Author: Li Baozhu
Editors: Li Weidong, xixi, Sanyang
The occurrence of earthquakes involves many variables, and "prediction" is challenging. However, significant progress has been made in predicting the number and intensity of aftershocks.
At 23:59 on December 18, 2023, a magnitude 6.2 earthquake occurred in Jishishan County, Linxia Prefecture, Gansu Province, with a focal depth of 10 kilometers. As of 06:00 on the 19th, 275 aftershocks had been detected locally. Coincidentally, a 7.6-magnitude earthquake occurred in the Noto Peninsula in Ishikawa Prefecture, Japan on the afternoon of January 1, 2024. As of 6:00 on the 2nd local time, 129 aftershocks with seismic intensity exceeding 2 degrees had occurred.
(A 2 on the Japan Earthquake Intensity Scale means that many people in quiet buildings can feel the shaking.)
Although it is not as severe as the main shock that caught people off guard, the power of the aftershocks cannot be underestimated. On the one hand, this is because the magnitude of the aftershocks of strong earthquakes is often not small; on the other hand, the secondary damage caused by aftershocks may further destroy already extremely fragile buildings, leading to the collapse of a larger area.
In addition, frequent aftershocks may cause the mountain structure to become unstable. In the event of precipitation, it may also cause secondary geological disasters such as landslides and debris flows. However, aftershocks are also often unpredictable because it is difficult to determine how remotely the main shock may have triggered fault zones in other areas in the short time following an earthquake.
In fact, people have never stopped exploring how to achieve earthquake prediction, especially the data processing and reasoning capabilities demonstrated by AI, big data and other technologies, which provide more flexible problem-solving ideas and have made gratifying progress. Nature reporter Alexandra Witze previously published a report in Nature, introducing the potential of machine learning models in predicting earthquake aftershocks and their intensity.
Original link:
https://www.nature.com/articles/d41586-023-02934-6
Forecasting model innovation powered by neural networks
Disasters often happen in the blink of an eye, and it is difficult to fight them with human power, so people are more eager to predict when earthquakes will occur and evacuate dangerous areas in advance. Although earthquake prediction with specific time and location is still difficult to achieve, aftershock prediction is possible with the support of deep learning, which will undoubtedly help post-disaster evacuation to a large extent and reduce casualties.
Just like large language models need to be trained with millions of words, sentences, paragraphs, etc., training earthquake prediction models also requires a large amount of past earthquake data to predict the probability of aftershocks. However, researchers have found that it is not easy in practice to clearly predict the detection indicators required for rare large earthquakes. Over the past few years, seismologists have used machine learning to discover small earthquakes that have never been detected in past earthquake records, thereby enriching existing data and providing new material for a second round of machine learning analysis.
The prediction model currently used by the United States Geological Survey (USGS) predicts possible earthquakes based on the magnitude and location of past earthquakes. Currently, there are three papers that use prediction methods based on neural networks to better capture the complex patterns of earthquake occurrence.
First, geophysicist Kelian Dascher-Cousineau of the University of California, Berkeley, and colleagues tested their model on data from thousands of earthquakes that occurred in Southern California between 2008 and 2021. The model outperformed the standard model in predicting the number of earthquakes that would occur over a rolling two-week period. In addition, the model better captures the range of possible earthquake magnitudes, reducing the chance of unexpectedly large earthquakes.
Currently, the earthquake evolution simulation method widely used in the industry is the ETAS (epidemic-type aftershock sequence) model.
Specifically, the study tested how the properties of the mainshock, as well as the background (depth, plate boundary type, etc.) and source (radiated energy, source size, etc.), affect the number of aftershocks and incorporated the neural-temporal point process model into the processes) are introduced into the standard earthquake prediction framework.
Source: Kelian Dascher-Cousineau’s GitHub profile https://keliankaz.github.io/academic-profile/
Secondly, Samuel Stockman, an applied statistician at the University of Bristol in the UK, also developed a model based on neural point processes, which performed well when training seismic data from central Italy in 2016-2017. Moreover, when the researchers reduced the training concentration The machine learning model performed better when the magnitude of the earthquake was .
The research has been published in Earth's Future. Research shows that the neural point process has better prediction performance for low-magnitude data than the earthquake aftershock statistical model ETAS (Epidemic-Type Aftershock Sequence), and the training speed is fast.
Paper address:
https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2023EF003777
Finally, Yohai Bar-Sinai, a physicist at Tel Aviv University in Israel, led a team to develop the encoder-decoder-based model FERN (Forecasting Earthquake Rates with Neural networks). This model also performed well when testing 30 years of earthquake data in Japan. to the standard model.
