1 · Detect
P waves travel at about 6 km/s, S waves at about 4 km/s. The gap between them grows by roughly one second for every 8 km of distance.
Earthquake ground-motion prediction
The first stations to record an earthquake already hold clues about what everyone else is about to feel. This work turns those first seconds into predictions: the peak shaking at stations the waves have not reached yet, and the rest of each waveform while the earthquake is still unfolding.
An earthquake early-warning system cannot stop the shaking. It buys the seconds needed to stop trains, shut off gas, park lifts and take cover, and real earthquakes show what those seconds are worth.
Of about 105,000 dead or missing in the 1923 Great Kantō earthquake, about 90,000 died in the fires that followed.
Cabinet Office of Japan, 2023 [1]
Share of 1906 San Francisco losses caused by fire: about $400M in total against $80M from shaking alone.
Derived from USGS figures [2]
Tokyo Gas meters that shut off automatically in the 2011 Tōhoku earthquake, triggered at about 200 gal.
Norito & Inomata, 2012 [3]
Shinkansen trains on 11 March 2011 began emergency braking about 10 s before strong shaking arrived.
Government of Japan, 2013 [4]
Speed of the Toki 325 when it derailed near the 2004 Chūetsu epicentre; none of 154 aboard were injured.
JTSB RA2007-8 [5]
BART trains automatically brake on an earthquake alert, a system in place since 2012.
BART, 2014 [6]
A ShakeAlert-connected water utility closes reservoir valves in under a minute; integration cost about $17,000.
PNWS-AWWA [7]
Fires after the 2011 Tōhoku earthquake that were electrical in origin, across 17 prefectures.
PreventionWeb, 2011 [8]
ShakeAlert alerts for M4.5+ events, Oct 2019 to Sep 2023, that were real earthquakes.
USGS 70263400 [9]
Share of Loma Prieta (1989) injuries caused by falls; in Northridge (1994) over half came from falling objects.
Allen & Melgar, 2019 [10]
Warning on Japan's east coast in 2011. JMA issued its alert 8.6 s after first detection.
JMA; USGS [11]
LA County Fire stations testing ShakeAlert to recall elevators, alert radios and start generators automatically.
Early Warning Labs / PNSN [13]
Action times are illustrative assumptions for this demonstration, not measured values.
Three steps run every second while the earthquake unfolds. The simulation below shows all three at once.
P waves travel at about 6 km/s, S waves at about 4 km/s. The gap between them grows by roughly one second for every 8 km of distance.
The first stations' P waves, distance to the rupture and station geometry give a PGA forecast for stations not yet reached.
Each station's recorded waveform is extended forward and refined every second as new data arrive.
The research is trained and tested on recorded strong-motion data from the NGA-West2 database, using rupture distance (Rrup) to relate each station to the fault. On this page, shaking levels follow BSSA14, a ground-motion model built from the same database, with realistic scatter between events and between stations. Every waveform is built from real recorded earthquakes: real P-wave windows and S-wave trains, aligned to each station's arrival times and scaled to its shaking, so that any epicentre and magnitude can be explored.
Stage 1 · The stations
Click or drag along the fault to choose where the rupture starts.
In order of arrival, first station at the bottom, as in the NGA-West2 plots.
Stage 2 · What the warning buys
Population (each figure is 25,000 people), critical infrastructure, and forecast then recorded PGA at each site.
Each place acts on what it has, its forecast shaking and its seconds of warning.
Stage 3 · Report, ready when the shaking stops
–
Estimated consequences in the four towns, from the shaking each one received. With early warning, people who get about 5 s or more can Drop, Cover and Hold On, gas valves close and power is cut before the shaking, so fewer people are hurt by falls and falling objects and fewer fires start.
Illustrative estimates from empirical loss curves of the PAGER type (Jaiswal & Wald), tuned to a modern, well-built region; 2.6 people per building. Real losses depend on building stock, time of day and local soils.
Warning time for each town, the shaking it received and how many of its safety actions finished before the S wave arrived.
Every earthquake is a bit stronger or weaker than average for its size. The first three stations' P waves reveal how this one compares. The model applies that to the expected decay of shaking with distance and forecasts peak ground acceleration (PGA) at stations tens of kilometres away, before the S wave gets there. Real stations scatter around any forecast, so each forecast carries a range.
–
Once a station's own P wave arrives, the model extends its recorded trace forward in time. The first forecasts are rough. They are refined every second, and about 10 seconds after the P wave they follow the true waveform closely. The amber band is the model's uncertainty.
–
The region, stations, towns, population and infrastructure are fictional. Shaking levels follow the published BSSA14 ground-motion model; waveforms and model forecasts are simulated for demonstration and are not real recordings or results from the research model. The statistics in section 01 are real and sourced above.
Video
The 3½-minute film tells the story in ten parts: what early warning is, why seconds matter, how the forecast works on a real earthquake's records, this dashboard and its three stages, our research on national television in Chile, and the papers behind it. To film your own scenario, set it up on the dashboard and press Record; it records to the end as long as this tab stays open, then Download appears.
No recording yetSet the magnitude, epicentre and early warning on the dashboard, then press Record.