Anticipating Earthquakes With the Help of AI

Major earthquakes can cause catastrophic loss of life and property, and they can occur seemingly at random in most parts of the world. For decades geologists and earthquake experts have been working to develop strategies to predict these damaging quakes, but with limited success.

Now, though, scientists from Stanford University and other earthquake research centers are applying advanced machine learning technologies developed for voice and image recognition to analyze seismic signals from millions of tiny temblors that go unnoticed every day – and create an innovative new model for accurately predicting large and potentially devastating earthquakes around the globe.

The Problem With Predictions

Throughout history, earthquakes have been sudden and terrifying events that can destroy entire communities and cultures, and people have sought ways to get advance warning of a coming quake from things like changes in animal behavior, or the shifting levels of well water. But even in more recent times, the US Geological Survey acknowledges that accurately predicting when and where an earthquake will strike is virtually impossible.

Today, earthquake preparedness models designed to mitigate the damage from a major quake include early warning systems and long-range forecasting, but neither of these strategies can predict a quake so precisely that communities can take safety precautions in advance. Many of the technologies used to understand earthquake behavior were developed in the 1980s, and little has been done since then to advance their capabilities. These technologies rely on data collected from sensors on known fault lines, which monitor the movement of seismic waves beneath the surface of the earth.

That kind of monitoring makes it possible to develop early warning systems that track the movement of the different types of seismic waves rippling out from a fault line. These systems, such as the one used by San Francisco’s famous BART lines, can alert people in the vicinity to be ready for an imminent quake.

A deeper analysis of quake behavior on a particular fault can allow earthquake experts to make long-term forecasts of the likelihood of a quake over a period of years, such as the conclusion of Los Angeles-area forecasters that the San Andreas fault is “overdue” for a major event. Likewise, scientists can warn that a small temblor could be a foreshock, or precursor, to a larger quake.

But the difficulty with accurate earthquake predictions arises from a combination of technological limitations and the geological complexity of earthquakes. Scientists simply haven’t had the means to delve deeper into the behavior of earthquakes and extract the data needed to create precise prediction models.

Understanding How Earthquakes Work

Earthquakes originate deep underground, or undersea. They arise from the movement of the Earth’s seven great tectonic plates—rigid pieces of crust that slide across the slow-moving lava of the mantle below. That movement creates energy that builds along the friction lines, and when the energy is released, it creates seismic waves that cause surface shaking that can radiate from the epicenter onto smaller faults nearby.

These seismic waves come in two stages. First, the primary, or P waves pull and compress the earth as they move along the fault line. These waves are followed by shear, or S waves, that move more slowly and push the earth up and down or side to side, causing more destruction to structures on the surface.

Significant earthquakes make headlines, but millions of small quakes go unnoticed every day. And while most efforts to understand earthquake behavior depend on data collected during larger temblors, new research suggests that data from these “microquakes” can provide key insights that lead to more accurate models of earthquake prediction. And when large amounts of data from these and other earthquake events are processed using the sophisticated neural networks of advanced machine learning, the result is a new model for accurately predicting, not forecasting, significant earthquake events.

Adapting AI Technologies for Earthquake Prediction

Geophysicists and other experts depend on seismic data to understand how earthquakes work, but interpreting the masses of data collected from sensors around the world can take weeks, months, or longer. Now, scientists at Stanford University are turning to artificial intelligence to create a new lab-tested model for earthquake prediction, using technologies derived from everyday AI-powered applications such as Alexa or Siri.

This new approach to understanding earthquake behavior arose from the efforts of Stanford researcher Moustafa Mousavi, who was working to create a novel way to automate earthquake detection and analyze the masses of data on seismic waves and other earthquake phenomena collected over time. By 2017, he was exploring ways to automate the process with the algorithms of machine learning.

Machine learning is the process in which algorithm-powered neural networks learn to make decisions on their own through training to recognize patterns in data. Once trained, these networks can process large amounts of data far faster than humans can, and they can recognize patterns too subtle for humans to see.

This kind of technology is used to create voice recognition applications like Alexa, which learn with repeated practice to separate and identify the sequences of sound patterns that create words and sentences, and to extrapolate from those sequences to recognize variations in tone and pronunciation.

From Sound Waves to Seismic Waves: Earthquake Transformer

Mousavi and his colleagues speculated that networks capable of recognizing sound wave patterns could also be trained to recognize the seismic waves generated by earthquake activity, even by very small tremors that might be missed by conventional sensors. Collecting and analyzing these weak signals could reveal previously unrecognized patterns that might point to an impending earthquake.

The resulting model, called Earthquake Transformer, can simulate the way humans look at data – taking a general survey and then focusing on specific areas of interest. And because Earthquake Transformer and similar networks can extract patterns from decades of collected data, it can offer new insights that fill in gaps in knowledge about the behavior of major quakes such as California’s powerful Loma Prieta event and extrapolate those findings to future events.

Earthquake Transformer was lab-tested using sets of unique training data that included a million hand-labeled seismograms collected over a span of two decades, excluding earthquake-prone Japan. Then, the model was tested on new data collected from a magnitude 6.6 quake recorded in Japan in 2000. Working with this previously unseen body of data, the model detected 21,092 seismic events. That was more than double the number detected by hand, and included mini-quakes previously undetected by human observers.

The technologies used to create and train Earthquake Transformer have also powered similar models at Los Alamos National Laboratory, where geologist Paul Johnson and other researchers have tested algorithms that, in laboratory settings, can accurately predict the occurrence of a specific kind of earthquake called a slow slip quake within a few days.

AI-powered earthquake detection is still in its infancy, and seismologists and other experts stress that the complex dynamics that produce earthquakes will never be entirely predictable. But with advanced neural networks capable of processing data from historic quakes and from the constant movement of microquakes, scientists are gaining new insights about the way earthquakes work—and moving closer to confidently predicting when and where a major quake might strike.

#earthquakeprediction #seismology #earthquaketransformer #artificialintelligence #AI #neuralnetworktraining #smartcities