How researchers across several organizations are Leveraging machine learning to gain insights about Earthquakes?
One can never know when we may befall victim to natural calamities. Every year we lose thousands of lives and witness damages due to floods, earthquakes, landslides, tropic storms and other disasters. With the advancements in modern technologies today, we can analyze data patterns and get an early warning about natural calamities like floods and tropical storms. However, predicting earthquakes have been challenging for researchers for quite some time, until recently. Enter machine learning.
Machine learning is a subtype of artificial intelligence that learns from the user data. Its algorithms can already predict the prices of stocks, help determine if an applicant should be offered loans, sift through huge chemical compound data to find cure for a disease. These applications point towards the vast capabilities of machine learning. Researchers assert that machine learning can be used to predict natural disasters.
Earthquakes typically happen when tectonic plates shift. When this occurs, a high amount of energy is released, which then creates the quaking movement of the earth beneath us. To predict them, researcher need enormous amounts of seismic data to analyze the magnitude and patterns of earthquakes – thus enabling in determining when earthquake may take place. For instance, a team of researchers from University of Cambridge, Los Alamos National Laboratory and Boston University, identified a hidden signal leading up to earthquakes, and used this ‘fingerprint’ to train a machine learning algorithm to predict future earthquakes. In their experiment study, the team used steel blocks to closely mimic the physical forces at work in a real earthquake, and record the seismic signals and sounds that were emitted. Then they employed machine learning to decipher the relationship between the acoustic signal coming from the fault and how close it is to failing.
The team had discovered a particular pattern in the sound, which was earlier dismissed as noise by geologists, and occurs long before an earthquake. The characteristics of this sound pattern can be used to give a precise estimate (within a few percent) of the stress on the fault (that is, how much force is it under) and to estimate the time remaining before failure, which gets more and more precise as failure approaches. The team believes this sound pattern is a direct measure of the elastic energy that is in the system at a given time.
There are other organizations too, who have worked on machine learning models that can help predict earthquakes. Appsilon Data Science has used PyTorch to create machine learning models that can assess structural damage by analyzing before-and-after satellite images of natural disasters. Meanwhile, in 2018, scientists from Harvard and Google published a paper in Nature, explaining how deep learning can help predict aftershock locations more reliably than existing models. They used dataset comprising of around 131,000 earthquakes and aftershocks to build a neural network and tested it on 30,000 events. To their surprise, the neural network model had predicted the aftershock locations more precisely than traditional methods.
Same year, researchers at Columbia University illustrated that machine learning algorithms could pick out different types of earthquakes by using catalog of three years of 46,000 earthquake recordings at The Geysers in California, one of the world’s oldest and largest geothermal fields.
Moreover, seismic data not helps in predicting earthquakes but also volcano eruptions and tsunamis.
The P wave And S Wave Dilemma
Earthquakes waves are generally of two distinct types- P waves (primary waves) and S waves (shear waves). P waves are compression waves that apply a force in the direction of propagation, transmit their energy quite easily through the medium and thus travel quickly. They are called primary waves since they are the first type of wave that arrive at seismic recording stations. Further, it can travel through solids, liquids, and even gases. In contrast, S waves shake the ground in a shearing, or crosswise, motion that is perpendicular to the direction of travel. So, though they travel 60% less slowly than P waves, they are more destructive as they travel side to side or up and down. And they transmit less energy, S waves are difficult to predict.
To offset this, a team of researchers at Stanford, built machine learning based, “Earthquake Transformer” that can also detect S waves. The primary objective was to see, how it worked with earthquakes not included in training data that are used to teach the algorithms what a true earthquake and its seismic phases look like. When tested on readings collected during Japan’s magnitude 6.6 Tottori earthquake, it isolated 21,092 separate events, more than twice what people had found in their original inspection. It only used data from only 18 of the 57 stations that Japanese scientists originally used to study the sequence, and completed analysis in mere 20 minutes which earlier took months of expert labor.
Low Frequency Waves
While P waves and S waves fall under high frequency waves and can be felt by humans, low frequency waves are more challenging to detect. Though high frequency waves tend to dissipate in the ground very close to the earthquake source and cause severe damages, low frequency waves don’t usually inflict much damage but can travel long distances.
Last year, MIT devised a neural network model that allows geologists to track low-frequency seismic waves. The researchers trained a neural network on hundreds of different earthquakes that they were able to simulate to enable the model mimic the different characteristics of the waves that occur due to earthquake and subsequently predict the waves occurring at low frequencies that had gone missing. As a result, when they presented the trained neural network with only the high-frequency seismic waves produced from a new simulated earthquake, the neural network was able to imitate the physics of wave propagation and accurately estimate the quake’s missing low-frequency waves. This implies that if the neural model is fed partial profile of an earthquake, it can fill in the gaps to help geologists have access to a more complete picture of earthquakes – which shall allow them to map the Earth’s internal structures.