Nobody Alive Has Felt the Earthquake ShakeAlert Is Built for, so Someone Checked Whether the Fake Ones Are Good Enough

No seismometer has ever recorded a great earthquake on the Cascadia Subduction Zone. The fault runs offshore from northern California to British Columbia, and the last time it broke along most of its length was in 1700, an event reconstructed from drowned coastal forests, buried marsh soils, and a tsunami that turned up in Japanese records with no local earthquake to explain it. Instruments arrived much later. Everything seismologists know about how a Cascadia megathrust rupture would shake Portland, Seattle or Victoria comes from physics, from analogues on the other side of the Pacific, and from computers.

ShakeAlert in the United States and the system operated by Ocean Networks Canada on the British Columbia side both work the same way in principle: read the first seconds of a rupture, decide fast how big it is going to get, and push an alert ahead of the shaking. Software like that has to be exercised against the thing it exists for. In Cascadia, the thing it exists for has never been recorded.

Synthetic Cascadia ruptures have been generated and used for years, and warning algorithms have been graded on them. What nobody had done was check the simulations themselves. Nine researchers led from the University of Oregon, with colleagues at Ocean Networks Canada and the University of Victoria, and at the Berkeley Seismology Lab, published that check on Aug. 3 in the open-access journal Seismica. Their statement of the gap is blunt: "no published study to date has validated simulations and their output for such applications."

The dataset behind the check holds 112 rupture scenarios spanning magnitude 6.6 to 9.4, built with one-dimensional semi-stochastic forward modelling. Long-period ground motion, the slow heave that matters to tall buildings and bridges, can be computed from the physics of a rupture. Short-period motion, the sharp rattle that trips a detection algorithm, is too chaotic for that, so it is generated statistically instead. Stitch the two together, and you get a waveform that behaves like a recording without ever having been one.

Each of the 112 scenarios produces waveforms at 191 sites strung between Oregon and British Columbia, with Washington in between. That is roughly the footprint of the operational warning networks themselves, which matters for readers on either side of the border: this is not a study of one city.

The authors did not ask whether their synthetic waveforms look convincing to a seismologist's eye. They asked whether the waveforms behave correctly on the three things early-warning software actually keys on: whether an event is detected at all, whether its magnitude is estimated correctly, and how strong the ground motion turns out to be. Those three decide whether an alert fires, how far across the map it reaches, and how many seconds it buys.

A great subduction earthquake takes minutes to finish tearing, so an algorithm reading its opening seconds is working from a fraction of the eventual rupture and tends to lowball it. A simulation set that made the early seconds of a magnitude 9 look tidier than they really are would flatter every algorithm tested on it, and nobody would know.

On the measures they checked, the synthetics held up. The paper concludes that the simulated data represent real earthquake scenarios well and make a valuable component in training warning algorithms for the region. Six of the scenarios were then run through two pieces of live operational software, the ShakeAlert EPIC algorithm and the Ocean Networks Canada algorithm, to demonstrate the kind of test a validated set makes possible.

Few moderate-to-large earthquakes have been instrumentally recorded in Cascadia, and no great one has. Validating synthetic Cascadia megathrust waveforms, therefore, cannot mean comparing them against recorded Cascadia megathrust waveforms, because there are none to compare against. What can be established is whether the simulations behave consistently and realistically on the metrics that decide an alert.

Diego Melgar, a co-author, is a developer of the simulation framework whose output is under validation here, which makes this in part a self-check. Three things cut against that. Seismica publishes its peer review reports alongside the article. The full simulated dataset is openly deposited, as is the modelling code. And the two algorithms used as the test bed were built by separate groups on opposite sides of an international border.