Energy Geotechnics – Wuttke, Bauer & Sánchez (Eds)

© 2016 Taylor & Francis Group, London, ISBN 978-1-138-03299-6

Monitoring elastic parameter changes in the vicinity of salt caverns due to

cyclic loading by seismic waveform inversion

D. Köhn, D. De Nil, S.A. al Hagrey & W. Rabbel

Institute of Geosciences, Department of Geophysics, Christian-Albrechts-University Kiel, Germany

K. Khaledi, D. König & T. Schanz

Faculty of Civil and Environmental Engineering, Chair for Foundation Engineering, Rock and Soil Mechanics,
Ruhr-University Bochum, Germany

ABSTRACT: Increased emission of greenhouse gases into the atmosphere lead to a strong requirement of
renewable energy resources. However, they are intermittent and need buffer storage to bridge the time-gap
between production and public demands. The storage of compressed gas energy in sealed underground structures
like salt caverns is one approach to bridge this time gap. The cyclic loading of compressed air in the cavern,
with periods ranging from hours to days, leads to a response of the host rock in the vicinity of the cavern. To
predict the mechanical behavior of rock salt an elastic-viscoplastic-creep model with damage is implemented in
the finite-element modeling software CODE-BRIGHT. The resulting changes of the elastic material parameters
after 400 loading cycles under extreme loading conditions are converted into seismic P- and S-wave velocity
variations. Changes occur in a thin layer with a maximum thickness of five meters near the ceiling and floor of
the cavern. Within this area P-wave velocities are reduced by 50 m/s to 160 m/s and S-wave velocities by 20 m/s
to 80 m/s, respectively. The local character of the material parameter changes prohibits a resolution by classical
traveltime based seismic tomographic approaches. Therefore, we develop an elastic Full Waveform Inversion
(FWI) strategy for two different acquisition geometries. Acquisition setup 1 consists of a vertical source and
receiver line inside the cavern, while in acquisition geometry 2 the receivers are moved into boreholes outside of
the cavern, resembling a classical cross-hole tomography. To mitigate the non-linearity of the inversion problem
due to strong multiple reflections of elastic waves radiated by the cavern, a time-damping approach is applied
to the recorded seismic data. Both acquisition geometries allow an accurate and highly resolved reconstruction
of the elastic material parameter changes in the vicinity of the cavern due to the cyclic loading process.

1 INTRODUCTION strong cycle loading, changes of the elastic material

parameters in the cavern wall are very small and not
Mitigation of anthropogenic Greenhouse gases, detectable by ”classical” traveltime based tomographic
including CO2 emissions in the terrestrial atmo- approaches. To solve this problem, we investigate the
sphere demands developments of viable alternative of application of a high-resolution elastic time-domain
renewable energy resources including hydroelectric, Full Waveform Inversion (FWI). In a first step the
biomass, solar, wind, marine (wave/tides) and geother- mechanical behavior of the rock salt is predicted by
mal sources. Most of these sources produce energy an elastic-viscoplastic-creep model including damage
only when suitable weather conditions are prevail- implemented in the finite-element (FE) modeling soft-
ing and not when energy is directly demanded. These ware CODE-BRIGHT. The FE modelling covers the
sources are intermittent and need buffer storage to whole cavern construction process, first filling and
bridge the time-gap between off-peak production and strong cyclic loading phase. Modelled changes of elas-
demand peaks. One storage option already commonly tic material parameters in the vicinity of the idealized
used for natural gas are salt caverns. Most cavern numerical cavern after the cyclic loading phase are
monitoring techniques are focused on the accessible used for the subsequent FWI of the seismic wavefield.
cavern space and the injection borehole. The condition The resolution of two different seismic acquisition
and shape of the cavern are mapped by sonar meth- geometries are compared. While the FWI of acqui-
ods in combination with mine surveying techniques at sition geometry 1 is mainly based on the reflected and
the earth surface to determine any ground movement refracted wavefield, acquisition geometry 2 adds trans-
(Reitze et al. 2014), while the sealing of the injection mitted waves. Advantages, disadvantages, resolution
borehole is verified by pressure monitoring. However, and areal coverage of the different setups are discussed
all these methods are not able to monitor changes for the synthetic cavern model example, as well as a
in the host rock surrounding the cavern. Even under checkerboard resolution test.

