Nanogirl Michelle Dickinson: Could AI predict post earthquake aftershocks?

The Iconic Bar and Restaurant in Christchurch after the February earthquake in 2011. Photo / File

Seven years on from the two destructive earthquakes in Christchurch more than 20,000 smaller earthquakes and aftershocks have kept the city shaking.

Nobody knows when or where the next big quake will hit our country, but a new artificial intelligence project has just proved itself to be the most accurate predictor of where post-quake tremors are likely to take place.

Aftershocks can occur for several years after an initial earthquake and although they still feel like earthquakes, they are called aftershocks because they are related to the initial event.

These aftershocks can be large enough to cause additional physical damage, however, the harm to the people in the city from constantly living in stress can be just as damaging.

A recent report from the Canterbury District Health Board shows that one in five Christchurch residents reported that they felt stressed most or all of the time from the anxiety associated with aftershocks, ongoing construction, and loss of facilities. These stresses have long-term health effects and quake-affected residents are twice as likely to be addicted to smoking and 40 per cent more likely to develop mental health conditions including major depression and post-traumatic stress disorder.

Although the probability of another big quake in the city is falling, not knowing when or where the next aftershock will be keeps the city and its residents in a constant state of high alert. Ideally, residents want to know three things about an upcoming aftershock.

The first is when it's going to occur, the second is where and the third is how big it will be.

Currently, researchers use empirical laws to help predict when and how big, but the location of an aftershock has proved difficult to determine.

New research published in the journal Nature shows a new artificial intelligence based system which could help with the where question and create more accurate predictions when it comes to aftershocks.

The work used a facet of artificial intelligence called deep learning. This is a more advanced form of machine learning where computers can learn from data sets to help them to solve new problems that they haven't been programmed to tackle.

Deep learning allows many possible results to be seen at once and creates a complex map connecting different factors.

In the study, researchers looked at more than 131,000 pairs of earthquake and aftershock readings using data collected from 199 real earthquakes. They then asked the deep learning network to predict the activity of 30,000 different pairs to test its accuracy.

Currently, the standard method used to predict earthquakes is the Coulomb failure stress change model.

In this study, it predicted 58 per cent of the aftershocks accurately. The new deep learning method, however, was accurate 85 per cent of the time when based on a grid of five square kilometres.

In addition to the increased computing power, the researchers also added a new calculation into the prediction model called the von Mises yield criterion. Although this is a standard model used in materials engineering that helps to predict when a material will break under stress, it hasn't been used much when it comes to modelling earthquakes.

More work is needed before the system can be deployed as it still takes too long to process the data in real time and only works on aftershocks with static stress, not dynamic stress.

However, when added to AI systems already in existence that can check multiple databases to guess which residents might have disabilities and need help with evacuation as well as AI that assesses building data, elevation and soil types to predict earthquake damage to buildings, the future of AI may be key to helping New Zealanders feel safer on their shaky land.

Dr Michelle Dickinson, creator of Nanogirl, is a nanotechnologist who is passionate about getting Kiwis hooked on science and engineering. Tweet her your science questions @medickinson.