Lingsen Meng

Research  /  04

Tsunami

Tsunami source imaging & warning

Near a subduction zone the tsunami can reach the coast within tens of minutes, so a warning has to rest on what can be measured in the first few. We work on two questions: how quickly the earthquake source can be characterised from seismic waves, and how accurately the initial sea surface can be recovered from tsunami records without assuming a fault model.

Warning from seismic arrays

P waves travel at about 8 km/s; a tsunami in deep water at a few hundred metres per second. Back-projection of the first minutes of seismic data from coastal arrays gives the extent of the rupture, which feeds a simplified source for tsunami simulation. We demonstrated this for the 2011 Tohoku-Oki, 2014 Iquique and 2015 Illapel earthquakes (An & Meng, 2016, GRL).

With local networks (0.7–3.5°), merging back-projections from several station groups avoids artefacts from overlapping phases and coda in single-array images. For the 2003 Tokachi-oki and 2011 Tohoku earthquakes, predictions were available 7 minutes after origin, with amplitude errors of −2.55 to 3.01 m and arrival-time errors of −8.7 to 10.0 minutes (Xie & Meng, 2020, GRL).

tsunami, ~0.2 km/s in deep water coastal seismic arrays P waves, ~8 km/s megathrust rupture
SchematicWhy seismic arrays can warn: the seismic waves reach coastal stations minutes before the tsunami reaches the coast. Schematic, not to scale.

Imaging the initial sea surface

Finite-fault tsunami inversions depend on the assumed fault geometry and rigidity, and computing Green's functions limits their resolution. Two alternatives avoid the fault model altogether.

Time-reversal imaging

Recorded tsunami waveforms are propagated backwards in time and add up constructively at the source, which gives the initial water elevation directly. It suits events with unknown fault geometry and can resolve small secondary sources; we applied it to the 2015 Illapel tsunami (An & Meng, 2017, GRL).

Adjoint-state full-waveform inversion

The adjoint method computes the gradient of the waveform misfit with respect to the whole initial sea surface at the cost of two simulations, which makes dense grids and many stations affordable. It was developed on synthetic sources and the 2014 Iquique tsunami (Zhou et al., 2019, JGR) and applied to near-field pressure-gauge records for rapid source characterisation (Xie et al., 2023, Earth Space Sci.). For the 2024 Noto earthquake, adjoint inversion of tsunami records across the Sea of Japan identified a landslide source in Toyama Bay, the first time a landslide source has been resolved by inverting water-elevation data (Mohanna et al., 2025, Earth Planets Space).

Tsunami propagation over real bathymetry

Warning needs many forward simulations quickly. We are training neural-operator (FNO) surrogates on simulations from our finite-difference solver; the animation below is one of those simulations. Red is sea surface above rest level, blue below.

Sea-surface height, Tohoku-like scenario

Scenario simulation
sea-surface height − to + m (saturated) · shading = water depth

A 2011-Tohoku-like two-patch thrust source (Mw 8.96), not an inversion of the 2011 event. Linear shallow-water equations solved with the group's finite-difference code on GEBCO 2026 30″ bathymetry, with the coast treated as a reflecting wall (no inundation); the solver was checked against GeoClaw and PCOMCOT. Frames every 3 min from 0 to 117 min, interpolated for display. Unpublished work in progress.

Key papers