PhD position AI4theSciences – Artificial Intelligence for Seismic Hazard Monitoring with InSAR

PhD position AI4theSciences – Artificial Intelligence for Seismic Hazard Monitoring with InSAR

A PhD position is available to join the Department in the Laboratoire de Géologie, mixing machine learning and InSAR.

Context – Motivation

Earthquakes are some of the most dangerous hazard on Earth while also being the least understood and predictable. Active faults accommodate the relative motion of tectonic plates that pave the surface of our planet. As these plates are elastic to first order and in constant, steady-state, motion, relative displacement at plate boundaries leads to stress increase within the first kilometers of the crust. Such stress is released by slip on faults between plates, either gently and slowly over days to months to years by a seismic slip or rapidly, within seconds, by a rupture generating devastating seismic waves. These earthquakes may suddenly rupture fault sections over hundreds of kilometers with meters of relative displacements, resulting in strong ground motion with destructive potential.

Over the last 2 decades, the expansion of space geodetic measurement techniques led to the discovery of an entire spectrum of modes of fault slip. Faults can be either locked most of the time or slipping slowly over decades. Faults can also slip aseismically for weeks to months during what is now recognized as Slow Slip Events (SSEs). First, recent observations suggest these slow slip events exist at all spatial and temporal scales and bear striking resemblance with classic earthquakes only slower. These slow slip events release a significant fraction of the stress, hence affect budgets of potential earthquakes along faults, interact with earthquakes via triggering and arrest mechanisms and influence the nucleation, the propagation and the termination of earthquakes, hence their magnitudes. From a fundamental point of view, understanding these events would lead to a better understanding of earthquakes and faulting. Second, numerical models of dynamic faulting predict earthquakes start with a nucleation phase that may span several minutes to days. Recent but few seismological and geodetic observations suggest a precursory phase to earthquakes may last from days to months with characteristic behavior such as foreshock migrations and slow slip events. Very few of these observations are available and it is currently impossible to determine in advance whether a slow slip event captured by geodetic data is the precursory phase to a large devastating earthquake or a harmless slow transient. Being able to mine within geodetic data to document all forms of slow slip events, whether they lead to an earthquake or not, would open the door to a quantification of the dynamics, hence potentially the understanding, of these transient deformation phase precursory to large earthquakes.

6 years ago, the European Space Agency triggered the golden age of Synthetic Aperture Radar observation of the Earth, with the launch of the Sentinel 1 constellation associated with a fully open policy on data distribution. Nowadays these satellites image every plate boundary region with a repeat frequency of a few days, allowing us to build time series of ground deformation with a millimeter precision with unprecedented spatial resolution (InSAR; ~100 m pixel size over hundreds of km). In the years to come, several space agencies, including NASA, the Canadian space agency and the JAXA will launch several satellites with comparable operating modes and open-data policies, perpetuating our ability to monitor ground deformation globally in time. We have built tools in our team to ingest, process and analyze these movies of ground deformation within the scope of the Geo4D ERC StG project and are in a position to monitor any plate boundary in continental region with great accuracy. However, despite efforts from the entire community of the last 20 years, these measurements are still polluted with atmospheric artifacts of amplitude comparable to the signals of deformation we arelooking for, requiring visual inspection by trained experts of the time series in order to assess the source of the signals, tectonics or artifacts. We are currently missing a tool that will clean our series of data and detect any transient slow slip event in InSAR time series, globally. Such tool would allow to build a collection of all slow slip events globally with the ambition to identify key differences between harmless slow slip events and the precursory phase to devastating earthquakes. In addition,such data set would allow to experiment existing mechanical models toward a complete understanding of the dynamics of fault slip.

Scientific Objectives, Methodology & Expected results

We have developed a prototype neural network designed for the removal of atmospheric artifacts within time series of data to highlight slow slip events. Our neural network, purely convolutional, trains onsynthetic time series of noise and surface displacements to separate the tectonic signal from the noise. Since we do not have ground truth in this enterprise, except in a few identified regions of the world, we have to use a synthetic, theoretical, training set. Our neural network is able to detect and highlight events that have been previously identified by experts and reveals the intriguing dynamics of slow slip along transform plate boundarieson newly processed datasets.

However, several issues remain and intense development and testing is required in order to push forward our approach.

  1. Significant improvements are required in the core of our algorithm. We are currently facing scaling issues that prevent us from detecting other forms of ground deformation than fault slip that reaches the surface. Solving this identified issue will allow to detect large scale tectonic deformation signals related to slip at depth (which is not currently possible with our method) and the inflation/deflation sources such as volcanic or hydrological ones. We should be in a position, at the end of the PhD thesis, to extract, in addition to tectonic signals, the deformation signature of volcanoes or crustal reservoirs for hydrology.
  2. We need to improve noise modeling for the training phase and various noise structures should be explored. We will base this part of the project on a collaboration with the Willow team (INRIA/ENS – PSL, PI Prof.J. Ponce), which develops a range of learning-based methods for image restoration tasks such as deblurring, denoising, and super-resolution.
  3. We will compare our results with the detections from a network of GNSS sites measuring ground deformation at discrete points on the ground along the North Anatolian Fault, in Turkey. This main plate boundary fault, threatening the mega-city of Istanbul, slips aseismically episodically along its central portion and we have installed a network of GNSS sites that measures ground deformation continuously. Ideally, the predictions of our neural network should match our detections from GNSS data.
  4. A global and continuous hunt for slow slip events within the entire Sentinel 1 archive over actively deforming regions will be conducted toward the understanding of the dynamics of fault slip and, potentially, the identification of key observations precursory to large earthquakes.

The project should be feasible in 3 years. The first methodological developments, including the detection of large scale signals (slip at depth and deflation sources) and noise structure examination should be done over the first year of the PhD. Validation of the method, which includes learning some concepts about GNSS data processing and analysis, should be done over the next 6 months. The large scale implementation, detection and subsequent modeling of slow slip events globally should then take one year, leaving 6 months for the preparation of the final manuscript and for finalizing the last publications.

In terms of skills, the PhD candidate will be fluent in InSAR processing, interpretation of ground deformation data (including tectonics, volcanic or hydrological sources), some concepts about GNSS data processing (done in the lab by our engineers) and of course, all things related to the development of the algorithms. The candidate will benefit from a stimulating environment in a lab where experimental rock mechanics experts collaborate with experts in theoretical solid mechanics, experts in geodesy, tectonics and seismology with regularvisiting researchers from all around France, Europe and the world.

International mobility

This project is based on 2 international collaborations which will directly benefit the PhD candidate, including the collaboration with the Los Alamos National Lab, NM, USA and with the Istanbul Technological University and Kandili Observatory colleagues, Pr. Z. Çakir and Dr. S. Ergintav, who are key in the setup and maintenance of the GNSS network we intend to use for the validation of our results.

This PhD thesis will benefit from our starting collaboration with the Inria team Willow in Paris (PI. Pr. J. Ponce, INRIA/ENS – PSL). We plan to attend at least 2 international conferences (American Geophysical Union and European Geosciences Union) and multiple workshops.

Thesis supervision

Romain Jolivet and Bertrand Rouet-Leduc

Learn more and apply on EURAXESS