WW3 → SWAN nested wave model · grid intercomparison

Hurricane Ike (2008)

Structured vs. unstructured regional grid sensitivity for a Gulf-of-Mexico–to–Galveston nested wave forecast — Hs, wave period, and mean direction at four locations spanning deep water to the harbor entrance, scored over a 72-hour window bracketing landfall after a 48-hour model spin-up.

Landfall
09/13 07:00 UTC
Intensity at landfall
95 kt / 950 mb
Saffir–Simpson
Category 2
Location
29.3°N 94.7°W

Key finding

After correcting a depth sign-convention bug in the unstructured WW3 grid setup (UNST%SF), the structured and unstructured regional grids agree closely at the reference points for Ike: within 2.9% at the deep-water reference point and 2.3% at the Galveston harbor entrance, for peak significant wave height. Nearer the storm core they diverge more: at Mid Gulf buoy 42001 the unstructured mesh peaks 10% higher (13.2 vs 12.0 m, against 9.25 m observed). Ike crossed the open Gulf on a broadly northwest track, sweeping hurricane-force winds across a wide swath of deep water before striking Galveston nearly head-on.

WW3 · Gulf of Mexico
structured 0.1° grid → boundary spectra
149 nest points → SWAN · Galveston
28,834-node mesh
WW3 · Gulf of Mexico
unstructured 55,691-node mesh → boundary spectra
149 nest points → SWAN · Galveston
28,834-node mesh (same)

Grids & meshes

Both grid types, for both models. WW3's two regional Gulf of Mexico grids (top row) are the actual grids used in this study's WW3 stage. SWAN's unstructured Galveston nest (bottom right) is the actual mesh used in this study's SWAN stage — SWAN's structured Gulf grid (bottom left) is shown for completeness (pywaves supports it) but was not part of the nested pipeline in this study, which always feeds SWAN the unstructured Galveston mesh regardless of WW3's grid type. Markers show the reference points sampled below.

WW3 structured Gulf of Mexico grid
WW3 · structured · 181×131 rectilinear, 0.1° (~10 km)
WW3 unstructured Gulf of Mexico mesh
WW3 · unstructured · 55,691 nodes, 108,734 elements
SWAN structured Gulf of Mexico grid
SWAN · structured · 181×131 rectilinear, 0.1° (illustrative, not used in the nest)
SWAN Galveston unstructured mesh
SWAN · unstructured · 28,834 nodes, 50,841 elements

Wind forcing

Both grids share identical forcing: a GAHM-style asymmetric parametric vortex fit to the HURDAT2 best track, blended with ERA5 ambient wind and pressure. Ike made landfall directly at Galveston, TX.

Wind fix applied for this storm. Buoy comparisons (below) showed GAHM overestimating peak wind by up to ~20-25% against Ike's own reported max sustained wind. Root cause: a Holland-B safety cap (ALLOWANCE) tuned against a weak, decaying storm (Harvey's stalled Aug-29 phase) permitted the profile to peak 25% above the storm's own reported vmax when it bound -- which it rarely did for Harvey, but did routinely for Ike's deeper, more intense pressure profile. Reduced from 1.25 to 1.05 (verified: peak/vmax ratio now 0.95-1.04 for both storms, vs. up to 1.24 for Ike previously) and both grids fully re-run. Bias improved most where the vortex core dominates (e.g. buoy 42001, bias 3.3→2.9 m/s); a residual positive bias remains at all three buoys, consistent with HURDAT2's 1-minute-sustained wind convention running higher than NDBC's 8-minute-average measurement -- a separate, unaddressed effect.

Physics fixes: bottom friction and nest time step

Found while investigating why this WW3→SWAN chain under-predicted Galveston relative to a coupled SCHISM-WWM hindcast of the same storm, even though WW3 alone matches or beats SCHISM-WWM at every other Gulf buoy. Both fixes are now the pipeline default for Ike; neither has been validated for Harvey.

