杏吧原创

We might be able to predict big聽earthquakes in the near future better than we thought, according to research that used machine learning to analyse decades of past earthquakes in California.

There have been many attempts to predict earthquakes using real-world signals, such as , or data catalogues, such as the time intervals between earthquakes, but there has been no .

One , first proposed in the 1970s, was that predictability based on data catalogues wasn鈥檛 possible, because the information was random and so didn鈥檛 contain useful details about future earthquakes.

Now, at the University of California, Davis, and his colleagues have developed an algorithm that gives the probability of an earthquake of magnitude 6.75 or greater happening at various times over a three-year period, in a region including California that covers 10 degrees of latitude and 10 degrees of longitude (about 1 million square kilometres). The model is based on records of earthquakes since 1970 and shows that past earthquake data isn鈥檛 random as previously thought, says Rundle.

Rather than look at the interval between earthquakes, as previous studies have, the team examined the number
of earthquakes per unit of time and then used a machine learning model to look for patterns in the data.

The group then applied statistical techniques that are common in economics, such as an exponential moving average, which gives recent events more significance than older ones.

The model gives 鈥渘owcasts鈥�, which are weather forecast-like probabilities of earthquakes happening in a short, upcoming period of time, such as a 90聽per cent chance of a major earthquake within three years.

The chance of a greater-than-magnitude-6.75 earthquake happening in the region varied over the three years from about 35 per cent to as much as 100 per cent. Although the team doesn鈥檛 give specific times for an earthquake, there is a kink in the hockey stick-shaped pattern of risk that the model produces as the chance of an earthquake rises over time, where the probability quickly tends to certainty. Rundle suggests that this kink is where people should start worrying. Not long after an earthquake, the risk drops again.

The team tested the algorithm鈥檚 鈥渟kill鈥� against the聽California data, rating it as 0.708聽鈥� perfect skill at predicting earthquakes was rated as 1, no skill as 0.5 and always being wrong as 0

Predictable pattern

The pattern of predictability also validates the hotly debated idea of a regional earthquake cycle, a repeating pattern of earthquakes in a particular seismic region, proposed more than a century ago.

鈥淭hat sort of cycle has always been elusive,鈥� says Rundle. 鈥淏ut what we鈥檝e done is show that you can see a regional cycle of activity.鈥�

The key to finding hidden patterns in data previously thought as random was a combination of better data, which began to be collected in California in the 1990s, and advances in machine learning models that are powerful for pattern recognition, says Rundle.

While the team has so far only tested its method on previously existing data, Rundle is keen to try it on other data sets, such as the extensive and high-quality seismological data of Japan, before seeing if the model can be used to mitigate the impact of future earthquakes.

Some people aren鈥檛 as convinced that this method will offer particular advances over past methods, or whether its probability-focused forecasts are useful for mitigating risks. 鈥淚t鈥檚 always good to know about these things with probability, but there鈥檚 not a lot you can do when things are relatively vague,鈥� says at the British Geological Survey.

Relying purely on data and not having a physical explanation of the underlying seismological processes is also a weakness for the model, says Musson. 鈥淓arthquakes are not numbers 鈥� earthquakes are real things. They have physical properties, and they have physical dimensions, and numbers are, at best, only a very approximate representation of the seismological reality.鈥�

Reference: