This article comes from the public account Netease Smart (ID: smartman163), the original author Ashley Smart, compiled Netease Smart, and Ai Faner released it with permission.
Last May, after 13 months of sleep, the ground rumbled beneath the Puget Sound in Washington. The quake started more than 20 miles below the Olympic Mountains and drifted northwest for a few weeks to Vancouver, Canada. It then briefly reversed direction and moved back to the US border. Overall, the one-month “earthquake” released 6 magnitudes of energy. By the end of the earthquake, Vancouver’s southern tip had advanced about one centimeter into the Pacific Ocean.
However, because the earthquake is so widely distributed in time and space, it is likely that nobody will feel it.
This type of ghost earthquake occurs deeper than a regular fast earthquake and is called “slow slips.” They occur approximately once a year in the Pacific Northwest, and along the fault, the Juan de Fuca plate slips below the North American plate. Since 2003, the region’s vast network of seismic stations has detected more than a dozen slow earthquakes. These events have been the focus of geophysicist Paul Johnson’s earthquake predictions for the past year and a half.
Johnson’s team is one of the few groups using machine learning to try to unravel the mystery of seismic physics and sort out the earthquake warning signals. Two years ago, Johnson and his collaborators used something similar to image and speech recognition.The model discovery algorithm successfully predicted earthquakes in a model laboratory system.
Johnson and his team now report in a paper posted on the website arxiv.org that they have tested their algorithm on a slow earthquake in the Pacific Northwest. The paper has not been peer-reviewed, but outside experts say the results are encouraging. According to Johnson, The algorithm can predict a slow earthquake a few days before the beginning of it. Maesen de Hoop, a seismologist at Rice University, said, “I think this is the first time that we have indeed made progress towards earthquake prediction.
Mostafa Mousavi, a geophysicist at Stanford University, called the new results “interesting and inspiring.” But he also emphasized that machine learning has a long way to go before it can reliably predict catastrophic earthquakes. Although there are still many difficulties to be overcome in research, machine learning may become a new breakthrough in such an area where scientists have hardly seen hope for decades.
Disastrous and slow earthquakes
The late seismologist Charles Richter (named after his Richter scale) pointed out in 1977 that earthquake predictions can be “for amateurs, weirdoes, and outspoken counterfeiters. Provide a happy hunting ground. ” Geophysicist Panayiotis Varotsos of the University of Athens claims that he can detect “earthquakes” by measuring “seismic electrical signals”. Physicist Brian Brady of the U.S. Bureau of Mines issued false alarms in Peru in the early 1980s. These events are based on a fragile notion that rock bursts in underground mines are a clear sign that an earthquake is imminent.
Johnson knows this twisty history. The term “earthquake prediction” is taboo in many places. Six scientists were convicted of manslaughter in 2012 (conviction was later reversed) for downplaying the possibility of an earthquake near the central Italian town of L’Aquila, and the area was damaged by a magnitude 6.3 earthquake a few days later. Some well-known seismologists strongly claim that “earthquakes are unpredictable.” But Johnson believes that earthquakes are a physical process, and in this respect, earthquakes are no different from stars collapsing or wind direction changes. Although he emphasized that his main goal was to better understand fault physics, he did not shy away from prediction problems.
▲ (Picture: Paul Johnson, a geophysicist at Los Alamos National Laboratory, took a picture in 2008 with a piece of acrylic plastic, which his team used to simulate an earthquake in the laboratory One of the materials.)
More than a decade ago, Johnson began researching “laboratory earthquakes”, which consist of sliding blocks separated by a thin layer of granular material. They recorded the acoustic signals emitted during the stick-slip cycle and found sharp spikes before each glide. These precursor events are similar to the seismic waves generated by the pre-earthquake. But just as seismologists have struggled to translate foreshocks into predictions of when the main shock will occur, Johnson and his colleagues cannot figure out how to turn precursory events into reliable predictions of laboratory earthquakes. “We went into a dead end to some extent,” Johnson recalled. “I can’t find any way to proceed.”
A few years ago, at a conference in Los Alamos, Johnson described and explained his dilemma, and theorists suggested that he use machine learning to reanalyze the data. So the scientists worked out a plan: record about 5 minutes of audio in each experiment, and then cut it into many small pieces. For each segment, the researchers calculated more than 80 statistical characteristics, including the average signal, the change in the mean, and information about whether the segment contains precursor events. Since the researchers analyzed the data after the fact, they also knew how much time elapsed between each sound clip and subsequent laboratory failures. With this training data, they used a “random forest” machine learning algorithm to systematically find feature combinations that are closely related to the time remaining before the failure. After looking at the experimental data for a few minutes, the algorithm can begin to predict the failure time based only on the characteristics of the acoustic emission.
