Periodic Reporting for period 2 - TECTONIC (The physics of Earthquake faulting: learning from laboratory earthquake prediCTiON to Improve forecasts of the spectrum of tectoniC failure modes: TECTONIC)
Berichtszeitraum: 2021-07-01 bis 2022-12-31
Machine learning techniques have been used extensively in seismology and earthquake physics. While the main applications have been in earthquake identification and seismic catalog development, there have been also many studies focused on learning how to improve our ability to forecast and post-predict earthquake occurrence. The latter has been driven in particular by the discovery that lab earthquakes can be predicted from the acoustic emissions AE (microlabquakes) that emanate from lab fault zones.
Our recent works conducted for TECTONIC show that the timing and magnitude of labquakes can be predicted from AE and also from acoustic interrogation of the lab faults zones. The latter involve active source acoustic measurements. We show that systematic changes in fault zone elastic properties can be used to predict labquakes. Our recent works also show that lab foreshocks and changes in fault zone elasticity can predict the state of stress on the lab fault zone.
From our initial work, two key questions have emerged. 1) What are the mechanisms that encode the patterns in lab foreshock activity to allow prediction of labquakes? Labquake prediction is possible for a wide range of AE frequency magnitude characteristics, so changes in b-value with time to failure is unlikely to be the prediction mechanism. Are there more complex patterns in AE energy release and possibly corresponding geometric structures within or along the lab fault zones? 2) How can the lab-based work be applied to tectonic faults? Is it possible to apply transfer learning or other techniques that would allow machine learning to forecast earthquakes and/or otherwise improve our understanding of earthquake physics?
We are working actively on these questions in addition to many other activities for TECTONIC. For Question 2, our recent work shows that deep learning (DL) methods can distinguish foreshocks from aftershocks of the 2016 M6.5 Norcia earthquake of Italy. We study events that occurred close to the Norcia mainshock and use very simple definitions of foreshocks and aftershocks; that is events that occur before and after the mainshock. The DL methods are being trained with ground truth results and tested on other data. We are using sections of full waveforms from local seismometers to train and test the model.
Our work is important for society in many ways. One of these is the simple value of encouraging creative thought. How and why do the natural phenomena around us work? Why does Earth's surface have the form that we see and what are the main processes that dictate the shape of landforms? What factors dictate the behavior of natural disasters such as earthquakes? Why, where and when do earthquakes occur? While we are still far from being able to forecast or predict the timing of earthquakes, we are closer than ever before. Recent advances have provided solutions that are at least 50 years old and have provided a framework for improving our ability to assess earthquake hazard. There are still many bridges to cross but we are moving toward solutions that will impact society and improve the security of our homes and workplaces.
Our recent works conducted for TECTONIC show that the timing and magnitude of labquakes can be predicted from AE and also from acoustic interrogation of the lab faults zones. The latter involve active source acoustic measurements. We show that systematic changes in fault zone elastic properties can be used to predict labquakes. Our recent works also show that lab foreshocks and changes in fault zone elasticity can predict the state of stress on the lab fault zone.
From our initial work, two key questions have emerged. 1) What are the mechanisms that encode the patterns in lab foreshock activity to allow prediction of labquakes? Labquake prediction is possible for a wide range of AE frequency magnitude characteristics, so changes in b-value with time to failure is unlikely to be the prediction mechanism. Are there more complex patterns in AE energy release and possibly corresponding geometric structures within or along the lab fault zones? 2) How can the lab-based work be applied to tectonic faults? Is it possible to apply transfer learning or other techniques that would allow machine learning to forecast earthquakes and/or otherwise improve our understanding of earthquake physics?
We are working actively on these questions in addition to many other activities for TECTONIC. For Question 2, our recent work shows that deep learning (DL) methods can distinguish foreshocks from aftershocks of the 2016 M6.5 Norcia earthquake of Italy. We study events that occurred close to the Norcia mainshock and use very simple definitions of foreshocks and aftershocks; that is events that occur before and after the mainshock. The DL methods are being trained with ground truth results and tested on other data. We are using sections of full waveforms from local seismometers to train and test the model.
Our work is important for society in many ways. One of these is the simple value of encouraging creative thought. How and why do the natural phenomena around us work? Why does Earth's surface have the form that we see and what are the main processes that dictate the shape of landforms? What factors dictate the behavior of natural disasters such as earthquakes? Why, where and when do earthquakes occur? While we are still far from being able to forecast or predict the timing of earthquakes, we are closer than ever before. Recent advances have provided solutions that are at least 50 years old and have provided a framework for improving our ability to assess earthquake hazard. There are still many bridges to cross but we are moving toward solutions that will impact society and improve the security of our homes and workplaces.
We have been developing and building the lab data set, constructing equipment to conduct new lab experiments at the Host Institution, working with seismic data and field observations from tectonic faults, and working to collect new geodetic data on the spectrum of fault slip behaviors on tectonics faults. We have published several papers, given talks at multiple scientific meetings, planned and hosted scientific meetings, and presented our project plans and results to the public. We are also: 1) training students and postdocs who will be the next generation of researchers in earthquake science and 2) leading a broad community collaboration in relation to the objectives of TECTONIC via a weekly, virtual seminar series that has been attended by more than 1000 participants
We are developing new methods to use machine learning to identify precursors to lab earthquakes and to predict the timing and magnitude of lab earthquakes as well as the fault zone stress state. One recent application is to map the distribution of a set of measurements onto another, such that for example a machine learning model trained on one dataset works on another. We want first to use this mapping from experiment to experiment, and then from experiment to field data. By the end of the project we expect to have developed methods to test the extent to which lab-based methods for prediction and forecasting can be applied to field data.