josiecerrell/can-it-ford-speed-surface

Can It Ford: a (vehicle speed x flow velocity) load surface

Hydrodynamic load on a passenger-car hull in a flooded roadway, measured as a function of two speeds kept as separate axes: the vehicle's speed over the ground, and the flow speed across the roadway.

Status: PROVISIONAL, and deliberately not frozen

This is a snapshot of an ensemble that is not finished. More seeds and at least one finer grid were planned when this card was written, and this snapshot does not include them. Every figure below is reproducible from the pinned source recorded at the bottom of this card, and the counts may grow.

Nothing here is a placeholder: the numbers are measured, not estimated. But if you are comparing against a later version of this dataset and the ensemble sizes differ, the later one is the better one.

Read this before you read a number

The vehicle is PRESCRIBED, not free. It is a rigid signed-distance-field collider moved along a path. It cannot slide, tip, float or be swept away, because those degrees of freedom are exactly what the scene removes.

No FORD or NO-FORD verdict is derivable from this dataset. It reports a LOAD. Whether that load moves a real vehicle depends on tyre-road friction, suspension and wheel state, none of which are in this scene. See Limitations.

Torque in the source is about the collider centre, not the centre of gravity, and is not carried into this table for that reason.

Why this dataset exists

A 105-paper literature search commissioned for this question found that vehicle-wading studies "reduce stability to failure thresholds (e.g., depth or depth x velocity), not a continuous safe-speed surface resolving vehicle speed independently from current velocity". This dataset is that surface.

That characterisation of the literature is a secondary-source claim. The papers were not read for this card, and the sentence above is quoted from the search's own summary.

The result, in one comparison

If a single relative speed determined the load, then two runs sharing v_rel_mag_ms would carry the same force. They do not.

	lower relative speed	higher relative speed
v_car, v_water (m/s)	2.2, 3.0	4.5, 0.5
|v_rel| (m/s)	3.720	4.528
settled window, 5 seeds	5177 +/- 4.0 N	5675 +/- 13.0 N

Across nine ways of splitting one fixed relative speed, the load spans S = (max - min) / mean of 0.76 to 1.28, and S grows with speed: 1 m/s -> S=0.76, 2 m/s -> S=0.97, 3 m/s -> S=1.07, 4.5 m/s -> S=1.12, 6 m/s -> S=1.28.

Three spreads, and they are not the same size

Conflating them is the easiest way to misread this dataset.

spread	what varies	size	is it an error bar?
seed	5 seeds, one cell	0.066 to 0.338 % (median 0.210)	yes, and it is tiny
split	how one |v_rel| divides into v_car and v_water	76 to 128 %	no, this is the result
window	measurement window f20-60 vs f250-400	-68.9 to +83.9 %	no, it means the load is still changing

Error bars drawn from seed scatter would be invisible and would imply the other two spreads do not exist.

A ratio quoted without its window is not weak, it is unfalsifiable

An earlier version of the source write-up reported this pair as 2.3x from the transient window f20-60. That figure has since been WITHDRAWN by its own author and marked in place rather than deleted, so the error stays visible.

The pair recomputed in every arm that holds both cells:

arm	grid	frames/discard	window	seeds	lower |v_rel|	higher |v_rel|	ratio
c3full	g64	60/20	transient	1	8621.4 N	3811.1 N	2.2622
L2full	g64	400/250	settled	1	5028.4 N	5534.7 N	0.9085
M1s*	g64	400/250	settled	5	5176.5 N	5675.3 N	0.9121
M2s*	g96	400/250	settled	2	5315.9 N	6246.4 N	0.8510

The transient arm discards 20 frames. This project's own settle audit found all 25 local runs need more than 8 discarded, minimum 29, median 48, so a 20-frame discard leaves a window that is still transient by the project's own criterion. The settled arms are the defensible ones.

The inversion is not one arm's opinion. It holds at 0.909 with one seed, 0.912 with five seeds at a different boundary-condition rate, and 0.851 on a finer grid, so it survives a change of seed, of BC rate, and of resolution. Per-cell seed spread is under 0.34 percent.

What this does NOT touch. The non-interchangeability result never rested on this pair, and the pair was always the weakest way to state it: the two cells differ in both variables and in |v_rel|. The iso-|v_rel| arcs hold relative speed exactly fixed and vary only the split, and that is where the result actually lives.

Reproducibility record

The same literature search reports that the field does not supply particle/grid counts, GPU model, wall time per simulated second, multi-GPU scaling or a runnable case in one place. So they are in one place here.

	n_grid = 64	n_grid = 96
water particles	41,636 to 41,674	164,351 to 164,382
cell size dx	0.147215 m	0.098143 m
water depth in cells	2.038	3.057
simulated time per run	14.545 s	13.333 s
mean wall clock per run	6.07 s	29.50 s
wall clock per simulated second	0.417 s/s	2.213 s/s
GPU memory used	about 630 MiB	(g128 is about 3 GB)
runs measured	156	20

Particle counts, dx and depth-in-cells are measured from the shipped table itself. The timings are from the source write-up.

