[ad_1]
This diagram shows how artificial intelligence improves tsunami prediction. This system stores the characteristics of past and simulated earthquakes and tsunamis. A network of devices monitors events. Based on the data, the algorithm predicts wave properties after just a few minutes of actual data. (contribution)
I know little about artificial intelligence. Media headlines may make it sound like AI is the greatest threat to human society or the solution to many problems. I thought it would be worth doing a little digging to see how it works in an area that I have some knowledge of.
If you’re like me, you know that AI has something to do with algorithms and can predict outcomes, but you probably don’t know much more than that. We delved a little deeper into the basics of AI to better understand how it is applied to earthquake detection, early warning, forecasting, and tsunami prediction.
Computers and algorithms have been part of seismology for more than half a century. In the late ’60s, early in graduate school, I would carry boxes of punched cards to a computer center, only to retrieve them hours later only to discover fatal errors that prevented the execution from completing. there was. Mine was the only relevant information, but it was often missing.
A key advancement in AI is the ability of programs to learn and adapt without human intervention. To do this, we need a huge amount of information about the task at hand, the machine learning part of the process. Traditional programming, like the one I used to do with a stack of punched cards, is similar to cooking. Carefully define what to include in your recipe and give detailed instructions on how the computer should process the data you provide.
Machine learning (ML) is the “training” part of the AI process, loading vast amounts of data into the system and identifying any number of patterns. Recognizing spam calls is an example that everyone with a smartphone uses on a daily basis. The dataset is phone numbers and past behavior. Some numbers can be identified as belonging to marketing companies. Others are identified from message content or user behavior. Every time you or someone else marks a message as spam, the message goes into a larger database and is flagged as potential spam the next time it is used. ML allows systems to learn new things as they get more data.
An AI step is an algorithm that tells the computer what to do with all the data it gets. For spam calls and emails, the process of notifying or sending a spam call to a spam folder is relatively simple. It’s not a perfect system. I skim through my spam folder every week or so to see if there’s anything important in there, and there are very few errors.
There are many other AI/ML applications that are already part of daily life. Banks and financial institutions are using AI to identify fraudulent transactions, and healthcare systems are incorporating AI into diagnostic and treatment options. If you use social media, the ads that pop up in your feed are based on what you click on or hide. I love crosswords, and searches for hip-hop stars or unusual clothing can sometimes result in a spike in unusual ad posts.
AI can be used to better understand what actually happened, predict what is most likely to happen next, or suggest what options are available in decision-making. It can be used as a writing tool for The accuracy and usefulness of any application is based on its data, and quantity and quality are key.
Turning to seismology, a quick look at the GEOBASE reference index reveals 400 citations involving AI or ML over the past five years. These are useful for identifying small earthquakes and aftershocks in noisy environments, for relationships between source and rupture characteristics, for predicting ground motion, and for identifying previously unnoticed signals that can provide insight into source dynamics. It covers a wide range of applications, including specific
The hottest topic in both media and research is the use of AI/ML as an earthquake prediction tool. Predicting the location and magnitude of earthquakes days or weeks in advance has been the holy grail of seismology for more than a century. A 1913 Scientific American article proclaimed, “Earthquake prediction is just around the corner.” That corner became a very large corner.
The USGS ShakeAlert system successfully predicts felt earthquakes in the seconds after rupture begins and before the strongest shaking arrives. This is a major advance, allowing critical systems to be powered down, people to fall into, take cover for, or grab onto. However, there is currently no reliable way to issue accurate alerts over longer time frames.
The idea behind using AI/ML is to identify subtle patterns that can lead to predictions with longer lead times. Many different approaches have been used, some taking a laboratory approach, others using GPS data of sensitive land-level changes, electromagnetic signals, seismic patterns, or slow-slip events. .
China is a world leader in AI-based earthquake prediction. The international competition, hosted by China, attracted 600 applicants from all over the world. A team from the University of Texas took first place in a seven-month test in China, accurately predicting 70% of earthquakes. They used historical earthquake data and developed pattern recognition algorithms to make predictions. The Texas team accurately estimated the magnitude and predicted location of 14 moderate earthquakes within 200 miles of their actual epicenters.
While the feat garnered praise and media attention, it was by no means ready for prime time. In the same time frame, it issued eight false warnings and missed one earthquake entirely. Predictions are only possible with higher levels of accuracy.
My interest in AI/ML peaked last week after listening to a talk on its application to tsunami prediction. Chris Moore at NOAA is working on using the first few minutes of a tsunami wave train with deep-sea instruments to improve estimates of tsunami wave height. Tsunamis are a problem because real-time and spatial distribution is very limited. His team enriched the real data by creating simulated sources from across the Pacific Ocean.
By training the program on thousands of simulated earthquakes, the group was able to obtain highly accurate predictions of entire wave trains by looking at just 12 minutes of real data. This means that the next time a major earthquake occurs somewhere in the Pacific Ocean, we will not only know when the first earthquake arrived, but also the arrival times of the second, third, fourth, and subsequent earthquakes, and which ones. Knowing what is likely to occur can affect us. maximum.
No matter how rosy the future of AI in earth sciences is, prolonged shaking of the ground will still serve as a warning to get out of the tsunami danger zone.
Note: For a more detailed overview of AI and machine learning, visit https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained.
Lori Dengler is professor emeritus of geology at Cal Poly Humboldt and an expert on tsunami and earthquake hazards. Have questions or comments about this column, or would you like a free copy of the prep magazine Living on Shaky Ground? Leave a message at 707-826-6019 or email Kamome@humboldt.edu.
[ad_2]
Source link
