About and tutorial

This service predicts the maximum water level offshore for a tsunami source that you define on a map. A fast neural-operator model, trained on numerical simulations of megathrust earthquakes in 20 subduction-zone regions, replaces a simulation that would take hours. It is a trial service for researchers: results are surrogate-model estimates, not warnings.

Tutorial in six steps

A 45-second recording of the real interface (no sound). Start the interactive tutorial on the map.

1. Pick a region. Click one of the 20 coloured boxes on the world map (or a name in the list). The map zooms to that subduction zone. Grey dashed lines mark segments that are not covered yet.
2. Paint the fault. The fault surface is drawn as cells of about 50 × 50 km. Click or tap a cell to add 1 m of slip; Shift-click or right-click removes 1 m (on a phone use the − Remove tool). The Brush tool sets a chosen value while you drag; "Fill selected", "Smooth", "Patch from Mw" and the presets help to start from a typical rupture. You can also enter fault parameters or upload a CSV.
3. Edit a single cell. Clicking a cell (with the Add, Remove, Brush or Select tool) also selects it. The "Selected cell" card in the panel shows its number, centre, depth, strike, dip, area and slip; type a value in m or cm and press Enter, use + and −, or press "Apply to all selected" after selecting several cells with the Select tool. Esc clears the selection. The table of painted cells stays available (collapsed).
4. Check Mw and the warnings. The panel gives the moment magnitude, ruptured area, mean and maximum slip. The notes below it repeat the checks of the server (rake, dip, centroid depth, distance to the training ruptures).
5. Submit. "Check and run" sends the source to the server, which checks it again; you confirm any warnings and the job is queued. Painting works without an account; running needs an account (registration with an academic e-mail address is automatic).
6. Read the result. The maximum water-surface elevation over 3 hours is drawn over the same map. The colour bar adapts to the event: the default range runs from 1 % of its maximum to the maximum (rounded), with a Log or Linear scale, your own minimum and maximum, and centimetres instead of metres when the maximum is below 0.5 m. "Go to maximum" flies to the peak; click a point to read the value in m and cm. Change the opacity and download the map, CSV, NPZ/NetCDF and README.

Model domain, absorbing strip and tested depths

In the region view a dashed frame marks the model domain: a box of 8 × 8 degrees on a 30 arc-second grid. The fast model and the solver it was trained on only know this box. Along all four edges an absorbing boundary strip of 60 cells (0.5 degree, about 56 km north-south and 30 to 55 km east-west depending on latitude) is shaded: the solver damps the water level there at every time step so that waves leave the domain without reflecting, and results inside the strip are not reliable. A source whose fault outline reaches into the strip gets a warning. An optional teal layer marks the offshore depth bands (1-10, 10-50, 50-200 m) in which the model's test error is measured; deeper water was not part of that test.

About this map. The relief inside the frame is drawn from the 30 arc-second bathymetry grid that the model and the solver use (ETOPO 2022 at 15 arc-seconds resampled to 30 arc-seconds; GEBCO 2026 for Tohoku), the relief outside from the 5 arc-minute world relief. That is why the picture is sharper inside the frame.

Solver animation (optional)

The fast model returns only the maximum over 3 hours. Tick "Full solver animation (a few minutes)" before you run a scenario to also get frames of the water-surface elevation every 5 minutes of model time (37 frames, 0 to 180 minutes, 1 arc-minute cells). The frames come from the same wet/dry shallow-water solver that produced the training data (CPU, about 4 minutes of computing), started from exactly the same initial condition as the fast model. The fast-model map appears first, then the frames. The result panel switches between the maximum (fast model), the animation, and the animation over the running maximum, with play/pause, a time slider (model time in minutes), speed and loop. The colour scale (Linear / Log, minimum, maximum, unit) applies to the frames too and can be changed without a new run. A comparison line reports how far the fast model's maximum is from the solver's in the 10-50 m depth band (median relative difference over 0.2 degree tiles).

Limits: 1 animation queued or running per user, 3 per user per 24 hours, 15 in the queue in total, at most 25 minutes of computing time per job, results kept 7 days. Animation jobs wait while the compute machine is busy. Animations count towards your daily number of events.

How the numbers are computed

• Magnitude. M₀ = μ Σ (area × slip) over the painted cells with rigidity μ = 40 GPa, the GeoClaw default (it makes one 50 × 50 km cell with 1 m of slip Mw 7.30). The live readout uses the GeoClaw constant, Mw = (2/3)(log₁₀ M₀ − 9.05); the server and the training sources use 9.1, so the review page can show a value that is 0.03 lower.
• Cells. The cells follow the trench axis and the USGS Slab2 interface (Hayes et al., 2018, doi:10.5066/F7PV6JNV) down to 60 km depth, the same geometry source and depth limit as the training ruptures. The models were trained on random rectangular ruptures (single or a few subfaults, lognormal slip), not on a fixed mesh. Each cell is a rectangular subfault given by the centre of its top edge (longitude, latitude, depth below the seafloor, strike, dip, length, width), and the interface dips to the right of the strike. Painting produces the same CSV format as the classic form; no separate physics is involved.
• Source to water level. The subfaults give the seafloor displacement (Okada 1985, elastic half-space, GeoClaw degree-to-metre convention), which is the model input. The model returns the maximum water-surface elevation over 3 hours on a 1 arcmin offshore grid (not run-up, not inundation).
• Guards. Rake outside 60–120 deg, dip outside 3–60 deg, centroid deeper than 80 km, or a centroid far from every training source of the region are refused; milder cases need your confirmation. The map shows the same messages while you paint; the server is the authority.
• Limits. 20 events per 24 hours, 2 queued or running jobs at a time, up to 2000 subfaults, Mw 5.0 to 9.7, source extent up to 1500 km. Training Mw is 7.6 to 9.0/9.1; outside it you get a warning.

The presets are illustrative uniform or tapered patches built from the cells of the region; they are not published slip models. Slip heterogeneity matters: uniform-slip ruptures tend to give lower offshore maxima than heterogeneous ones of the same Mw.

Data and attribution

• Basemap: relief and coastlines rendered by us from NOAA NCEI ETOPO 2022 (5 arcmin worldwide, 15 arcsec in the regions) and GEBCO 2026 (Tohoku). Not for navigation.
• Fault geometry: USGS Slab2. Map library: MapLibre GL JS (BSD-3-Clause), served from this site; no external tile servers or fonts are used.

Back to the map · Advanced form · Register