Bottom friction. WW3's coded JONSWAP friction default (GAMMA = -0.067, also what the SCHISM-WWM hindcast uses) damps shelf-transiting swell more than SWAN's own default (-0.038). Switching WW3 to SWAN's value improved Galveston (structured-grid RMSE 1.33→0.97 m, bias -0.92→-0.54 m, measured directly on WW3's own grid) and was neutral-to-better at every one of the other 8 Gulf NDBC stations checked (pooled RMSE 0.98→0.93 m) — not a Galveston-only trade-off.
SWAN nest time step. The nest's BSBT propagation scheme is unconditionally stable but diffusive at high Courant numbers; running it at a 20-minute compute step instead of the 60-minute output stride recovered part of the under-predicted landfall peak (structured-grid Galveston peak 4.50→5.62 m, obs 6.03 m) without moving the pre-landfall bias much.
Combined result, still short of a full fix. With both fixes together in the full nested chain, Galveston bias is -0.88 m (structured) / -0.76 m (unstructured), RMSE 1.33 / 1.21 m, peak 5.26 / 5.30 m against 6.03 m observed — better than before either fix (RMSE was 1.64 m), but each individual fix's gain, measured in isolation, was larger than what survives once both run through the full WW3→SWAN nest together: a standalone SWAN run covering the whole Gulf on the same mesh and wind, with no WW3 stage or nest boundary at all, gets Galveston to RMSE 0.70 m under the same conditions. That gap between the nested chain and a single-domain SWAN run is a real, unexplained difference between WW3's and SWAN's wave physics at this location, and was not pursued further here.

Run design: 48-hour spin-up

Both WW3 and SWAN start from zero wave energy at 09/09 12:00 UTC, 48 hours before the 72-hour scoring window (09/11 12:00–09/14 12:00), and every statistic and chart below is scored on that window only.

Why the spin-up matters. The first version of this comparison started the models cold at 09/11 12:00, but Ike was already a mature Gulf storm: buoys read 7.7 m (42001), 2.9 m (42019) and 1.8 m (42035) at that moment while the model read 0 m, and the 13–16 s swell Ike had radiated ahead of landfall was missing for the first day (model peak periods 2–7 s vs. 13–16 s observed). With the spin-up, structured-grid Hs correlation rose from 0.67 to 0.85 at Mid Gulf, 0.67 to 0.91 at Freeport and 0.59 to 0.69 at Galveston, and the Freeport first-day bias went from −2.4 m to 0.0 m.

Wave parameters by location

Hs, period, and mean wave direction (MWD, 0–360° raw — a jump across the wrap point is not a real direction change) at each of four locations, from deep Gulf water through the SWAN nest boundary to the harbor entrance. Period uses Tp (peak period, from WW3's peak frequency output) for the two Gulf points and T01 (mean period) for the two Galveston points, the field SWAN's own output carried for this run. Hover any chart for hourly values.

Deep water

WW3 · Gulf of Mexico

92.5°W, 26.5°N · regional WW3 grid, open water (~2,900 m)

Structured Unstructured

Hs · significant wave height

peak 11.87 / 12.21 m

Tp · wave period

peak 16.4 / 16.9 s

MWD · mean direction

0–360°, circular

Continental shelf

WW3 · Gulf of Mexico

94.5°W, 28.5°N · regional WW3 grid, shelf water (~30-35 m)

Structured Unstructured

Hs · significant wave height

peak 8.11 / 8.25 m

Tp · wave period

peak 17.2 / 17.9 s

MWD · mean direction

0–360°, circular

Open boundary

SWAN · Galveston

94.28°W, 28.51°N · SWAN mesh, nest boundary node (~35 m)

Structured Unstructured

Hs · significant wave height

peak 4.12 / 4.31 m

T01 · wave period

peak 9.4 / 10.6 s

MWD · mean direction

0–360°, circular

Harbor entrance

SWAN · Galveston

94.79°W, 29.285°N · SWAN mesh, nearshore node (~1.8 m)

Structured Unstructured

Hs · significant wave height

peak 0.95 / 0.98 m

T01 · wave period

peak 8.8 / 10.6 s

MWD · mean direction

0–360°, circular

Peak value summary

LocationHs struct. (m)Hs unstruct. (m) DifferencePeriod struct. (s)Period unstruct. (s)
Deep water11.8712.21+2.9%16.416.9
Continental shelf8.118.25+1.8%17.217.9
Open boundary4.124.31+4.6%9.410.6
Harbor entrance0.950.98+2.3%8.810.6
Reading the difference column: a few percent is the expected, physically reasonable sensitivity to regional grid resolution — the unstructured mesh resolves shelf and nearshore bathymetry more finely than the 0.1° (~10 km) structured grid. Before the UNST%SF fix, the same comparison showed the unstructured grid underestimating the structured grid by a factor of 5–6×, caused by ~98% of the unstructured mesh's nodes being silently marked dry.

Observed vs. modeled — NDBC buoys

Real wave-height observations from three NDBC buoys spanning the same range as the synthetic reference points above, compared against both grid types at their nearest node. Bias, RMSE, and correlation computed on hourly-matched Hs over the 72-hour scoring window.

Mid Gulf

NDBC 42001 · -89.64°W 25.92°N

Compared against the regional WW3 Gulf grid, hourly, nearest valid observation within 30 min.

