Validation
What the models get right, what they miss, and why
A twin is only as useful as its error bars. This page puts every model output next to an observation or a published benchmark, shows how the 2D answer moves with grid resolution, and records what was fixed from the first prototype.
2025 event: modelled vs observed
Baseline · 10 min · n = 0.03 · runout 600 m| Quantity | Observed | Empirical chain | 2D · 120 m | 2D · 60 m | 2D · 30 m arm | Calibrated · 30 m | Assessment |
|---|---|---|---|---|---|---|---|
| Run-up on the opposite wall | 481 m (470–500) | 247–494 m | 43 m (toe 51 m) | 76 m (toe 94 m) | 135 m (toe 207 m) | 385 m (toe 248 m) | Partial Empirical envelope brackets the observation. 2D wave run-up (≥ 300 m from the slide) reaches 80% of it with the calibrated slide (28% with the baseline) and is not grid-converged. Run-up was not used in calibration, so this is an independent check. |
| Wave at Sawyer Island | ≥ 30 m | a(x) ≈ 81 m | 5.5 m | 12.4 m | 14 m | 30 m | Miss Baseline slide at 30 m: 47% of the lower bound, changing 13% from 60 m, so the shortfall is the assumed slide, not resolution. The calibrated slide meets it, but this is the calibration constraint, not an independent check. |
| Wave height at the mouth (Harbor Island) | 3–4.6 m (10–15 ft) | — | 0.3 m | 0.7 m | outside domain | — | Miss Finest grid gives 18% of the reported height. Over ~45 km the first-order wet/dry fronts, the reconstructed entrance sill and missing dispersion all act to damp the wave. |
| Arrival at the mouth | 19–34 min (woken 05:45–06:00) | — | 22 min 13 s | 21 min 34 s | outside domain | — | Consistent First water level above 0.25 m at Harbor Island. Travel time depends on depth, so this tests the reconstructed bathymetry more than the slide. |
| Near-field breaking wave | ~100 m (Science study simulations) | H_M = 167 m, a_M = 134 m | resolved as a bore | resolved as a bore | resolved as a bore | resolved as a bore | Consistent Heller–Hager impact-zone height compared with the published simulated breaking wave. |
| Wave energy / slide kinetic energy | ≈ 2–30% in laboratory studies (Fritz et al. 2004) | — | 1.2% | 2.5% | 3% | 4.6% | Consistent Rises from 1.2% to 3% as the grid refines: less numerical dissipation in the impact zone. |
Observed values: USGS event page, Alaska Earthquake Center and Shugar et al. (2026, Science). 2D columns 120 / 60 / 30 m use the baseline (assumed) slide, from `pnpm validate:model`; the calibrated column uses the best slide from the inverse problem below (5 min, upper arm), from `pnpm calibrate:slide`. Verdicts grade the calibrated slide where it is an independent test (run-up, energy) and the baseline elsewhere: grading the calibrated slide against its own constraint (Sawyer Island) would be circular. Every verdict and note is computed from the generated data.
Benchmark: Lituya Bay 1958
| Quantity | Published | This twin | |
|---|---|---|---|
| Impulse product P | 3.37 | 3.38 | Match |
| Slide Froude number F | 2.66 | 2.66 | Match |
| Max amplitude a_M | 143 m (152 m physical model) | 144 m | Match |
| Max height H_M | 179 m (162 m physical model) | 180 m | Match |
| x_M (location of a_M) | 1232 m | 1234 m | Match |
| Run-up envelope vs 524 m observed | 524 m | 314–560 m | Match |
Benchmarks the empirical chain (not the 2D solver). Inputs exactly as published by Heller & Hager (Vs = 92 m/s, h = 122 m, s = 92 m, b = 823 m, mₛ = 82.6 × 10⁹ kg, α = 40°); lib/physics/impulse.test.ts pins these values.
