AN ADAPTIVE FREQUENCY DOMAIN FINITE ELEMENT MODEL

FOR SURFACE WAVE TESTING OF PAVEMENTS

N. Ryden1 and M. Castaings2

1
Engineering Geology, Faculty of Engineering, Lund University, 22100 Lund, Sweden
2
Université Bordeaux 1, Laboratoire de Mécanique Physique, UMR CNRS 5469, France

ABSTRACT. The proposed adaptive frequency domain finite element (FE) model makes it possible
to simulate wave propagation in a pavement structure over a wide frequency range using a normal
size computer. The geometry and mesh size are optimized and regenerated for each frequency using a
simple script in a commercial FE software package. Results from the FE model agree well with
normal mode dispersion curves for a 3-layer reference pavement model. The proposed approach is
fast and computationally efficient and intended to be used as a forward model for iterative surface
wave inversion.

Keywords: Surface Waves, Finite Element Modeling, Pavement, Dispersion

PACS: 43.35.Cg, 43.20.Ks, 43.20.Mv

INTRODUCTION

Aging infrastructure and increasing loads fuel the need for efficient non-destructive
quality control/assurance of civil infrastructures. Surface wave testing of pavements is a
promising non-destructive technique for obtaining both thickness and velocity (stiffness)
of all the layers in a pavement construction [1].
Surface wave techniques are based on the dispersive properties of surface waves in
a layered medium. Dominating phase velocities are measured along the surface over a
wide frequency range in order to estimate the elastic structural properties of the unknown
layered medium. Recent developments have been focused on data processing [2,3] along
with optimization techniques for iteratively finding the best theoretical match to the
measured response [4,5]. However, the forward model used to predict the theoretical
response of a layered structure during the iterative inversion process, is still based on
normal mode dispersion curves for a layered half-space [6]. These traditional matrix based
techniques are fast and efficient for normal mode dispersion curve calculations, but are
only valid for far field propagation of surface waves in horizontally infinite homogeneous
layers. The traditional matrix techniques cannot be used to accurately predict the effect of
2D and 3D effects from the actual geometry, variable material properties, scattering from
cracks and/or other anomalies.
For future research and development of surface wave techniques it is important to
move towards more realistic 2D and 3D theoretical models. Finite Element (FE) models
have been used to study 2D effects on surface wave propagation [7,8]. However, these

CREDIT LINE (BELOW) TO BE INSERTED ON THE FIRST PAGE OF EACH PAPER

EXCEPT FOR ARTICLES ON PP. 19 - 32, 217 - 223, 371 - 377, 1091 - 1098, 1099 - 1106,
1190 - 1197, 1719 - 1726, 1792 - 1799, 1870 - 1877, 1878 - 1885, and 1902 – 1909

CP1096, Review of Quantitative Nondestructive Evaluation Vol. 28,

edited by D. O. Thompson and D. E. Chimenti
© 2009 American Institute of Physics 978-0-7354-0629-2/09/$25.00

1481
studies have been based on time-marching finite element or finite difference modeling
using only a few elastic layers without large velocity contrasts and without frequency
dependent material properties. The computational load increases immensely for problems
with large velocity contrasts and an increasing frequency range since the range of
wavelengths increases with both of these factors. More over it is difficult to model realistic
frequency dependent material properties in time domain. An efficient alternative based on
frequency domain modeling has been presented in [9]. The problem is decomposed into a
few stationary frequencies and by satisfying infinite boundary conditions the solution can
be transformed back to time domain using an inverse Fourier transform. With this
approach it is straight forward to implement user defined frequency dependent velocity
and damping properties, as well as properties varying in any direction of space, even
within any layer.
It is the purpose of this paper to demonstrate an efficient frequency domain finite
element model for surface wave testing of pavements. The viscoelastic frequency domain
approach [9] has been extended using an adaptive geometry and mesh model to
accommodate the extremely wide range of wavelengths used in surface wave testing of
pavements. However, the same approach can also be used for normal soil site applications.
The model is intended to be fast enough to be used as a forward model during the iterative
inversion procedure of surface wave data.

