The Hualien Earthquake Was Reported 20 Seconds in Advance. Is Accurate Earthquake Prediction Not Far Away?

Using machine learning technologies such as deep learning and neural networks to analyze and study earthquake issues can reveal the value of some data that people cannot see, and achieve higher accuracy in the prediction of aftershocks and micro-earthquakes.

At 13:01 on April 18, a 6.7 magnitude earthquake suddenly struck Hualien, Taiwan. This was the largest earthquake in Taiwan in the past 20 years, and many areas near the Taiwan Strait felt the tremors. Fortunately, there were only a few accidental injuries in this earthquake.

According to media reports, New Taipei City received a warning 20 seconds before the earthquake was felt. Former New Taipei City "Mayor" Zhu Lilun was being interviewed at the time, and the interview footage also recorded the moment he received the warning.

The message received in advance in the video is called earthquake early warning. It means sending an alert to areas far from the epicenter in advance after an earthquake occurs, usually tens of seconds in advance. Although this technology is mature, it can only help areas on the edge of the earthquake zone, while the epicenter area can only leave it to fate.

Unlike earthquake early warning, earthquake prediction is to accurately predict the time, location and level of an earthquake before it occurs, so that countermeasures can be arranged in advance. However, due to the complex causes of earthquakes and scarce data, we still cannot accurately predict earthquakes to this day.

However, it is gratifying that although the problem of earthquake prediction has not been solved, in recent years, scientists have begun to try to use machine learning technologies such as deep learning and neural networks to analyze and study earthquake problems, and have achieved good prediction results in forecasts such as aftershocks and microearthquakes.

Harvard and Google team up to use machine learning to predict aftershocks

Earthquakes are usually not isolated events. After the "main shock" (usually a news headline-level event), there is often a series of "aftershocks". These aftershocks are numerous, and large aftershocks can cause serious repeated injuries. A typical example is the 5.12 Wenchuan earthquake, which was followed by tens of thousands of aftershocks, posing a great threat to rescue work.

Therefore, the detection of aftershocks is also an important part of earthquake prediction. Under traditional methods, there are some empirical judgment rules and methods for the time and level of aftershocks, but they usually cannot accurately predict the location and require cumbersome procedures in operation.

Harvard University collaborated with machine learning experts from Google to try to use deep learning to predict the location of aftershocks. Their research made a breakthrough, and the final results were published in Nature in August 2018.

Their database contains information about at least 199 major earthquakes that occurred around the world. Based on this database, they applied a neural network model to analyze the relationship between the static stress changes caused by the main shock and aftershock locations. The algorithm can identify useful patterns from the data information.

They eventually obtained the optimal aftershock location prediction model. Although the system still needs to be improved, it means that this is a step forward in this direction.

The research also had an unexpected benefit: It helped the team identify the physical quantities involved in earthquakes, which is important for earthquake research. When neural networks are applied to data sets, they can gain insight into the specific combination of factors that are critical to prediction, rather than just treating the predictions as numerical values.

Meade, one of the team members, once explained: "Traditional seismologists are more like pathologists. They study what happens after catastrophic earthquake events. We don't want to do that. We want to be epidemiologists. We want to understand the triggers and causes of these events."

It is expected that in the future, machine learning can uncover the mysteries behind earthquakes and reduce the damage they cause.

Learning to predict earthquakes from 550,000 samples

Based on the AI models proposed by Harvard and Google, researchers at Stanford University also created an artificial intelligence model that focuses on detecting and predicting microearthquakes, and ultimately achieved a high accuracy rate.

Microearthquakes, or low-intensity earthquakes, are those with an instantaneous magnitude of 2.0 or less. Such earthquakes are less destructive, but they are sometimes missed by earthquake monitoring systems due to factors such as background noise, small events and false alarms.

The AI model built by Stanford University, called Cnn-Rnn Earthquake Detector (CRED), can accurately detect information about microearthquakes through continuously recorded historical data.

The system consists of two types of neural network layers:Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). CNN extracts features from earthquake sensors, while RNN can combine memory and input data to improve the accuracy of its predictions and learn sequence characteristics similar to seismographs.

These two constitute aResidual learning framework,This is done to alleviate the problems of overfitting and other problems that occur in multi-layer neural networks. In this way, neural networks can maintain their accuracy while learning more detailed features from the dataset. In addition, it will be easier to optimize.

To train and validate the earthquake detection AI system, the researchers collected continuous recorded data from Guy-Greenbrier, Arkansas, in 2011, which contained 3,788 events, as well as 550,000 30-second earthquake maps from 889 monitoring stations in Northern California, including three indicators.

50,000 samples out of 550,000 data points were used to evaluate performance. The results showed that the network model was able to accurately identify earthquake signals regardless of the magnitude of the earthquake, whether it occurred locally, or whether there was strong background noise. More importantly, AI only needed partial records to detect earthquakes.

When fed continuous data from the Guy-Greenbrier dataset, the model, which took nearly an hour to train on a computer, detected 1,102 micro- and large earthquakes caused by hydraulic fracturing, wastewater injection, and tectonic plate movement, including 77 that had never been recorded before.

The report said that in all tests, the learning model showed "superior" performance compared to two widely deployed earthquake systems because the computer model can analyze some data values that people cannot see.

also,The scalability of the model has also been improved.“Once trained, the model can be applied in real time to seismic data streams,” they wrote. “The seismic signals, based on their spectral structure, are modeled with high resolution and a low rate of false positives.”

The team believes that the machine learning model could be easily scaled up to multiple sensors, allowing for real-time monitoring in tectonically active areas and could also serve as the basis for an early earthquake warning system.

If the judgment of small earthquakes is accurate enough, then it will be of great significance for the prediction model to be used in the prediction of large earthquakes.

Earthquake prediction may be possible in the future

Machine learning techniques can be used to preserve simulated records of past earthquakes in large quantities. As the media that records this data gradually degrades, seismologists are racing to preserve this valuable information.

Some researchers are using machine learning algorithms to sift through seismic data to better identify earthquake aftershocks, volcanic seismicity and to monitor precursors to shaking that signal deformation at plate boundaries where huge earthquakes could occur.

Other researchers are using machine learning techniques to locate earthquake origins and distinguish small quakes from other seismic "noise" in the environment.

For a long time, some scholars believed that it was impossible to completely predict earthquakes. But from the current research results, perhaps predicting earthquakes is no longer "impossible". Through accurate prediction of aftershocks and micro-earthquakes, perhaps in the near future, the problem of predicting large earthquakes can be solved.

Natural disasters are uncontrollable. We can only hope that we can use the power of modern technology to prevent natural disasters from causing harm to anyone.

Look forward to the day when machine learning-based predictions help deploy emergency services and provide evacuation plans for areas at risk of aftershocks.