Grid convergence (2D solver)
| Grid | Cells | Steps | Wall time | Energy | Conversion |
|---|---|---|---|---|---|
| 180 m | 33,984 | 600 | 2.3 s | 2.88 TJ | 0.6% |
| 120 m | 76,680 | 681 | 6.1 s | 6.18 TJ | 1.2% |
| 60 m | 307,430 | 1,365 | 55.4 s | 12.92 TJ | 2.5% |
| 30 m | 279,000 | 2,541 | 56.3 s | 15.55 TJ | 3% |
Same slide on four grids; the 30 m run uses the upper-arm domain at the DEM's native resolution. Halving the cell from 60 to 30 m changes the wave run-up by 78%, so run-up is not grid-converged, while the wave at Sawyer Island moves only 13%. Run-up on a 40° wall needs finer cells than the 30 m terrain data and non-hydrostatic physics.
Sensitivity to the assumed underwater runout
How far the block travels after entering the water is not observed. On the 30 m upper-arm grid (5 min), varying it from 300 to 1000 m moves the wave run-up over a 37 m range (smoothly): a second-order effect next to grid resolution. Water piled against the wall right next to the block ("toe") is reported separately because it measures the assumed kinematics, not the wave.
Far field: the fjord mouth
| Gauge | 120 m: max · arrival > 1 m | 60 m: max · arrival > 1 m |
|---|---|---|
| South Sawyer terminus (2025) | 32.3 m · 4 s | 30.5 m · 8 s |
| Sawyer Island | 5.5 m · 3 min 40 s | 12.4 m · 3 min 30 s |
| Upper-fjord bend | 2.6 m · 7 min 17 s | 4.5 m · 7 min 8 s |
| Mid-arm vessel route | 2.1 m · 10 min 59 s | 4.1 m · 10 min 38 s |
| Tracy Arm entrance sill | 0.9 m · — | 1.7 m · 19 min 40 s |
| Harbor Island | 0.3 m · — | 0.7 m · — |
| Harbor Island, observed | 3–4.6 m · first water 19–34 min (model, > 0.25 m: 22 min 13 s / 21 min 34 s) | |
35 simulated minutes. Kayakers camped on Harbor Island reported 10–15 ft waves and were woken around 05:45–06:00, 19–34 min after the 05:26 failure (Alaska Earthquake Center).
Inverse problem: which slides reproduce the observations?
18 of 128 members acceptedTravel azimuth controls the wave reaching Sawyer Island more than any other input: slides aimed down-fjord send their energy toward the island, slides aimed straight across spend it on the opposite wall. The accepted travel azimuth press against the edge of the tested range, so the true value may lie beyond it.
| Parameter | Prior | Accepted p10 · median · p90 |
|---|---|---|
| Volume | 55 M m³ – 100 M m³ | 67 M m³ · 86 M m³ · 93 M m³ |
| Impact velocity | 50 m/s – 110 m/s | 55 m/s · 82 m/s · 102 m/s |
| Thickness | 60 m – 160 m | 98 m · 124 m · 152 m |
| Width | 500 m – 1100 m | 706 m · 894 m · 991 m |
| Underwater runout | 300 m – 1000 m | 501 m · 856 m · 954 m |
| Travel azimuth | 175° – 240° | 225° · 234° · 238° |
| Confirmed at 30 m | Sawyer Island | Wave run-up (toe) |
|---|---|---|
| 233° · 138 m thick · 63 m/s · 82 M m³ | 45 m | 226 m (439 m) |
| 230° · 100 m thick · 88 m/s · 88 M m³ | 35.2 m | 350 m (502 m) |
| 231° · 145 m thick · 86 m/s · 90 M m³ | 30 m | 385 m (248 m) |
128-member Latin hypercube over six slide parameters, each run for 5 min on the upper arm at 60 m. 10 pilot members rerun at 30 m give the coarse→fine correction (Sawyer Island ×1.162, R² 0.995); members are accepted when the corrected wave at Sawyer Island reaches the observed ≥ 30 m. Run-up is reported but not used as a constraint because it has not converged with grid size. Regenerate with pnpm calibrate:slide.
