Subscriber OnlyInnovation

How close is a reliable prediction system for earthquakes?

Substantial investment needed to build a comprehensive global multi-sensor network

A professor of mathematics in Utrecht, Buys Ballot, founded the world’s first meteorological centre in January 1854, analysing weather reports from four stations along the Dutch coastline. A few months later Robert Fitzroy, the captain of HMS Beagle during the historic global circumnavigation with Charles Darwin, founded the UK’s Meteorological Office.

By June 1860 Ballot had established a regular service to warn of storms, based simply on deviations from average barometric pressure. The next year Fitzroy also started a storm-warning system after a disastrous shipwreck off Anglesey. Unlike Ballot’s four Dutch stations, Fitzroy connected 50 weather reporting stations around the British Isles by telegraph to his office in London. As a byproduct of his analysis of atmospheric data, Fitzroy also inaugurated a daily 24-hour national weather forecast, first published by the Times in London that August. Although his forecasts were reasonably correct, there was considerable public criticism of inaccuracies. Fitzroy died by suicide in April 1865.

Weather forecasting today is remarkably accurate, saving countless lives at sea, on land and in the air. It catalyses economic activity for farming and the marine, and warns of potential damage to physical assets from storms. Given the phenomenal success of weather prediction, why do we not yet have a similar prediction system for earthquakes since they too can cause immense loss of life and economic disaster?

For centuries humans have observed some animals appear to become alarmed a few seconds, or occasionally hours or days, before an earthquake. However, other factors may also cause animals to behave unusually when an earthquake is not imminent. Animals are thus not considered a reliable predictor.


Predicting events is the curse of forecasting, as Fitzroy observed. Failure to predict an earthquake or reassuring the public about an earthquake that then happens, have led to criminal proceedings against scientists (in China for the Tangshan earthquake in 1976, and in Italy for the Abruzzo quake in 2009). Equally, predicting an earthquake that does not happen can cause a panicked population to flee and unnecessarily disrupt economic activity. Aesop’s fable of The Boy Who Cried Wolf comes to mind.

Weather forecasters generally benefit from a far more consistent range of data and readings than the variability in available earthquake data culled around the globe

Substantial research into earthquakes has yet to lead to a reliable prediction system. Radon gas emissions from pre-seismic rock fracturing, changes in the density of crystalline rock, changes in water ions in stressed rock, electromagnetic anomalies in geoelectric voltages, thermal deviations from average land and sea temperatures, and monitoring of both seismic vibrations and magnetic anomalies, are all used in various combinations across the globe.

Mathematical models, similar in spirit to the huge numerical models of the atmosphere now used regularly for weather forecasting, have been conceived. But in general, complex models do not yet seem to give more accurate predictions than simpler approaches using probabilities based on average historical occurrences.

The challenges compared with weather forecasting include both the relative paucity of sensor data and rate of occurrence. Weather forecasters generally benefit from a far more consistent range of data and readings than the variability in available earthquake data culled around the globe. Severe weather events are also increasing, with some 7,400 worldwide in 2000-2019. Severe seismic activity, of Richter magnitude of 7.0 or more, by contrast occurred about 300 times globally over the same period. There are thus many more opportunities to analyse storms than earthquakes.

Earthquake research continues. In many fields machine learning can detect intricate data patterns sufficient for predictions, often compensating for a lack of a workable mathematic model. We may not yet fully understand earthquake physics but the rapid advances in machine learning approaches perhaps suggest we might not need to do so, and can use data mining instead.

Tectonic stress changes can cause electrically charged ripples that propagate from the Earth’s surface into the atmosphere. Last July a team of Israeli researchers announced they could predict a major earthquake up to 48 hours in advance with 80 per cent accuracy, having retrospectively tuned a machine-learning algorithm on archived ionosphere data filtered for other effects such as solar flares. A Beijing team has also reported that ionosphere changes were discovered in historical satellite data, up to two weeks before two large Chinese earthquakes in 2021 and 2022.

A recent paper from Northwestern University in Illinois claims improved predictions by modelling partial releases of seismic stresses, whilst Los Alamos Laboratory in New Mexico has been using surrogate data from in-house analogies to compensate for the lack of real seismic data available for machine learning.

Fitzroy likened his systematic collection of weather data from multiple viewpoints to the four Synoptic Gospels, the telling of the life of Christ from different perspectives. His synoptic weather forecasting is now the norm. If we are to advance earthquake prediction, a substantial international investment will be needed to build a comprehensive global multi-sensor network – a synoptic earthquake network.