April 15, 2021 by Lianne Frith
While there has been much progress in disaster predictions in certain areas, such as hurricanes, earthquake detection is falling behind. Now, engineering teams are stepping up to the plate, hoping that they can use AI areas, such as machine learning and complex algorithms, to create the beginnings of a solution.
According to the U.S. Geological Survey, there are 20,000 earthquakes around the globe each year, 16 of which are at the magnitude of seven or above. With such a high frequency, you would think that we’d have a better idea of how to predict them, but sadly, that is far from the case. In 2009, Italian scientists were convicted of manslaughter for failing to predict the L’Aquila earthquake that killed over 300 people; however, they were later acquitted, further reflecting the fact that earthquake prediction is far from cut and dry.
The fact is that no scientific body has ever accurately predicted a major earthquake, despite their said prevalence. The question, of course, is why not? And what can we do about it?
The world map superimposed with colour-coded circles that represent the location and intensity of earthquakes from 1900 to 2017
Image credit: Wikimedia Commons
Earthquake prediction has provided a productive hunting ground for amateurs, attention-seekers, and reputable scientists alike. Theories have been put forward that suggest the use of seismic electric signals, rock bursts in underground mines, and radon gas emissions. However, all of them have fallen flat, leaving some prominent seismologists to declare that earthquakes categorically can’t be predicted.
Whether using patterns from past earthquakes or creating mathematical models of tectonic plate movement, what we need is a warning sign that could translate into a date, time, location, and magnitude. However, all studies have faced similar challenges:
Geological connections between changes in nature and earthquakes haven’t been found with reliable evidence to prove that one must occur with the other.
Earthquakes occur miles below the surface, so early indicators (if they exist) may not be able to be detected.
Drilling boreholes into fault zones is extremely expensive and complex.
Modelling how rocks and minerals behave at increased temperatures and pressures towards the earth’s surface is challenging in lab conditions due to limited sample sizes.
Even if an early warning sign shows some promise of detecting a quake, there is not an easy way to decipher whether it will be a major or minor one.
A diagram that represents the general plate tectonic formations most commonly found in the earth’s oceanic crust
Image credit: Wikipedia Commons
What we would need to reliably predict earthquakes are unequivocal precursory signals. Unfortunately, to date, all signals studied occur erratically, and we don’t even know if there is a clear-cut signal there to be found in the first place.
With the threat that earthquakes may put to humans and nature, there is a clear need for a solution—if not a solution to predicting earthquakes, then at least a way to provide an early warning system of such a potential disaster. This is where engineers are stepping into the domain of earthquake detection, using machine learning and algorithms to crunch data, discover patterns, and ultimately warn communities.
Artificial intelligence is well suited to earthquake prediction thanks to its ability to uncover previously overlooked patterns in complex datasets. Seismic events are intrinsically complicated, with a huge volume of variables. Accordingly, AI is being used to try to make sense of the makeup of the ground, interactions between plates, and the way that energy propagates in waves.
Another notable engineering approach includes the use of algorithms, sensors, data processing, and IoT communication systems. While these technologies don’t allow earthquake prediction, they can work together to provide an early warning.
Research published in Nature Geosciences shows the discovery of seismic signals that accurately predicted a fault line’s slow slippage known to precede large earthquakes. The study used a random forest machine-learning algorithm to systematically look for combinations associated with time before fault failure. The statistical feature that the algorithm leant towards wasn’t a precursor event itself, rather the variance of the signal in the moments before failure. The volume of the fault’s acoustic signal was previously discarded as noise, but with the algorithm, the acoustic signal was found to accurately predict lab-based simulations and forecast, using past data, a prediction within a few days.
Research published in Nature has shown that AI can predict aftershock locations more reliably than existing models. The team trained a neural network to look for patterns in 130,000 aftershock events. The deep learning model used the Von Mises criteria to predict when materials will break under stress and apply such an approach to earthquake science. The model provided greatly improved accuracy in predicting aftershocks caused by static stress, although there is also dynamic stress to consider. And, while the model is too slow to work in real-time, it does point at the potential of machine learning to increase the accuracy of predictions.
The OpenEEW project, manufactured by the sensor and data company Grillo in collaboration with IBM, aims to accelerate the standardisation and implementation of early warning systems. The goal is to develop integrated capabilities of detecting and analysing earthquakes so all communities can be alerted. Grillo has used an IoT-based approach to reduce the traditional costs of seismometers and telecommunications. Public alerts can be delivered via social media platforms, a mobile app and an alarm device, within milliseconds of an earthquake being detected. And, while that may only give communities seconds or minutes to react, being given any warning is better than having no time to prepare for such a disaster.
Image to represent a seismic wave that is emitted during an earthquake
Image credit: Pixabay
In a field where scientists have struggled for decades, it seems that engineers are giving us a few glimmers of hope. Without a doubt, machine learning has provided a pathway, allowing us to recognise patterns in audio data and find signals that haven’t been previously identified.
However, we still have a long way to go. While success has been seen in predicting slow slips, there isn’t the same volume of training data for the biggest and rarest earthquakes. We are yet to learn if seismic patterns before small earthquakes are statistically similar to larger ones; and on top of this, there is no way to know whether there will always be an element of randomness, no matter how well we train machines.
At best, we may be able to predict big earthquakes within a measure of weeks, months, or years. While that may not be fast enough for mass evacuation, it could increase public preparedness. And, meanwhile, the provision of open-source sensor implementation, earthquake detection, and alarm sounding should offer a feasible solution to communities—a benefit that was previously reserved only for those with substantial budgets.
