The earth shakes and the machine learns

Researchers from Harvard and MIT trained a neural network to detect weak earthquakes that existing methods miss ● They tested it on an area in central Oklahoma that started shaking due to the introduction of excess water from oil drilling into dry wells and FRACKING

The most recent development of man-made seismic activity in the central United States has resulted from the injection of water produced as a byproduct of pumping oil under high pressure into underground spaces in nearby wells that are no longer active. Some of the artificial earthquakes are caused by oil and gas extraction using a method known as fracking. Central Oklahoma is one of the areas where a large number of earthquakes have been measured as a result of human activity.

Although as humans we cannot feel the absolute majority of these earthquakes, it is important to discover them as well. This detection can help determine what causes earthquakes, and develop tools to prevent them or at least prepare for them.

A group of researchers from the Massachusetts Institute of Technology, MIT, and Harvard University have found a way to use artificial intelligence to detect earthquakes. Their work was published on February 14, 2018 in the scientific journal Science Advances.

Researchers Thibaut Perol, Michael Garby and Marine Danola write in the article that "between 2008 and 2017, at least nine earthquakes with a magnitude greater than 5.0 occurred in the United States that could have been caused by disposal wells. Most earthquake methods are calibrated to detect medium and large earthquakes. As a result, They tend to miss many of the weak earthquakes It is the key to understanding the causes (natural or human) and ultimately to finding ways to mitigate the seismic damage."

According to them, "Over the past decades, the volume of seismic data has grown exponentially and created a need for more efficient algorithms to discover and locate the location of earthquakes."

"Today's most advanced methods scan the wealth of continuous seismic records in search of repeated seismic signals. We leveraged the latest advances in artificial intelligence and created ConvNetQuake – a neural network for detecting the magnitude and location of earthquakes that left a single waveform imprint. We applied the technique to study the activity The man-made seismicity in Oklahoma we found to occur more than 17 times than those recorded by the Oklahoma Geological Survey. Our algorithm also runs much faster than the old methods."

Real activity or just noise?

The artificial intelligence-based algorithm makes it possible to determine whether a certain seismic activity really occurred or whether it is "noise" in the data.

Conventional earthquake detection methods fail to detect events hidden within even moderate levels of seismic noise. In the usual method, a comparison is made between waves (Waveform) which can be used to detect earthquakes originating in a certain area. This is done by detecting repeated earthquakes. Waveform correction is the most effective method for identifying these repeating earthquakes in seismograph data."

"Although the method is safe and reliable, it requires enormous computing power, so it is impractical to use it for long time series. One approach to reducing the computing requirements is to choose a small series of sample waveforms as a template and adapt the calculation on it to that of the entire data series. For this method There are many limitations, and its accuracy depends on the number of these patterns. However, the location information is lost during the transfer of the database to representative waveforms."

Earthquake prediction - the holy grail of the industry
According to the researchers, "We treated earthquake detection as a supervised classification problem and thus designed the ConvNetQuake network. ConvNetQuake was trained on a large dataset of raw seismic waveforms (as captured by seismographs) and learns how to distinguish between seismic noise and earthquake signals. The waveforms are no longer classified by their similarity to other waveforms, as in previous studies. Instead, we analyze the waveforms with a collection of local non-linear filters."

"During the training phase, the filters are adapted to select features in the waveforms that are most relevant for classification. This process bypasses the need to store a growing library of shape templates. Due to this representation, our algorithm very well detects earthquake signals that it did not see during training. It is more accurate than the newest algorithms And allows to quickly arrive at the analysis of the location probability results of an earthquake source from one station."

"Given the limitations of our algorithm, we ran it on the human-caused earthquakes in central Oklahoma and found that it reveals earthquakes that do not fit the standard descriptions."

The holy grail of this scientific branch is the ability to predict earthquakes. The researchers claim that their system still cannot predict strong earthquakes before they occur. Anyway, without the neural networks it would not even have been possible to think about it even a year or two ago.