Model architecture
As shown in the figure above, the model input is encoded by a neural network, generating a latent representation of the construction state, which is then passed to the decoder network. The advantage of this approach is that different data sources and schemas can be naturally incorporated and added to the model via data source-specific encoders. Furthermore, the same encoding state can be used as input to multiple prediction heads (decoders) for different prediction tasks. The research has been published in Nature.
Paper address:
https://www.nature.com/articles/s41598-023-38033-9
Intelligent earthquake - AI real-time earthquake monitoring system
People's expectation to "prevent disasters before they happen" is a huge driving force for the continuous upgrade of earthquake early warning systems. However, we must first be clear that what is currently in use and continuously iterated is earthquake early warning systems, not earthquake predictions.
Although there is only one word difference between the two, the technical difficulty and actual effects are completely different. Earthquake early warning refers to issuing an alarm from a few seconds to tens of seconds in advance before the seismic waves propagate to the fortified area after an earthquake occurs, so as to inform people to take emergency measures to reduce casualties; earthquake prediction refers to the prediction of earthquakes that have not yet occurred . earthquake events that may occur but are likely to occur.
Picture source: Weibo@chargehorn
From the makeup of the ground in different areas to the types of interactions between seismic plates and the way seismic waves travel through the Earth, there are so many variables involved in assessing earthquakes that people need to fully understand them all to make an accurate judgment. Therefore, "prediction" is not easy, but "early warning" is not.
To view "earthquake early warning" from the perspective of a model, it is first necessary to ensure the timely input of seismic data, secondly to quickly and accurately process the ongoing earthquake data, and then infer the rupture direction and speed of the fault, and finally to transmit it in real time through communication means. to the disaster area. This process is comparable to a race against death, and only a few seconds of escape time can be gained. Data shows that when an earthquake occurs, if the earthquake warning information can be received 3 seconds in advance, casualties will be reduced by 14%, and if the earthquake warning information can be received 10 seconds in advance, casualties will be reduced by 39%.
Currently, earthquake early warning systems have been deployed in many places around the world, but the time it takes to receive early warning information is mostly 3-10 minutes. Japan's REIS earthquake early warning system can calculate the location and magnitude of an earthquake 5 seconds after receiving a seismic wave signal, and estimate the source mechanism of an earthquake rupture in about 2 minutes; the US National Geological Survey's automatic early warning system takes 3-5 minutes Report earthquake information; in 2021, my country released the world's first artificial intelligence "real-time" earthquake monitoring system-Smart Earthquake.
This system was developed by Professor Zhang Jie's team at the University of Science and Technology of China in cooperation with the China Earthquake Administration. It can obtain the three elements of an earthquake within 1 second - time (moment of earthquake), space (location of earthquake source), intensity (magnitude of earthquake), and Obtain the earthquake focal mechanism, that is, fault rupture direction, velocity and other information, thereby realizing real-time and automatic detection and release of earthquake information.
Investigating its principle, "intelligent earthquake" is mainly based on deep learning capabilities, based on millions of earthquake data collected in the database, combined with seismological theory, to quickly process ongoing earthquake data.
What's more important is how to issue early warning information as soon as possible after detecting earthquake data.
During the Gansu earthquake, users in Xining, Chengdu and other places near Jishishan County received mobile phone earthquake warnings, ranging from 120 seconds to 240 seconds. Many netizens lamented the power of domestic mobile phones, but in fact, the credit More credit should be given to the China Earthquake Early Warning Network jointly built by the Chengdu High-tech Disaster Reduction Institute and the China Earthquake Administration for providing early warning information. Among them, the early warning network issued an early warning 12 seconds in advance to Linxia City, 56 kilometers away from the epicenter, and 29 seconds in advance to Lanzhou City, 110 kilometers away from the epicenter.
According to Wang Dun, director of the Chengdu High-tech Disaster Reduction Institute, the principle of earthquake early warning technology is that electromagnetic waves propagate much faster than seismic waves. Before the earthquake shear waves reach the warning target area, electromagnetic waves with faster propagation speed are used to issue early warnings to potentially affected areas. At present, domestic mobile phones such as Huawei and Xiaomi have been connected to the earthquake early warning function of the China Earthquake Early Warning Network. This is also the 80th time that the China Earthquake Early Warning Network has warned of devastating earthquakes.
Human beings are so insignificant in the face of natural disasters, but with the continuous upgrading of technologies such as AI, big data, and the Internet of Things, we are also arming ourselves with technology and becoming increasingly powerful. With the accelerated iteration of emerging technologies, people will continue to optimize earthquake early warning systems and work towards the goal of "predicting earthquakes"!