FINITE ELEMENT MODEL

The equations of dynamic equilibrium can be solved in either space-frequency or space-

time domain using finite elements. The accuracy of wave propagation simulations using FE
models depends heavily on the space-time discretization of the model [10,11]. Each wavelength
(λ) is approximated with a couple of finite elements and it is generally recommended to use at
least 10-12 degrees of freedom (DOF) per wavelength. In a 2D model with axial symmetry the
total number of DOF’s then becomes a function of the length (R), depth (Z) and the minimum
wavelength (λmin) present in the model.
Meshing a typical 3-layer pavement construction in 2D using a fixed geometry and
mesh size can easily generate several million DOF’s. If this type of model is solved using a
time-marching technique (typically from 0 to 0.030 s) several thousand time steps may be
necessary. However, if the same model is solved in frequency domain it may be enough to
solve for 50 discrete frequencies [9]. By using this method the total time for wave propagation
simulations can be drastically reduced turning hours into seconds. This approach is especially
appealing for surface wave inversion since the inversion procedure is performed in frequency
domain. With an adaptive mesh it also becomes easier to optimize the mesh and absorbing
boundaries regions for each frequency, avoiding artifacts like mesh dispersion typically
encountered at the high frequency limit using time marching techniques [10].

Absorbing Regions

Energy absorbing regions are used to avoid reflections at the boundaries in FE

wave propagation simulations. For frequency domain models it is important to use
excellent absorbing boundaries to simulate an infinite space avoiding potential artifacts
from reflections at the boundaries. Viscoelastic absorbing regions have been successfully
used by [9,11,12]. In this approach the loss factor (η) of the complex viscoelastic modulus
is gradually increased from 0 to 1 over the absorbing region. Compared to perfectly
matched layers this approach requires more elements but are more robust and easy to use
since the absorption over a fixed distance is not dependent on the direction of wave-fronts

1482
propagation, i.e. they also absorb waves having energy and phase velocities of opposite
signs. In a 2D axial symmetric case, Equation 1 is used to define the complex modulus in
the r-direction inside the absorbing region and is given as:

⎡ ⎛ r − ra ⎞ ⎤
3

E = Ereal ⎢1 + i⎜⎜ ⎟⎟ ⎥ (1)

⎣⎢ ⎝ Rabsr ⎠ ⎦⎥

where r is the position along the r-direction and ra and Rabsr are the starting position and
the length respectively. A similar equation is used for the depth (z) direction. The length of
the absorbing region should be 3 times the longest wavelength in the material [9], see
Figure 1. In this specific pavement application the velocity decreases from layer to layer in
the z-direction making λmax and hence Rabsr shorter and shorter with depth.

Adaptive Geometry and Mesh Size

The fundamental difficulties with FE modeling of surface waves in a pavement

structure originates from the large velocity contrast between layers and the wide frequency
range utilized. The top pavement layer can have a maximum compression (P) wave
velocity of more than 4000 m/s while the bottom subgrade material can have a low shear
(S) wave velocity of less than 100 m/s. To resolve the top one or two meters of a complete
pavement construction a sensor array of 4 to 5 meters along the surface and a frequency
range of 50 to 20 000 Hz is usually necessary. This velocity and frequency range
corresponds to minimum and maximum wavelengths of 0.005 to 80.0 m. A fixed model
geometry and mesh size that could accommodate this complete range of wavelengths
simultaneously would require almost 3 billion DOF’s and cannot be solved for using a
normal size computer. Even if the mesh in each layer is adjusted according to the layer
velocities it is still almost impossible to solve for the complete frequency range using one
fixed model and a normal size computer.
To solve the problems above and develop an efficient FE model for surface wave
simulations we propose to use an adaptive geometry and mesh size for each frequency.
The modeled space is divided into an analyzed region and an absorbing region, Figure 1.
The analyzed region is the space of interest and the absorbing region is only used to
simulate infinite boundary conditions. The analyzed region is kept fixed but the size of

FIGURE 1. Schematic of the adaptive FE model with analyzed and absorbing regions.