Johnson and his colleagues chose to use the random forest algorithm for prediction in part because it is relatively easy to interpret (compared to neural networks and other popular machine learning algorithms). The algorithm basically works like a decision tree, where each branch splits the dataset based on some statistical feature. Therefore, the tree keeps a record of the features of the algorithm used to make the predictions, as well as a record of each feature in helping the algorithm to derive the relative importance of these predictions.
▲ (Picture: In an experiment conducted at Los Alamos National Laboratory, polarizing lenses show the stress buildup of a model tectonic plate sliding laterally along a fault line.)
When researchers at the University of Los Alamos explored the inner workings of their algorithm, they were surprised by what they found: the statistical features that the algorithm relied on most for its predictions had nothing to do with precursor events before the laboratory earthquake. Instead, it’s about variance, which is a measure of how the signal fluctuates around the mean. In addition to the signal appearing just before the failure, the variance starts small and then gradually rises during earthquake preparations. As long as this change is known, the algorithm can make reasonable guesses. This finding is potentially significant. For decades, prospective earthquake forecasters have been studying foreshocks and other isolated earthquake events. And the results of Los Alamos show that everyone is looking to the wrong place-the key to prediction is the more subtle information broadcast during the relatively calm period between the two major earthquake events. To prove that machine learning can predict real earthquakes, Johnson needs to test real fractures. What better place than the Pacific Northwest?
Out of the laboratory
Most places where a magnitude 9 earthquake occurs on Earth are subduction zones, where one tectonic plate subducts under another. The subduction zone in eastern Japan caused the Tohoku earthquake and subsequent tsunami that destroyed Japan’s coastline in 2011. One day, the Cascadia subduction zone will also destroy Puget Sound, Vancouver, and the surrounding Pacific Northwest.
From Cape Mendocino, Northern California to Vancouver, the Cascadia subduction zone extends about 1,000 kilometers along the Pacific coastline. The last damage occurred in January 1700, it triggered a magnitude 9 earthquake and tsunami and reached the coast of Japan. Geological records indicate that such large earthquakes occur about every five thousand years in faults, more or less hundreds of years apart, which is one of the reasons seismologists pay so close attention to slow earthquakes in the region.
Slow earthquakes downstream of the fault in the subduction zone transfer a small amount of stress to the brittle crust above, and then a fast and catastrophic earthquake occurs there.
But for Johnson, the focusThere is another reason for slow earthquakes: they produce a lot of data.
For comparison, there have been no major catastrophic earthquakes in the fault zone between Puget Sound and Vancouver over the past 12 years. Over the same time span, the fault produced more than a dozen slow-slip zones, each of which was recorded in a detailed earthquake catalog. This seismic catalog corresponds to the real-world recordings from Johnson Lab’s seismic experiments. Just as with sound recordings, Johnson and colleagues sliced seismic data into small segments and characterized each segment with a set of statistical functions. They then provided the training data to the machine along with information on past slow earthquake times.
After training with data from 2007 to 2013, the algorithm was able to make predictions for slow earthquakes between 2013 and 2018 based on data recorded months before each event. The key feature is seismic energy, which is closely related to the variance of the acoustic signal in the laboratory. Like variance, seismic energy climbs in a characteristic way before each slow slide.
Cascadia’s predictions are not as accurate as laboratory earthquakes. Compared with the new results, the correlation coefficient between the degree of agreement between the characterization prediction and the observation is much lower than that in the laboratory research. Johnson said that despite this, the algorithm was able to predict one of five slow earthquakes that occurred between 2013 and 2018, and could almost accurately predict the start time.
Machine learning provides us with an entry point into the data search to find things we have never discovered or never seen before. But people still have a long way to go.
Profound facts
The goal of earthquake prediction has never been to predict slow earthquakes. Instead, it is intended to predict sudden and catastrophic earthquakes that pose a danger to life and limbs. For machine learning methods, this raises a paradox: the largest earthquakes, the earthquakes that seismologists most want to predict, and the rarest. How does a machine learning algorithm get enough training data to predict them?
