its configuration, final status, wall time, the GPU it ran on, and a per-frame time series of probe depths and speeds and the force and torque read from the held hull.

These are research records, not validated results. Read the caveats before using any number.

Contents

path	what
runs_index.csv	one row per run: stage, status, frames, wall time, GPU, host, vehicle, depth, speed, grid spacing, particle count, engine commit, and the SHA-256 of its time series
runs/stage1*/<run>/config.json	the run's full configuration as recorded by the driver (home-directory paths shortened to $WORK and $SCRATCH)
runs/stage1*/<run>/run_status.json	final status (OK, or RUNNING for the one run stopped by its time cap), frames completed and planned, wall time
runs/stage1*/<run>/timeseries.csv	one row per frame (30 frames per simulated second), byte-identical to the file written on TACC
runs/stage1*/<run>/reduced.json	window means and standard errors (drag coefficients removed; see caveats)
runs/stage1*/*provenance*.txt	the job's start time, node, GPU list, driver, software versions and script hashes, written on the compute node; home-directory paths are shortened in these copies, so they do not match the on-node checksums
runs/vista_gb_probe/	a 60-frame Vista GB200 test (job 1050374) of the Stage 1d driver (sha256 46f9ed09), engine 627367ec, with a 5 s GPU monitor log; one network-check line was removed from its log
viewer/	particle frames for three Stage 1e runs (a fixed random 120,000 of 7.8 million particles for the two flowing runs and 80,000 in still water, every 5 s), six near-car frames at the end of the 1.5 m/s run (every 16th particle), a surface height map per frame, and the Yaris hull mesh, as used by the Space
SHA256SUMS	checksums of every file above

timeseries.csv columns: t (s); for probes ctrl, up4, up2, up1, dn2, side, up4in, up4out, ctrlin, ctrlout: u_* speed (m/s), d_* count depth (m), h_* surface-tracked depth (m, Stage 1e only); Fx, Fy, Fz force on the hull (N) and Tx, Ty, Tz torque (N m) read from the collider; controller, damping, energy and particle-containment diagnostics; wall_s (s).

Hardware and software (read from the files written on each compute node)

stage	TACC job	machine, queue	node	GPUs on the node	runs	approx. SU
Stage 1	3482073	Lonestar6, gpu-a100-dev	c301-002	3 x NVIDIA A100-PCIE-40GB, driver 570.195.03	12	1.1
Stage 1b	3483326	Lonestar6, gpu-a100-dev	c301-002	same	21	4.7
Stage 1d	3483537	Lonestar6, gpu-a100-dev	c301-003	same	7	4.4
Stage 1e	3484188	Lonestar6, gpu-a100-dev	c301-004	same	6	5.55
GB200 test	1050374	Vista, gb	c672-009	4 x NVIDIA GB200 (189,471 MiB each per nvidia-smi, 184 GiB per PyTorch), driver 590.48.01, CUDA 13.1	1	0.25

Software on Lonestar6: Python 3.9.7, PyTorch 2.8.0+cu128, NVIDIA Warp 1.12.1, NumPy 2.0.2. On Vista: Python 3.12.13, PyTorch 2.11.0+cu128, NVIDIA Warp 1.15.0. Engine: WARPMPM (kks32/mpm-engine) at commit 627367ecc0d022b366f825b1c3f60c37f286f1e2 for every run. Allocation BCS20003 (PI Krishna Kumar). Stage 1 used one job's provenance file for all its runs; the per-run job was not recorded. Every run records engine commit 627367ec, and its recorded engine status lists one untracked folder, test_meshes/ (46 of 46 runs).

Measured cost on Lonestar6 A100s: 2.98 to 4.89 ns per particle per substep across the 45 completed runs (3.27 to 4.89 in Stages 1d and 1e), all overheads included. On the Vista GB200 test the timestep loop ran about 2.7 times faster per GPU than on an A100, in one short run.