NDBC observed Structured Unstructured

Hs · wave height

struct: bias +0.66m rmse 1.70m corr 0.90 (n=73)
unstruct: bias +1.19m rmse 2.24m corr 0.90 (n=73)

Wind speed

struct: bias +1.68m/s rmse 3.33m/s corr 0.96 (n=73)
unstruct: bias +1.64m/s rmse 3.60m/s corr 0.95 (n=73)

Tp · peak period

struct: bias +0.87s rmse 2.08s corr 0.69 (n=73)
unstruct: bias +1.35s rmse 2.23s corr 0.76 (n=73)

MWD · direction

circular; bias/RMSE only

struct: n=0, insufficient data
unstruct: n=0, insufficient data

Freeport, TX

NDBC 42019 · -95.34°W 27.91°N

Compared against the regional WW3 Gulf grid, hourly, nearest valid observation within 30 min.

NDBC observed Structured Unstructured

Hs · wave height

struct: bias +0.58m rmse 0.89m corr 0.89 (n=72)
unstruct: bias +0.73m rmse 0.97m corr 0.90 (n=72)

Wind speed

struct: bias +2.54m/s rmse 3.24m/s corr 0.97 (n=73)
unstruct: bias +2.34m/s rmse 2.98m/s corr 0.97 (n=73)

Tp · peak period

struct: bias +1.22s rmse 2.67s corr 0.77 (n=72)
unstruct: bias +1.62s rmse 3.12s corr 0.78 (n=72)

MWD · direction

circular; bias/RMSE only

struct: bias -9.22deg rmse 39.39deg (n=72)
unstruct: bias -6.54deg rmse 37.34deg (n=72)

Galveston, TX

NDBC 42035 · -94.41°W 29.23°N

Compared against the SWAN Galveston mesh, hourly, nearest valid observation within 30 min.

NDBC observed Structured Unstructured

Hs · wave height

struct: bias -0.88m rmse 1.33m corr 0.75 (n=65)
unstruct: bias -0.76m rmse 1.21m corr 0.78 (n=65)

Wind speed

read from the actual GAHM+ERA5 forcing field, not SWAN's own BLOCK wind output -- that diagnostic was found not to reflect the input forcing at all (verified by injecting an extreme value into the wind file and confirming it never appeared in the output)

struct: bias +3.14m/s rmse 4.82m/s corr 0.90 (n=72)
unstruct: bias +3.14m/s rmse 4.82m/s corr 0.90 (n=72)

RTP · peak period

struct: bias -1.36s rmse 3.69s corr 0.64 (n=65)
unstruct: bias +0.01s rmse 2.69s corr 0.80 (n=65)

MWD · direction

circular; bias/RMSE only

struct: bias -3.18deg rmse 44.49deg (n=65)
unstruct: bias +11.95deg rmse 46.12deg (n=65)
What is still wrong, after the fixes above. (1) Galveston remains under-predicted, though less than before: Hs bias is -0.88 m (structured) and -0.76 m (unstructured), peak 5.26 / 5.30 m against 6.03 m observed — see "Physics fixes" above for what was tried and what remains unexplained. (2) Mid Gulf is over-predicted during core passage (peak 12.0 / 13.2 m vs. 9.25 m), which tracks a residual wind overshoot near the core that the GAHM fix reduced but did not eliminate.

Accuracy and runtime, summarized

Raw comparison stats, no verdict — structured and unstructured each win some buoy/parameter comparisons and lose others; see the buoy cards above for the full breakdown per location and parameter.

Buoys where structured has lower RMSE
2 / 3
against NDBC Hs
Buoys where unstructured has lower RMSE
1 / 3
against NDBC Hs
Structured runtime
116 min
WW3 regional, 120 h incl. 48 h spin-up
Unstructured runtime
329 min
2.8× the structured runtime
Why does peak-period correlation differ between the two grids, when Hs correlation barely does? For Ike, correlation against NDBC's peak/dominant period is higher for the unstructured grid at 3 of the 3 buoys (mean gap 0.08), while Hs correlation between the two grids differs by only 0.02 on average at the same buoys. Hs is an integral over the full frequency spectrum, so small differences in local spectral shape between the structured (uniform 0.1°) and unstructured (55,691-node, spatially varying resolution) WW3 solutions mostly average out; peak period picks the single frequency bin with the most energy, so the same differences can shift which bin wins. At Galveston specifically, the structured-forced nest's peak period is worse (RTP bias -1.4 s, corr 0.64) than the unstructured-forced nest (bias +0.0 s, corr 0.80). Why the two WW3 grids differ in swell delivery has not been diagnosed, and the companion report for Hurricane Harvey shows a different split between buoys, so this is not evidence that either grid is generally better.