1483
each element and the absorbing regions are optimized for each modeled frequency. This
means that the total model and element size is big when we solve for low frequencies and
small when we solve for high frequencies. Since the analyzed region is kept fixed we can
still extract the response at the surface over the same fixed distance of interest for all
frequencies.
Using a commercial FE package [13] an adaptive frequency domain model is set-
up with a 2D axial symmetric geometry. The mesh is made of 2nd order quadratic elements
using 4 elements per wavelength (λmin) in each layer. The adaptive geometry is based on
equally sized elements and a fully fixed analyzed region in the r-direction (R=5.0 m). A
semi fixed depth of the analyzed region (Z) is defined as:

Z = h1 + h2 + 2λ3max (2)

where h1 and h2 represents the thickness of layer 1 and 2 respectively and the depth of the
bottom layer (half-space) is calculated as 2 times the maximum wavelength in the 3rd layer
(λ3max), see Figure 1. The element size in the r-direction is determined from the minimum
wavelength in the complete model. In the z-direction the element size is optimized based
on λmin in each layer. This approach saves a lot of computational power since the total
geometry is reduced at higher frequencies when wavelengths are comparatively small.
A vertical unit amplitude force is applied at the free surface at r<0.05 m, i.e.
simulating a circular source with 0.1 m diameter at the centre of the axial symmetry.
Calculated vertical displacement components (w) are extracted along the surface from
r=0.05 to r=5.00 in 0.05 m increments. The axial symmetry of the problem causes a
frequency independent geometrical decay in guided wave amplitudes equal to 1/√r.

Synthetic Reference Layer Model

This pavement layer model (Table 1) has been studied in [14] where normal mode
dispersion curves were calculated in the complex wave number domain. Here the same
layer model is used as a reference to validate results from the proposed adaptive frequency
domain FE model. The 3-layer pavement model will be used to compare phase velocities
and relative amplitudes, between the FE and the matrix method.
Figure 2 shows a comparison of the required number of DOF’s as a function of
frequency for the reference layer model. The fixed geometry (time domain) alternative
requires 3 billion DOF’s for all frequencies. The adaptive geometry and mesh approach is
most efficient at low frequencies requiring less than 1 million DOF’s below 9 kHz. At 20
kHz the adaptive approach still requires almost 5 million DOF’s. Figure 3 shows the
frequency dependency of the total geometry further illustrating the need for an adaptive
geometry.

TABLE 1. Reference pavement layer model, from [14].

Property\Layer Layer 1 Layer 2 Layer 3

Thickness (m) 0.20 0.60 ∞
VS (m/s) 1400.0 500.0 100.0
VP (m/s) 2914.3 1040.8 208.2
Density (kg/m3) 2000 2000 2000
Poisson’s ratio 0.35 0.35 0.35
E-modulus (GPa) 10.5840 1.3500 0.0540

1484
FIGURE 2. Required number of DOF’s for different frequencies.

FIGURE 3. Total size of the adaptive geometry as a function of frequency.

RESULTS

The proposed adaptive FE model is used to simulate wave propagation in the

reference layer model. The model is solved at 400 discrete frequencies from 50 Hz to 12
kHz using a unit amplitude vertical force at the surface (r<0.05 m). Figure 4 shows an
example of the vertical displacements in the complete model at 1000 Hz.
In Figure 5 the frequency phase velocity spectrum have been calculated from the
real part of the solved vertical displacements along the surface (r=0.05-5.00 m) using the
Multichannel Analysis of Surface Waves (MASW) transformation technique [2]. Dark
regions in the image correspond to frequency phase velocity combinations with good
coherence along the surface, revealing the dispersive nature of different modes of guided
waves in the reference layer model. As expected the main trend of the dark regions in
Figure 5 follow a trend similar to the A0 Lamb mode dispersion curve of the top layer as a
free plate [14]. The asymptotic straight part above 10 kHz approaches 1309 m/s which
agree well with the theoretical Rayleigh wave velocity of the top layer.