Caveats (they travel with the data)

• No drag coefficient is valid. Stages 1b and 1d used a particle-count depth; where it was checked, in the Stage 1d 40 s runs, it overstated the free surface 4 m upstream by 4 to 24 percent. Stage 1e measured the surface directly, but no run was steady and its grids disagree. All are removed from reduced.json.
• buoyancy_ratio_ctrl and buoyancy_ratio_lattice in reduced.json are the vertical force over rho g V_sub, an uncalibrated readout diagnostic. In flowing runs the vertical force includes lift, so they are not a buoyancy.
• Stage 1 forces are withdrawn as force data: its speed controller let the flow spin down.
• Forces are not grid-converged. The Stage 1e grid ladder (0.075, 0.05, 0.0375 m) gives a three-grid (Celik et al. 2008) grid convergence index of 221 percent for Fx, apparent order 0.41, outside the asymptotic range; the least-squares method needs a fourth grid. Quote no absolute force as a physical load.
• The vertical force is not calibrated. In still water, Stage 1e's reference box reads 6.5 percent above Archimedes at the measured surface (3.9 percent at the lattice depth) and the Yaris 14.4 percent (11.4 percent). Do not derive buoyancy or flotation from Fz.
• The Stage 1d and 1e runs at 1.5 m/s and 0.30 m choke at the car. The Froude number averaged over the strips beside it exceeds 1 (1.37 to 1.44 in Stage 1e), as a one-dimensional criterion predicts (flow-area blockage 9.9 percent against a 1.2 percent limit at Froude 0.874). The 1.0 m/s runs stay subcritical in that average, although the water just past the car's ends locally reaches about 1.07.
• Torques Tx, Ty and Tz are taken about the collider centre, which sits on the floor under the middle of the car (mid-length, mid-width, floor height), not about the centre of mass or a wheel line.
• Count depths (d_*) overstate the free surface where particles crowd, which happens beside the car and grows with time. Prefer h_*, which only Stage 1e has.
• No run is steady by the campaign's pre-registered test, and no FORD or NO-FORD verdict is computed here.
• The channel has a frictionless floor and walls and a 6 m velocity-relaxation zone at its periodic seam; it is a confined channel, not an open road.
• Stage 1 to 1d runs predate the floor fix: water leaked through the floor plane. In Stages 1 and 1b it stayed below the plane; in Stage 1d a guard re-inserted it as a film just above it. Both bias depths and forces. A floor offset of 0.05 cell in Stage 1e cut floor crossings about 300-fold.

Use

import pandas as pd
idx = pd.read_csv("hf://datasets/josiecerrell/can-it-ford-steady-force/runs_index.csv")
ts = pd.read_csv("hf://datasets/josiecerrell/can-it-ford-steady-force/runs/stage1e/Pf_d030_u100_W20_dx050_fo005/timeseries.csv")

Related

• Interactive page: https://huggingface.co/spaces/josiecerrell/can-it-ford
• Code, analysis and paper: https://github.com/jcerrell-IS/can-it-ford (paper PDF). The steady-force driver and stage plans are not yet public, so the script hashes in the provenance files can be checked but not yet rerun by others.

Licence and citation

CC BY 4.0. Vehicle geometry is derived from the NCAC/CCSA 2010 Toyota Yaris finite-element model (CCSA at George Mason University, sponsored by FHWA; DOI 10.13021/G8JS5D). Josie Cerrell, Claremont McKenna College; NSF REU Site: Cyberinfrastructure Research for Societal Advancement (Award #2447887), GeoElements Lab, TACC / UT Austin.

@misc{cerrell2026canitfordsteadyforce,
  author       = {Cerrell, Josie},
  title        = {Can It Ford: steady-force simulation records},
  year         = {2026},
  publisher    = {Hugging Face},
  note         = {NSF REU Site: Cyberinfrastructure Research for Societal Advancement, Award \#2447887; GeoElements Lab, TACC / UT Austin},
  howpublished = {\url{https://huggingface.co/datasets/josiecerrell/can-it-ford-steady-force}}
}