You are using an old browser with security vulnerabilities and can not use the features of this website.
Geoscientists have a problem: The phenomena they research are so complex that it is difficult to track them down. New sensor technology, satellite surveillance, and computer models of the processes in the earth’s interior bring them closer to these phenomena but create another obstacle: Even the most modern high-performance computers take a very long time to calculate the complex models. Artificial intelligence (AI) may provide a remedy. Dr. Hannes Vasyura-Bathke is developing a system that will take only a few second to complete modeling that normally keeps a computer busy for weeks - and will even do a more accurate job.
Hannes Vasyura-Bathke is a geoscientist. His research focuses on the physics of earthquakes and volcanoes. This is a field of research that has significantly benefited from technological developments in recent years. Computer programs are used to model processes and phenomena that cannot be detected directly because they sometimes happen several kilometers deep into the earth's crust. Meanwhile, the researchers are able to draw initial conclusions about what happens in the earth’s interior when there is a quake on the surface. Where do rock masses break? Where and how far do they move? The extent, displacement, and magnitude of seismic events can be simulated on the basis of measured data. Despite the new opportunities that big data and computer models have opened up, geo-researchers are already bumping up against new limits. And they are technical in nature: mathematical modeling on the computer, i.e. the search for seismic source parameters and their inaccuracies, which explain the measurement data similarly well, is complex and can sometimes takes weeks because millions of models with slightly different parameters have to be calculated over and over again. “This is a bottleneck hampering our progress,” says Vasyura-Bathke. “Actually, you do not want to do the same calculation over and over again; you want to get straight to the solution.”
Vasyura-Bathke is developing an AI system that performs the modeling and gets smarter with each run. “We do not calculate other solutions. We model the same physics but faster,” he says. If the method is successful, a few seconds might be enough to create a model from new data. The “system” is an artificial neural network. “Imagine it as a small brain,” explains Vasyura-Bathke. The system’s special feature is that the AI does not save the millions of calculations it performed to retrieve them when needed. Instead, it learns to recognize patterns in the data. “To identify a banana, we do not need 30 pictures of a banana, but the banana’s characteristic features.” It is exactly these criteria that the network should identify and remember - and that it has to learn.
Technically, the AI consists of mathematical algorithms, filters that are being applied to measured data. Only when there are many filters stacked behind one another do they form a neural network. It works in a way similar to the face recognition tool on Facebook or the tool to unlock smartphones, only with satellite images and seismic wave recordings. The AI system has to be trained to learn the patterns. "For the methods currently in use, the measurement data is entered into a program and then we check to see if the calculated model fits the measurements. To train the AI, we turn this process around,” he explains. To begin with, he “feeds” the system with validated data, i.e. pairs of measured data and calculated models that are known to be correct.
“We actually have only a few hundred of these.” Many more are needed. Therefore, they simulate the data and add a “noise” – sources of error, deviations that always occur and affect the measurements. “Initially, the system is often wrong. But then you tell it how much it is off, and it corrects the filters.” The AI calculates millions of slightly varied models based on the known data pairs and gradually refines the filters guided by its coach. This process takes a long time but will ultimately save a lot of time because when successfully trained, the AI can match the data to the physical quantities of an earthquake. In the end, Vasyura-Bathke is not training a room-filling supercomputer, but a computer program that is barely more than a few megabytes.
Vasyura-Bathke came up with the idea while studying. “Nowadays you have to program everything yourself anyway. If you look around the relevant forums, sooner or later you will inevitably come across ‘machine learning’.” Only now is the technical development mature enough for such a project, he explains. “Machine learning has made significant progress in recent years, for example in terms of how the AI is able to remember patterns.” The project breaks new ground, and not only for Vasyura-Bathke. It is difficult because it requires expertise from a wide range of subject areas. Ideally, one would be a geodesist, geophysicist, and computer scientist all in one. It took Vasyura-Bathke years to work through all these fields, he says. At the same time, he found competent partners: at the University of Potsdam, the Machine Learning Group headed by Prof. Dr. Tobias Scheffer at the Institute of Computer Science and the geophysicist Dr. Matthias Ohrnberger from the Institute of Earth and Environmental Sciences, as well as the working group for seismology led by Prof. Torsten Dahm at the GFZ German Research Center for Geosciences. They are all helping him to develop and train his AI system.
Like every student, the AI also has to overcome learning difficulties. Currently it is ‘struggling’ with the complexity of the task. “The network has difficulties with concurrent learning parameters with different mathematical units,” says Vasyura-Bathke. While the orientation of an earthquake surface in space is measured in degrees, displacements of rock masses are quantified in meters. Combining these units is a challenge. The researchers do not know yet how many modeling algorithms can be combined in an AI system. “We’re the first to try something like this - the combination of geoscience and machine learning is just in its infancy. It is entirely possible that different networks will be developed for different problems. But it might also be conceivable that in the end one network will do everything.”
Together with the researchers Marius Kriegerowski from the University of Potsdam and Gesa Petersen from the German Research Centre for Geosciences, Vasyura-Bathke has taught the AI system to find a location based on the three spatial coordinates - latitude, longitude, and depth. In its first test, it had to evaluate data for the localization of earthquakes in the Vogtland region, where seismic waves have been recorded over long periods at various measuring stations. Their analysis allows conclusions about where and when earthquakes occur. So far, the localization of earthquakes still requires a lot of “manual work” and uses standard methods, which allow a very accurate analysis, but also involve a considerable amount of work, according to Vasyura-Bathke. “Our neural network is able to accomplish the task with similar accuracy but about 100 to 1000 times faster than comparable methods, which enables a sustainable and holistic data analysis even with increasingly larger seismological data volumes!” Vasyura-Bathke is now sending another network ‘back to school’ because every scientific question requires the training of a new network. The researcher estimates, however, that it will probably take years until one of them is ready for automated, operational use. But then they should be able to model more complex phenomena - and continue learning.
Artificial Intelligence support for rapid analysis of earthquakes and volcanic activity
Funding: Geo.X The Research Network for Berlin and Brandenburg
Participants: University of Potsdam; Helmholtz Centre Potsdam – GFZ German Research Centre for Geosciences
Dr. Hannes Vasyura-Bathke studied geophysics und earned his doctorate in geophysics at the University of Potsdam in 2013. Since 2017, he has been research assistant at the Institute of Geosciences.
The work on localizing earthquakes by using convolutional neural networks was published in 2018:
Marius Kriegerowski, Gesa M. Petersen, Hannes Vasyura‐Bathke, Matthias Ohrnberger (2018); A Deep Convolutional Neural Network for Localization of Clustered Earthquakes Based on Multistation Full Waveforms. Seismological Research Letters ; 90 (2A): 510–516. doi: https://doi.org/10.1785/0220180320
Text: Matthias Zimmermann
Translation: Susanne Voigt
Published online by: Marieke Bäumer
Contact to the online editorial office: onlineredaktionuni-potsdamde