45
 Figure 2. The loading pattern applied to the inner boundary
of the cavern.

step is assumed to be 160 days. That means the whole

cavern is excavated after 4 years.
Debrining phase: the debrining phase has been
modeled in 9 steps from top to bottom of the cav-
ern. In each step, the brine pressure is replaced by the
gas pressure which is equal to the weight of brine col-
umn from ground surface to the current brine level.
Figure 1. Geometry and boundary conditions of the ideal- Thus, when the pressure of injected gas is equal to the
ized cavern model. brine pressure at the bottom of cavern, the whole brine
is ejected. In this numerical example, each debrining
2 NUMERICAL CAVERN MODEL UNDER step takes 20 days and the whole process is carried out
CYCLIC LOADING within 6 months.
First filling phase: in this phase, the pressure inside
A typical salt cavern with a simplified geometry has the cavern reduces to the minimum pressure of the
been modeled in GID software. GID is used as the cavern. It is assumed that the pressure reduction is
pre-processor and post-processor for the finite element carried out within 5 days.
solver CODE-BRIGHT. For a detailed description of Cyclic loading phase: to model the cyclic operating
the underlying mechanical model we refer to Khaledi condition, the internal pressure of the cavern fluctuates
et al. (2016). The axisymmetrical model with a height within a predefined range. Furthermore, it is assumed
of 800 m and a width of 500 m is shown in Fig. 1. The that the pressure variations during the charge and dis-
cavern itself has a diameter of 37.5 m and a height of charge of the cavern are linear. In this paper, an extreme
233 m and its top and bottom have a spherical shape. loading scenario is defined for the cyclic loading
On the upper model boundary a load of 10 MPa is operation. That means, the stresses around the cavern
applied which represents the overburden load at the top during the cyclic loading phase may locate inside the
of the model. The vertical displacement at the model dilatancy zone. Thus, the unfavorable consequences
bottom is restrained. The density of compact rock salt such as the damage development and microcracking
is assumed to be ρsalt = 2000 kg/m3 and the numerical can be experienced. In this scenario, the pressure varies
simulation is done at constant temperature T = 318 K. between 3 and 8 MPa. The duration of each cycle
The initial stress field is assumed to be isotropic (i.e. is assumed to be 1 day and 400 loading cycles are
σxx = σyy = σzz ). simulated. Fig. 2 describes schematically the loading
The following simulation phases have been consid- pattern which has been defined to simulate the cavern
ered to model the construction process as well as the excavation process as well as its cyclic loading oper-
cyclic loading operation. ation. Fig. 3 shows the final Lamé parameter distri-
Initial phase: At time t = 0 it is assumed that no bution with maximum reductions of 1.1 GPa for λ and
excavation has been performed. For this reason, a uni- 0.7 GPa for μ with respect to the undisturbed host rock.
formly increasing load equal to the geostatic pressure Changes occur at the curved ceiling and floor of the
is applied to the inner boundary of the cavern. Since cavern in a layer with a thickness of approximately 5 m.
the initial stress state is isotropic, the principal stresses
around the cavern are identical before the excavation.
Leaching phase: in order to model the leaching pro- 3 THEORY OF SEISMIC MODELLING AND
cess, the applied load inside the cavern which is equal FULL WAVEFORM INVERSION
to the geostatic pressure is gradually reduced to the
brine pressure (with ρbrine = 1100 kg/m3 ). This pro- Before applying the elastic FWI to the CODE-
cess has been performed in 9 excavation steps from BRIGHT cavern modelling results, we shortly
bottom to the top of the cavern. The duration of each summarize the fundamental physics of the seismic

