Research

My research focuses on simulating ground motions generated by earthquakes, monitoring and understanding the processes occurring in the shallow subsurface, and on improving passive seismology studies using deep learning.

1) Ground motion simulation/prediction using the ambient seismic field.

The ever-increasing number of large-scale structures, such as high-rise buildings, oil tanks, and suspension bridges, requires an accurate prediction of long-period ground motions to evaluate seismic hazard. The weak attenuation of long-period seismic waves combined with their amplification by sedimentary basins combined are a real threat for large-scale structures in metropolitan areas.

The ambient seismic field recorded by seismometers, which can be described as tiny vibrations of the ground mainly caused by ocean waves and human activities, can be used to retrieve the response of the Earth, also called impulse response function (IRF), between the two sensors with seismic interferometry. Several empirical studies demonstrated that the amplitude and the phase information of the IRF can be preserved. By regarding one station located close to an earthquake epicenter as a virtual source and another one as the receiver, we can simulate the long-period ground motions of the earthquake.

This technique has been used to simulate long-period ground motions (longer than 1 s) generated by moderate (M_w 5-6) earthquakes, where the point source hypothesis can be made, which occurred in different geological contexts. Viens et al. (2015) demonstrated the feasibility of simulating ground motions of a moderate M_w 5.0 offshore subduction earthquake that occurred along the Nankai Trough (Fig. 1). I also applied this method to a very dense seismic network in the Tokyo metropolitan area to investigate the response of the Kanto sedimentary basin (Viens et al., 2016a). Results show that the long-period amplification caused by the basin can be retrieved from the ambient seismic noise (Fig. 2). As most of buildings are sensitive to high-frequency waves (shorter than 1 s), I coupled the long-period simulations to high-frequency stochastic modeling in order to simulate broadband waveforms that agree well with those of moderate crustal earthquakes in Japan (Viens et al., 2014).

For large earthquakes, however, the point source assumption used to simulate moderate earthquakes does not hold and finite rupture modeling needs to be taken into account. I developed a technique to simulate the long-period ground motions from large earthquakes and successfully applied it to past onshore (Viens et al., 2016b) and offshore events (Viens and Denolle, 2019).

Fig. 1: Simulation of long-period ground motions at stations in the Nobi basin and surrounding area (triangles) for a Mw 5.0 earthquake (star). The color of the triangles on the map (a) indicates the percentage difference between the PGVs of the extracted ambient noise impulse response functions (ANIRFs) and the earthquake records computed using equation 2. (b) Calibrated ambient noise impulse responses (red lines) and earthquake records (black lines), bandpass filtered between 5 and 8 Hz in velocity for vertical components, are shown as a function of the distance from the virtual source (TK4OBS). The blue dashed line represents the wave propagation speed of 2.7 km/s and correlation coefficients (CC) are given in parentheses (From Viens et al., 2015).

Fig. 2: 5% damped pseudo-velocity response at a period of 6 s computed from (left) records of a Mw 5.8 earthquake (red star) and (right) simulated waveforms from the ambient seismic noise in the Kanto Basin, Japan. The TENNOD station (blue triangle), which is located near the earthquake epicenter, is the virtual source and black triangles are the receiver stations. Red and blue indicate a strong and weak wave amplifcation, respectively (modified from Viens et al., 2016a).

Fig. 3: (top row) Distribution of the ratio between simulated and observed long-period PGVs of the 2008 Mw 6.9 Iwate-Miyagi Nairiku earthquake (PGVsim∕PGVobs) for (a) vertical, (b) radial, and (c) transverse components. Simulated PGVs are from the waveforms computed using the uniform source model and a rupture velocity of 1.95 km/s. Green triangles indicate that the ratio is between 0.5 and 2; blue triangles are used for ratios between 0.33 and 0.5, and between 2 and 3; and red triangles illustrate ratios smaller than 0.33 or larger than 3. (middle row) Residuals (log10(PGVobs∕PGVsim)) versus distance from the fault plane. The colors of the circles correspond to those of Figure 11 (top). The mean of the data is indicated by the solid line and the dashed and dash-dotted lines denote 1 and 2 standard deviations (𝜎), respectively (from Viens et al., 2016b).

