Earthquake prediction
Nobody has ever predicted a major earthquake's time, place and size; only long-term probabilities and seconds-long early warnings work.
Open in the interactive tree →Attempts such as Parkfield (a quake was predicted by 1993 and arrived in 2004) and Haicheng (1975, disputed) did not produce a reliable method. Stress deep inside faults cannot be measured directly.
As of October 2026
The USGS says that neither it nor any other scientists have ever predicted a major earthquake. Working tools are hazard forecasts (decades), aftershock forecasts and early-warning systems that give seconds after a rupture starts. The February 2023 Turkey-Syria earthquakes killed about 59,000 people. A University of Texas machine-learning trial in China reported forecasting 70% of earthquakes about a week ahead (14 hits, 1 miss and 8 false alarms in seven months), but the team said it is not yet known whether the method works elsewhere.
What is missing
- Continuous measurement of stress and fluid pressure deep in fault zones
- Reproducible precursor signals (none has proven reliable so far)
- Enough large events for statistics, since they recur over centuries
- Physical models of how a rupture starts
- Dense seismic and GNSS networks in poorer high-risk regions
Becomes possible once solved
- Evacuation and shutdown of critical systems before a quake
- Targeted retrofitting and insurance based on real timing
- Far fewer deaths in earthquake regions
Open steps
- Complete small-quake catalogues High AI leverageDetect and locate the many small quakes standard methods miss, to image fault structure and track slow slip near large faults.
- Tests for reliable precursors Low AI leverageSearch long seismic, GPS and fluid records for signals that repeatedly precede large quakes, with strict out-of-sample tests.
- Fault stress and fluid pressure Medium AI leverageInfer the stress state and fluid pressure of a fault from surface signals, lab ruptures and simulations, since deep faults cannot be drilled everywhere.
- How a rupture starts Low AI leverageExplain how slow slip turns into fast rupture, from lab experiments and physics-based simulations, to know what to look for.
- Short-term probability forecasts Medium AI leverageImprove aftershock and short-horizon probability forecasts and early-warning triggers using richer catalogues.
Where AI could help
Medium AI leverage. AI finds many more small quakes and can search for precursors, but no reliable precursor is known and deep-fault stress cannot be measured.
- Detect and locate tiny quakes to image active faults and slow-slip events in near real time
- Search continuous seismic, GNSS and fluid data for repeatable precursors with strict out-of-sample tests
- Improve aftershock and short-term probability forecasts and early-warning triggers
- Learn from lab and simulated ruptures to infer fault stress state from surface signals
Shown so far
- In August 2021, a Nature Communications review reported that machine-learning earthquake catalogs typically contain at least ten times more earthquakes than standard ones, sharpening images of faults. source
- In October 2023, a University of Texas AI scored 14 hits, 1 miss and 8 false alarms forecasting quakes a week ahead in a seven-month trial in China; the team says it is untested elsewhere. source