46
Figure 3. True distribution of the Lamé parameters λ (a) and μ (b) in the cavern host rock after modelling the construction
process and strong cyclic loading operation.

forward problem and introduce the theory of full parallelized by domain decomposition using the Mes-
waveform inversion. sage Passing Interface (MPI). For a detailed descrip-
tion of the forward code, we refer to Bohlen (2002)
3.1 The seismic forward problem and Köhn et al. (2014).
The behavior of seismic waves in a 2D isotropic linear- 3.2 2D elastic full waveform inversion
elastic medium for the PSV case can be described by
the following equations of motion The classical FWI approach relies on the minimization
of the data residuals delta δv = vmod − vobs between
modelled seismic data vmod and field data vobs to
deduce high resolution models of elastic material
parameters in the subsurface. To solve this nonlin-
ear optimization problem an appropriate objective
function E has to be defined.

The objective function Eq. (2) can be minimized by

iteratively velocity Vs) at iteration step n, starting
with an initial background model m0 using the quasi-
Newton limited memory Broyden-Fletcher-Goldfarb-
Shanno (l-BFGS) method (Nocedal & Wright 2006)

where λ, μ denote the Lamé parameters, (vx , vz )

particle velocity vector, σxx , σzz , σxz stress tensor
components and (fx , fz ) directed body force vector,
respectively. For the numerical solution of eqs. (1) we
use a time-domain 2D finite-difference (FD) scheme where the product of the inverse Hessian H −1 with
with 2nd order operators in time and 4th order in space the gradient ∂E/∂m is iteratively approximated by
on staggered grids (Virieux 1986, Levander 1988). At finite-differences. The time-domain gradients can be
all boundaries convolutional PML (C-PML) absorb- effectively calculated using the adjoint state method
ing boundary conditions are implemented (Komatitsch (Tarantola 2005, Köhn et al. 2012). The step length
& Martin 2007) to avoid artificial boundary reflec- μn is estimated by an inexact parabolic line-search
tions.To reduce computation time, the resulting code is (Nocedal & Wright 2006).

47
4 MONITORING OF ELASTIC PARAMETER
CHANGES IN THE CAVERN MODEL

4.1 Model parametrization and synthetic seismic

data acquisition
Because the FWI approach is based on a model
parametrization in terms of P-wave and S-wave veloc-
ities, the modelled Lamé parameter distribution λ, μ
in the vicinity of the cavern after the modelled con-
struction process as well as the strong cyclic loading
operation from section 2 is converted to seismic
velocities via:

Figure 4. Seismic acquisition geometries used for the

FWI monitoring strategy: pure reflection seismic (a) and
cross-hole tomography geometry (b). White squares and
green dots denote the positions of source and receiver lines,
The changes of seismic velocities with respect to the respectively.
initial state of the host rock Vp0 = 4100 m/s, Vs0 =
2193 m/s