The Los Alamos team bets that their algorithm does not actually require training in catastrophic earthquakes to predict them. Recent studies have shown that seismic images before minor earthquakes are statistically similar to those of major earthquakes, and that on any given day, dozens of minor earthquakes may occur on a fault. A computer trained on thousands of small earthquakes may be sufficiently versatile to predict large earthquakes. Machine learning algorithms may also be trained on computer-simulated fast earthquakes that could one day serve as a substitute for real data.
But even so, scientists will face a thought-provoking fact: Although the physical process of pushing faults to the edge of an earthquake may be predictable, most scientists believe that the actual triggering of an earthquake involves a random Sex. Assuming this is the case, no matter how well the machine is trainedHow good, they may never predict earthquakes as scientists predict other natural disasters.
In the best case, predictions of large earthquakes may have time limits of weeks, months, or years. Although such predictions may not be used to coordinate large-scale evacuation on the eve of the earthquake, it can increase the public’s preparation period, help government officials to modify unsafe buildings in a targeted manner, and otherwise mitigate the damage of catastrophic earthquakes.
Johnson thinks this is a worthwhile goal. But he knew it would take time. “I’m not saying to predict earthquakes in my lifetime, but no matter how much we explore, we are moving forward.”
Slow earthquakes downstream of the fault in the subduction zone transfer a small amount of stress to the brittle crust above, and then a fast and catastrophic earthquake occurs there.
But for Johnson, the focusThere is another reason for slow earthquakes: they produce a lot of data.
For comparison, there have been no major catastrophic earthquakes in the fault zone between Puget Sound and Vancouver over the past 12 years. Over the same time span, the fault produced more than a dozen slow-slip zones, each of which was recorded in a detailed earthquake catalog. This seismic catalog corresponds to the real-world recordings from Johnson Lab’s seismic experiments. Just as with sound recordings, Johnson and colleagues sliced seismic data into small segments and characterized each segment with a set of statistical functions. They then provided the training data to the machine along with information on past slow earthquake times.
After training with data from 2007 to 2013, the algorithm was able to make predictions for slow earthquakes between 2013 and 2018 based on data recorded months before each event. The key feature is seismic energy, which is closely related to the variance of the acoustic signal in the laboratory. Like variance, seismic energy climbs in a characteristic way before each slow slide.
Cascadia’s predictions are not as accurate as laboratory earthquakes. Compared with the new results, the correlation coefficient between the degree of agreement between the characterization prediction and the observation is much lower than that in the laboratory research. Johnson said that despite this, the algorithm was able to predict one of five slow earthquakes that occurred between 2013 and 2018, and could almost accurately predict the start time.
Machine learning provides us with an entry point into the data search to find things we have never discovered or never seen before. But people still have a long way to go.
Profound facts
The goal of earthquake prediction has never been to predict slow earthquakes. Instead, it is intended to predict sudden and catastrophic earthquakes that pose a danger to life and limbs. For machine learning methods, this raises a paradox: the largest earthquakes, the earthquakes that seismologists most want to predict, and the rarest. How does a machine learning algorithm get enough training data to predict them?
The Los Alamos team bets that their algorithm does not actually require training in catastrophic earthquakes to predict them. Recent studies have shown that seismic images before minor earthquakes are statistically similar to those of major earthquakes, and that on any given day, dozens of minor earthquakes may occur on a fault. A computer trained on thousands of small earthquakes may be sufficiently versatile to predict large earthquakes. Machine learning algorithms may also be trained on computer-simulated fast earthquakes that could one day serve as a substitute for real data.
But even so, scientists will face a thought-provoking fact: Although the physical process of pushing faults to the edge of an earthquake may be predictable, most scientists believe that the actual triggering of an earthquake involves a random Sex. Assuming this is the case, no matter how well the machine is trainedHow good, they may never predict earthquakes as scientists predict other natural disasters.
In the best case, predictions of large earthquakes may have time limits of weeks, months, or years. Although such predictions may not be used to coordinate large-scale evacuation on the eve of the earthquake, it can increase the public’s preparation period, help government officials to modify unsafe buildings in a targeted manner, and otherwise mitigate the damage of catastrophic earthquakes.
Johnson thinks this is a worthwhile goal. But he knew it would take time. “I’m not saying to predict earthquakes in my lifetime, but no matter how much we explore, we are moving forward.”