Peak observed GPU memory across the whole session was 4,069 MiB of 97,871 MiB. Memory was never the binding constraint, and it is more useful to say so than to imply the card is reporting a limit that was approached.

• GPU: NVIDIA GH200 120GB, driver 590.48.01, 97,871 MiB, TACC Vista, partition gh. ONE card. Single GPU, not multi-GPU.
• Engine: warpmpm on NVIDIA Warp 1.15.0. NOT Genesis.
• Rigid body: Yaris hull as an SDF collider at --sdf-res 32.

Two qualifications, both from the source write-up: the water particle count varies with the seed (41,636 to 41,649), and every timing was measured with up to four concurrent jobs sharing the one card, so each is an upper bound rather than a dedicated-card benchmark.

Limitations, without which the data overstates itself

1. Prescribed body. Stated above. This is the big one.

2. No wheels, no suspension, no rolling degree of freedom. The hull is a solidified particle cloud. Converting a load into a movement verdict needs a tyre-road coefficient that this scene does not contain, and the published values span a wide range across different conditions that must not be merged:

   condition	coefficient	source
   free-rolling rolling resistance, handbrake disengaged	mu_R = 0.0242 to 0.0250	Nihei et al 2025
   locked-wheel static, sand and gravel worst case	mu_s ~ 0.30	Smith, Modra and Felder 2019
   locked-wheel static, wet AND dry concrete	~0.78	Smith, Modra and Felder 2019

   Nihei et al 2025, "Full-scale experimental assessment of passenger vehicle stability in flooding flow", 10.1016/j.rineng.2025.107189. Title and standing checked against the Crossref record rather than assumed from a resolving link. It carries an erratum, 10.1016/j.rineng.2025.107527 (2025-10-01), which has not been read for this card; the figures above are from the original article. Not retracted. The same paper reports C_D = 1.38 +/- 0.18 and that mu_R decays to about 40 percent of its initial maximum, so a criterion built on peak mu_R is unconservative.

   Sliding friction and rolling resistance are different physical quantities; the span from 0.0242 to 0.78 is a factor of 32 across conditions, not a disagreement about one number. This project's own register records that quoting 0.30 as "the wet-road value" is refuted, and that Nihei's mu_R decays to about 40 percent of its initial maximum, so a criterion built on peak mu_R is unconservative.

3. Three hulls appear in the source table and they are not interchangeable. The canonical Yaris hull carries 327,212 vertices (655,308 faces), against Silverado meshes of 2,108 and 48,706 vertices, a ratio of 155x. The hull column separates them. Note the naming is misleading: the file called "yaris_coarse" has far more vertices than the mesh called Silverado "fine".

4. Resolution is characterised, not converged. An n_grid=96 surface (2 seeds per cell) differs from the five-seed n_grid=64 surface by -19.5 to +40.6 percent across 20 cells (median 5.0). The g96 seed spread is 0.026 to 0.515 percent, so the grid effect is far larger than the seed noise at either resolution. No grid-converged claim should be read off this dataset: two grids cannot establish convergence, and this project's own record shows the displacement measure is non-monotone under refinement.

5. fz_settle_over_analytic_diagnostic is a diagnostic, not a validation. It sits near 2.05 for the Yaris rows. Do not read it as a buoyancy check.

6. Depth is fixed at 0.3 m in every row. This is a two-speed surface, not a three-parameter one.

Files

• load_surface.csv, 368 records, one per run.
• surface_cells.csv, the 20-cell settled surface with per-cell mean, standard deviation and seed count.
• iso_vrel_arcs.csv, the split spread at each measured |v_rel|.
• window_comparison.csv, the same cells in both measurement windows.

Column notes

v_car_ms is the vehicle's speed over the ground along its own long axis. v_water_ms is the flow speed across the roadway, broadside to the vehicle. Both are ground-frame. v_rel_angle_deg_from_broadside is 0 when the relative velocity is purely broadside. force_horiz_mag_N is the magnitude of the time-mean horizontal reaction force, and equals hypot(force_mean_x_N, force_mean_y_N) to machine precision (verified on all 368 rows, worst relative mismatch 2.1e-16). family and family_role identify which experiment a row belongs to; do not pool families without reading family_note.

Provenance

Built from git blob 7eee079a7085c98e140ca61ea708ebb70003d71e (data/r9_speed_surface.tsv at commit 159bf7d in the development repository), pinned by content so it can be re-resolved.

No image, texture, HDRI or mesh is included: asset licence provenance for this project is unresolved, and the build refuses those paths rather than warning.