Larzavan is an open, global earthquake early-warning system for every damaging earthquake, from magnitude 5 to 9 and above. Its engine is Zelzana, which estimates how big an earthquake is, second by second, while it is still happening.
Nobody can predict earthquakes: no method has ever reliably forecast one, and Larzavan does not try. Larzavan starts the moment an earthquake begins. The earthquake's waves travel through the ground at a few kilometres per second, while a warning travels at the speed of light. That gap saves lives in three ways:
- Seconds before the shaking. Cities tens to hundreds of kilometres away can get several to tens of seconds of warning: enough to stop trains and lifts, halt surgery, shut gas valves and get people under tables. Japan's warning system slows its bullet trains this way.
- Minutes before a tsunami, the biggest life-saver. Tsunami waves often reach the nearest coasts 15 to 40 minutes after a great earthquake. How big the warning is, and how many people are told to leave, depends on getting the magnitude right in the first minutes.
- The first hour after. Knowing where the shaking was worst sends rescuers to the right places first.
Moderate earthquakes cause havoc too, and they are far more frequent: the 2003 Bam earthquake (magnitude 6.6) killed about 26,000 people, and the 2011 Christchurch earthquake (6.2) killed 185. Since 2000 the USGS has recorded 44,290 earthquakes of magnitude 5 to 6, 3,601 of 6 to 7 and 400 of 7 and above. For magnitude 5 to 7, today's magnitude methods work well; what limits warnings is speed, the blind zone close to the epicentre where shaking arrives before any alert, and too few sensors near cities. Larzavan answers with the fastest possible detection, dense low-cost and citizen sensors near cities, on-site warning (each sensor warns its own location from the first P wave), and shaking forecasts for each city.
Early-warning systems measure an earthquake in its first seconds. For the largest earthquakes, the ones that cause tsunamis, those first seconds are not enough: the rupture is still growing, and the magnitude estimate stalls far below the true size. Japan's first estimate for the 2011 Tohoku earthquake was 7.9; it was a 9.0, so the first tsunami warning was too low. Zelzana is being built to give a correct magnitude, with an honest uncertainty, for earthquakes of magnitude 7.5 to 9 and above, worldwide, from open data.
Phase 1 replays every magnitude 8+ earthquake since 2000 (28 in the USGS catalogue) second by second, with the classic early-warning magnitude (peak P-wave displacement, Kuyuk & Allen 2013), using open recordings within 250 km. The replay is generous to the classic method: P waves are picked near their theoretical arrival and data arrive with no delay.
Controls, where the classic method is known to work
(results/phase1/controls.md): on the 2014 South Napa, California
earthquake (magnitude 6.0) it gives 5.90 four seconds after the origin and stays within 0.3; on the
2019 Ridgecrest foreshock (magnitude 6.4) it ends at 6.20. On the Ridgecrest mainshock (magnitude
7.1), with 78 stations, it already stalls at 6.46.
Results for all 28 earthquakes, from a public Kaggle run, are in
results/phase1/summary.md:
- Only 8 of the 28 have any open recording within 250 km. Sumatra 2004, Maule 2010 and Tohoku 2011 have none.
- On all 8, the classic magnitude ends between 5.5 and 6.9 for earthquakes of magnitude 8.1 to 8.8, and is never within 0.3. On the 2014 Iquique, Chile earthquake (magnitude 8.2), 20 stations give 5.97.
These are the two problems Larzavan is built to solve: the magnitude stalls on the largest earthquakes, and open sensors are missing where they strike.
Phase 2 replays the magnitude from GNSS peak ground displacement, second by second, on the 29
earthquakes (magnitude 6.0 to 9.0) of the open Ruhl et al. (2019) high-rate GNSS database. Each
earthquake is scored with coefficients refitted on the other 28, so the method never sees the
earthquake it is tested on. Results, from a public Kaggle run, are in
results/phase2/summary.md:
| Earthquake | Magnitude | Classic seismic (Phase 1) | GNSS, held out (Phase 2) |
|---|---|---|---|
| Tohoku 2011 | 9.0 | no open station | 8.82 |
| Maule 2010 | 8.8 | no open station | 8.80 |
| Illapel 2015 | 8.3 | 6.21 | 7.54 |
| Tehuantepec 2017 | 8.2 | 6.59 | 7.97 |
| Iquique 2014 | 8.1 | 5.97 | 7.54 |
- GNSS closes most of the gap on the largest earthquakes: Tohoku and Maule end within 0.2.
- It is not enough on its own: some magnitude 8 earthquakes still end 0.5 to 0.8 low, it takes one to four minutes to settle, and on moderate earthquakes such as Napa 2014 the classic method is faster and more accurate. Over all 29, 15 end within 0.3, with a mean error of 0.31.
- The replay uses post-processed positions with no delay, so it is generous to GNSS.
Next, Phase 3: synthetic megaquakes, so later layers can learn from far more large earthquakes than nature has recorded.
- No prediction. Larzavan warns about an earthquake that has already started.
- No public alerts of its own. Alerts reach people through each country's agency; Larzavan publishes research estimates and a public scorecard.
- No unverifiable claims. Every number comes from public code, public data and a public run.
pip install -e ".[dev]"
pytest
python scripts/phase1_megaquakes.py --data-dir data --out-dir results/phase1
python scripts/phase1_megaquakes.py --only us20003k7a --providers IRIS # one earthquake
python scripts/phase2_gnss.py --data-dir data --out-dir results/phase2Seismic data come from the open FDSN data centres (IRIS/EarthScope, GEOFON, ORFEUS, RESIF, INGV and others) through ObsPy, and the catalogue from the USGS. GNSS data come from the Ruhl et al. (2019) database on Zenodo (CC BY-SA 4.0), downloaded on first use.
See docs/PLAN.md.
MIT. Author: Sankur Kundu.