are shown in Fig. 6(a) and (b), respectively. Near

the curved ceiling and floor of the cavern, P-wave
velocities are reduced by 50 m/s to 160 m/s and S-
wave velocities by 20 m/s to 80 m/s, respectively. To
test the applicability of elastic FWI to resolve these
small material parameter changes, synthetic datasets
for two different acquisition geometries are calculated
by solving the forward problem (1). Acquisition setup
1 consists of a vertical source and receiver line inside
the cavern (Fig. 4a). 113 airgun sources radiate a pres-
sure wavefield, which covers a frequency range from
10 Hz to 100 Hz. The seismic waves are recorded by
226 hydrophones. This acquisition configuration has
the advantage, that no monitoring boreholes outside
of the cavern are required, sources and receivers can
be setup via the injection tube, similar to the equip-
ment used for sonar surveys. On the downside, this Figure 5. Seismic section for shot 1 (acquisition geome-
geometry can only record the reflected, refracted and try 1).
diffracted wavefield. To investigate to what extent this
limitation has an effect on the FWI resolution, we also 4.2 Application of elastic FWI
test a second acquisition geometry (Fig. 4b), where
560 multi-component receivers are placed into two To mitigate the non-linearity of the inverse problem,
boreholes outside of the cavern, while the shot posi- introduced by the complexity of the recorded wave-
tions inside the cavern are unchanged. This resembles field, we apply a sequential time-damping approach
a more classical cross-hole tomography approach with for the FWI (Brossier et al. 2009). Each trace of the
a much higher ray-coverage than acquisition geometry seismic section is damped from the first time sam-
1. As an example Fig. 5 shows the recorded pressure ple by the exponential function exp(−γt), with γ =
wavefield excited by shot 1 for acquisition geometry 1. 160, 80, 40, 20, 0. Starting model for the inversion is
Note the dominant direct wave and its reflection from the initial state of the host rock Vp0 , Vs0 including the
the cavern floor, the weak but still distinct refracted cavern. We assume that the geometry of the cavern was
wave and numerous reflections and their correspond- accurately estimated by sonar techniques and therefore
ing multiples. Due to the strong material parameter do not allow any model changes within the cavern. The
contrast at the interface between air-filled cavern and FWI result using acquisition geometry 1 are shown in
host rock, it is also likely that significant P to S and S Fig. 6 (c+d). Despite the limitation to reflected and
to P conversions occur. refracted seismic waves, the FWI could successfully

48
Figure 6. Cavern model: Comparison of the true changes of P-wave velocity δVp (a) and S-wave velocity δVs (b) with the
corresponding FWI results using acquisition geometry 1 (c+d) and acquisition geometry 2 (e+f).

reconstruct distribution and magnitude of the seismic models and inverted by FWI as in section 4.2. The
velocity variations due to the cyclic loading of the results for acquisition geometry 1 (Fig. 7 c+d) show
cavern. Introducing transmitted seismic waves with that the P-wave velocity checkerboard can be resolved
acquisition geometry 2 does not significantly improve in a layer with a thickness of roughly 2 m along the
the FWI result (Fig. 6 e+f). The S-wave velocity vari- circumference of the cavern. In case of the S-wave
ations show a little bit less artifacts below the cavern velocity model, this layer extents to a thickness of
floor. about 7 m. Acquisition geometry 2 (Fig. 7 e+f) is not
able to add a significant improvement in terms of reso-
lution for the P-wave velocity model in the vicinity of
4.3 Resolution analysis the cavern compared to acquisition geometry 1. This
For a quantitative estimation of the FWI resolution can be explained by the long wavelength of the P-wave
and areal coverage of the two different acquisition within the host rock of about 40 m. Nevertheless, near
geometries we apply a checkerboard test. Except for the monitoring boreholes some tiles can be resolved
the air-filled cavern the homogeneous seismic back- by FWI, maybe due to possible trade-offs between the
ground velocity models Vp0 , Vs0 are covered with Vp and Vs model. Tiles of the Vs model are visible in
5 m × 5 m large tiles (Fig. 7a+b). For comparison the a layer with a thickness of 12 m around the cavern and
minimum wavelength of the P-wave in the cavern also in a maximum radius of 5 m around the receiver
equals 3 m, in the host rock 40 m, while the minimum boreholes. Midway between boreholes and cavern sur-
S-wavelength is 20 m. In each tile the P-wave veloc- face, the checkerboard resolution is strongly reduced.
ity varies between ±150 m/s, the S-wave velocity by In summary, both acquisition setups are able to mon-
±80 m/s. Corresponding to the approach in section itor the thin layer in the cavern vicinity affected by
4.1, synthetic data is calculated for the true seismic

49
Figure 7. Checkerboard test: Comparison of the true changes of P-wave velocity δVp (a) and S-wave velocity δVs (b) with
the corresponding FWI results using acquisition geometry 1 (c+d) and acquisition geometry 2 (e+f).