2) Non-linear site response induced by strong ground motion from Megathrust earthquakes.

Soft shallow layers such as sedimentary basins can significantly amplify incoming seismic waves. For weak ground motions, soft soils behave as an elastic material and the amplification of ground motions is linear. However, strong ground shaking can induce a non-linear response of the soils that needs to be accurately characterized to mitigate seismic risk.

I estimated the non-linear response of shallow soils during the 2011 M_w 9.0 Tohoku-Oki earthquake by computing the spectral ratio between surface and borehole records (Fig. 4). The non-linear soil response is characterized by a de-amplification and a shift of the amplification peak to lower frequencies (red curve in Fig. 3b). I also observed that shallow soils need several months to heal after the mainshock (blue and brown curves in Fig. 4b). These results have been published as a report for the ONAMAZU Japan (JST) – France (ANR) research project.

During my postdoc at Harvard University, I developed a technique that uses the continuous records of one seismic station located near the Earth surface to monitor the processes that occur in shallow soils around the sensor (Single-station cross-correlation (SC) method). I used one of the densest seismic networks in the world, composed of 244 seismic stations in the Tokyo Metropolitan area, to investigate the seismic velocity changes caused by the 2011 M_w 9.0 Tohoku-Oki earthquake. I observed large coseismic velocity drops in the subsurface that correlate well with the strength of ground shaking and local site conditions (Video 1). Such results are critical to better understand the processes occurring in the near surface where human-made constructions are located.

Fig. 4: (a) Ground acceleration of the 2011 M_w 9.0 Tohoku-Oki event recorded by horizontal components at the IBRH16 KiK-net station (Ibaraki Prefecture) in a ~300 m depth borehole (black) and at the surface (grey). (b) Linear soil responses computed from weak ground motions that were recorded 1 month before (green) and after the mainshock (blue and brown). The red curve is the non-linear soil response during the mainshock.

Video 1: Relative velocity changes in the near-surface of the Tokyo metropolitan area during the 2011 Mw 9.0 Tohoku-Oki earthquake. Single-station cross correlation (SC) functions are computed using the continuous data of the MeSO-net network. The velocity changes (dv/v) are computed from the SC functions.

3) Long term monitoring of the Earth near surface

The Earth's physical properties can be affected by tectonic and environmental forcing through time. I applied the SC technique to 8 years of continuous records at MeSO-net stations (Viens et al, 2018, AGU). Seasonal variations as well as a clear sensitivity to precipitation and to the 2011 Mw 9.0 Tohoku-Oki earthquake strong ground motions can be observed at this station (Fig. 5).

Fig. 5: (top) Relative seismic velocity change (dv/v) at the NS7M station calcultated from the average of the Z-N and Z-E SC functions (1-day stack) from 2010 to the end of 2017. The location of the NS7M station can be seen in Fig. 4. (Bottom) Temperature (orange) and precipitation (blue) at the Funabashi weather station located 5.3 km away from the NS7M station

4) Deep learning applied to passive seismology

To improve the precision and temporal resolution of seismic monitoring studies, I developed a new denoising technique based on deep learning. This method takes advantage of convolutional denoising autoencoders and significantly improves the monitoring of the Earth's near surface. In Viens and Van Houtte (2019, GJI), we developed and applied an algorithm, called ConvDeNoise, to denoise SC functions in the Tokyo metropolitan area.

Fig. 6: Topographic map of a part of the greater Tokyo area, Japan including MeSO-net stations (white circles) and the three stations used in this study (red circles). The blue triangle represents the Funabashi weather station operated by JMA and rivers are shown by white lines. For the NS7M, STKM and NMDM stations, the 20-minute resolution dv/v(t) calculated from the (top) raw and (bottom) denoised with ConvDeNoise SC functions are shown. Note that the y-axis representing the dv/v variations is different for each station. The color of each circle represents the correlation correlation after stretching. The duration of the rainfall event that occurred during this period is highlighted by the light blue area and the heavy rain period is shown by the dark blue area in each subplot (This is Figure 9 from Viens and Van Houtte, 2019 GJI).

Google Sites

Report abuse