Second benchmark: Taan Fiord 2015
| Estimator | 5% · median · 95% | vs 193 m |
|---|---|---|
| Müller, decayed wave H(x) | 130 m · 181 m · 241 m | Brackets |
| Synolakis, impact-zone wave H_M | 297 m · 424 m · 584 m | Outside |
Taan is matched by the decayed-wave estimator; Tracy Arm's 481 m only by the impact-zone one (Impulse Lab). Consistent with Tracy's trimline being a near-field run-up next to the slide, and Taan's a propagated wave on a gentler fan.
Higman et al. (2018): ~51 M m³ entered ~90–100 m of water at ≥ 36 m/s; run-up directly across the fjord reached 193 m. Thickness, width and crossing distance are not published, so 4,000 Monte Carlo draws span them (see lib/physics/montecarlo.ts).
Bathymetry provenance
| Survey | Year | Locality | Ship |
|---|---|---|---|
| H13006 | 2018 | Tracy Arm East | NOAA Ship Rainier |
| H13007 | 2018 | Tracy Arm South | NOAA Ship Fairweather |
| H11998 | 2008 | Southern Holkham Bay | NOAA Ship Fairweather |
| H11760 | 2007 | Northern Endicott Arm | NOAA Ship Fairweather |
| F00503 | 2005 | Entrance to Tracy Arm | NOAA Ship Fairweather |
Deepest point 383 m, against the 290 m assumed by the earlier parametric reconstruction.
Depths from NOAA NOS multibeam BAG surveys (NCEI), shifted MSL (MLLW + 2.617 m, Juneau 9452210). Unsurveyed water, mainly the fjord head exposed after 2018, takes a 30° trough profile capped at the deepest survey within 1 km. The Simulator's "Survey" layer maps it cell by cell.
Audit of the Streamlit prototype
Domain was not Tracy Arm Fixed
Found: The processed topo-bathymetry was clipped to longitudes −134.15° to −133.55°, which lies west of the fjord (Stephens Passage). Tracy Arm runs from Holkham Bay (−133.62°) east to the Sawyer glaciers (−133.1°).
Now: Domain clipped in UTM 8N from Holkham Bay to the South Sawyer terminus; landmarks placed from OpenStreetMap / GNIS coordinates.
Bathymetry raster was a scanned map of another fjord Fixed
Found: The "bathymetry" GeoTIFF held 0–255 pixel values from a map image; its legend, crediting the Greenland Institute of Natural Resources, is visible in the raster.
Now: NOAA multibeam surveys (2005–2018) for 84% of the fjord’s water, shifted from MLLW to MSL; the rest (mainly the post-2018 fjord head) is trough-filled from nearby surveyed depths.
Impulse product parameter terms Fixed
Found: Relative slide mass was inverted (ρw·b·h²/mₛ instead of mₛ/(ρw·b·h²)), the angle term used cos(2α/3) instead of cos(6α/7), and the amplitude used 0.25·P^⅘ instead of (4/9)·P^⅘.
Now: Published Heller & Hager (2010) form; reproduces their Lituya Bay worked example to two decimals.
Solver and run-up definition Fixed
Found: Central differences without upwinding or wet/dry treatment (unstable for a 100 m bore), and run-up taken as "η − bed" along a slice, which is water depth rather than run-up.
Now: Well-balanced, positivity-preserving finite volumes (MUSCL + hydrostatic reconstruction) with wetting/drying; run-up is the highest water level reached on the shore.
Re-implementing the twin was also a review. Each item below was found by inspecting the prototype's data and code, then fixed here.
Known limitations
- Hydrostatic shallow-water physics: no frequency dispersion, so the trailing wave train is under-resolved compared with Boussinesq models.
- The slide is a prescribed rigid-shape bed deformation, not a coupled granular flow; its momentum transfer to the water is only through displacement, and its underwater runout is assumed (see the sensitivity panel).
- Bathymetry is surveyed (NOAA, 2005–2018) except where the glacier retreated after 2018; depth at the 2025 fjord head is trough-filled and is the largest geometric uncertainty.
- Slide scar, trimline and vessel-route positions are approximate; no surveyed coordinates are public yet. The Science study supplement should replace them.
- Still water at mean sea level: no tide, no glacier-ice interaction and no Coriolis.
- Not a warning or forecasting system. Outputs are for research, education and screening only.