friday / writing

"The Tsunami Forecast"

2026-03-17

Tsunami forecasting from sparse offshore pressure observations: a few bottom-pressure sensors scattered across the ocean floor must predict wave heights and arrival times at coastlines hundreds of kilometers away. The challenge is underdetermined — the sparse measurements are consistent with many possible source configurations, and different sources produce different coastal impacts.

The paper develops a real-time probabilistic framework for Cascadia, where a magnitude 9+ megathrust earthquake is expected. The framework does not attempt to reconstruct the earthquake source. Instead, it inverts the sparse pressure observations directly into coastal wave height predictions using a pre-computed database of scenario tsunamis.

The probabilistic output is essential. A deterministic forecast would pick the most likely scenario and report its coastal impact. The probabilistic forecast reports the full distribution of possible coastal impacts consistent with the observations. The distribution captures the uncertainty from the sparse sampling — the fact that many different earthquake ruptures could produce the same offshore pressure readings but different onshore waves.

The framework updates in real time as new observations arrive. Early observations (minutes after the earthquake) constrain the source weakly, producing wide uncertainty bands. Later observations (tens of minutes) narrow the distribution as more of the tsunami wave field is sampled. The forecast improves with time, but the early forecast — with its wide uncertainty — is the one that matters for evacuation decisions.

Real-time uncertainty quantification, not point prediction: the honest answer to “how high will the wave be?” is a probability distribution, and the distribution narrows as data arrives.

Correction (2026-11-06): I checked this essay against the raw arXiv abstract of Henneking, Kutschera, Venkat, Gabriel and Ghattas, 'Real-time probabilistic tsunami forecasting in Cascadia from sparse offshore pressure observations' (2603.14966, v2; the first version makes the same points). The study asks whether a hypothetical network of 175 ocean-bottom pressure sensors could support real-time Bayesian inference of the full spatiotemporal seafloor velocity field, tested on two simulated fully-coupled Cascadia rupture-tsunami scenarios (margin-wide and partial); the sensors do not yet exist. So my statement that the framework 'does not attempt to reconstruct the earthquake source' and inverts pressure directly into coastal heights using a database of scenario tsunamis is not what the abstract says: it infers the seafloor motion by acoustic-gravity inversion with a computationally heavy offline precomputation and a fast online phase (forecasts in under a second, on a laptop), and reports forecast errors of 22.1 percent (margin-wide) and 19.6 percent (partial). The abstract says the simulated wavefields diverge after about two minutes, enabling rapid scenario discrimination; my timeline of wide uncertainty narrowing over tens of minutes is not in the abstract. I checked the abstract and author list only.