the strong cyclic loading operation, especially S-wave receiver lines. The cylinder-symmetric cavern geom-
velocity changes. etry and material parameter distribution is a strong
simplification compared to a 3D asymmetric cavern.
Nevertheless, we assume that a similar seismic FWI
5 CONCLUSIONS monitoring concept could be applied to more complex
cavern geometries, if acquisition geometries based
We presented a feasibility study for a cavern mon- on seismic arrays and beam-forming techniques are
itoring concept based on seismic FWI, focusing on adapted to the problem.
elastic parameter changes within the host rock under
strong cyclic loading. For a realistic prediction of
the rock salt response, the cavern excavation and
strong cyclic loading operation is modelled using an
elastic-viscoplastic-creep model with damage, imple-
mented in the FE modeling software CODE-BRIGHT.
While the mechanical changes are very small and
localized in a thin layer around the cavern, the applied ACKNOWLEDGMENTS
seismic FWI approach is able to resolve these vari-
ations. As demonstrated by a checkerboard test, the This study has been carried out within the frame-
resolution of a pure reflection based acquisition setup work of the ANGUS+ research project (Bauer et al.
within the cavern is comparable with an acquisition 2013) funded by the German Federal Ministry of Edu-
geometry based on transmitted waves, requiring addi- cation and Research (BMBF). The FWI inversions
tional monitoring boreholes near the cavern to host the were performed on the NEC-HPC-Linux-Cluster at

50
Kiel University. The 2D FWI code DENISE Black- caverns in rock salt subjected to mechanical cyclic load-
Edition is available under the terms of GNU GPL 2.0 ing. International Journal of Rock Mechanics and Mining
at https://github.com/daniel-koehn. Sciences 86, 115 – 131.
Köhn, D., D. De Nil, A. Kurzmann, A. Przebindowska, &
T. Bohlen (2012). On the influence of model parametriza-
tion in elastic full waveform tomography. Geophysical
REFERENCES Journal International 191(1), 325–345.
Köhn, D., A. Kurzmann, D. De Nil, & L. Groos (2014).
Bauer, S., C. Beyer, F. Dethlefsen, P. Dietrich, R. Duttmann, DENISE - User manual. available at http://www.geophysik.
M. Ebert, V. Feeser, U. Görke, R. Köber, O. Kolditz, uni-kiel.de/%7Edkoehn/software.htm.
W. Rabbel, T. Schanz, D. Schäfer, H. Würdemann, & Komatitsch, D. & R. Martin (2007). An unsplit convolutional
A. Dahmke (2013). Impacts of the use of the geological perfectly matched layer improved at grazing incidence for
subsurface for energy storage: an investigation concept. the seismic wave equation. Geophysics 72(5), 155 – 167.
Environmental Earth Sciences 70(8), 3935–3943. Levander, A. (1988). Fourth-order finite-difference P-SV
Bohlen, T. (2002). Parallel 3-D viscoelastic finite-difference seismograms. Geophysics 53(11), 1425–1436.
seismic modelling. Computers & Geosciences 28(8), Nocedal, J. & S. Wright (2006). Numerical Optimization.
887–899. Springer, New York.
Brossier, R., S. Operto, & J. Virieux (2009). Seismic imaging Reitze, A., F. Hasselkus, & P. Wurmbauer (2014). Today’s
of complex onshore structures by 2D elastic frequency- technology for monitoring caverns. BHM Berg- und
domain full-waveform inversion. GEOPHYSICS 74(6), Hüttenmännische Monatshefte 159(4), 149–153.
WCC105–WCC118. Tarantola, A. (2005). Inverse Problem Theory. SIAM.
Khaledi, K., E. Mahmoudi, M. Datcheva, & T. Schanz (2016). Virieux, J. (1986). P-SV wave propagation in heteroge-
Stability and serviceability of underground energy storage neous media: velocity-stress finite-difference method.
Geophysics 51(4), 889–901.

51