1485
FIGURE 4. FE solution, real part of the vertical displacement, at 1000 Hz.

FIGURE 5. Frequency phase velocity image obtained from the FE simulation.

Figure 6 shows the results from [14] using the global matrix method [6]. Figure 6a
shows the normal mode dispersion curves (solid lines) where the dominating mode of
propagation at each frequency has been marked with a cross. The dominating response
along the surface jumps from different modes of propagation. Figure 6b shows the relative
amplitude between different modes and frequencies after 5.0 m propagation.

FIGURE 6. Normal mode dispersion curves from the reference layer model (Table 1) along with relative
amplitudes at 5.0 m distance from source, from [14].

1486
FIGURE 7. (a) Normal mode dispersion curves from the reference layer model (Table 1) along with (b)
relative modal amplitudes at 5.0 m distance from source. Phase velocities and total amplitudes (from all
modes) computed using the adaptive FE model are marked with circles.

Figure 7a shows a comparison of phase velocities between the adaptive FE model

and the normal mode matrix method. Circles represent the extracted maxima from the
MASW transformation (dark peaks in Figure 5). As shown in Figure 7a phase velocities
calculated from the FE model compares well with the normal mode dispersion curves and
jumps from mode to mode in a similar pattern as predicted in Figure 6a. A small deviation
from the exact normal mode dispersion curves can be observed in the transition zones
between the different branches of dispersion curves. This deviation is likely to be caused
by interference of different modes with comparable amplitudes.
Total vertical amplitudes at r=5.0 are plotted in Figure 7b along with the predicted
curves from Figure 6b. The general trend of the extracted circles from the FE model agrees
well with the predicted amplitudes from the normal mode dispersion curves. It should be
noted that extracted circles from the FE model corresponds to total displacements while
the lines in Figure 7b corresponds to displacements from each individual mode.

CONCLUSIONS

The proposed adaptive FE model makes it possible to simulate wave propagation in

a pavement structure over a wide frequency range using a normal size computer. The
geometry and mesh size are optimized and regenerated for each frequency using a simple
script in the commercial FE software package [13]. Results from the FE model agree well
with normal mode dispersion curves for a 3-layer reference pavement model. The
proposed approach is fast and computationally efficient enough to be used as a forward
model for iterative surface wave inversion. By solving the model in space-frequency
domain it is also straight forward to implement frequency dependent material properties as
well as properties varying in any direction of space, even within a layer.

1487
REFERENCES

1. S. Nazarian, D. Yuan and V. Tandon, Transportation Research Record 1654, pp. 50-60
(1999).
2. C. B. Park, R. D. Miller and J. Xia, Geophysics 64, pp. 800-808 (1999).
3. C. P. Lin and T. S. Chang, Soil Dynamics and Earthquake Engineering 24, pp. 877-
886 (2004).
4. K. S. Beaty, D. R. Schmitt and M. Sacchi, Geophysical Journal International 151, pp.
622-631 (2002).
5. N. Ryden and C. B. Park, Geophysics 71, pp. 49-58 (2006).
6. M. J. S. Lowe, IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency
Control 42, pp. 525-542 (1995).
7. N. Gucunski, V. Ganji and M. H. Maher, Soil Dynamics and Earthquake Engineering
15, pp. 223-231 (1996).
8. C. P. Lin and C. H. Lin, Soil Dynamics and Earthquake Engineering 27, pp. 541–552
(2007).
9. M. Castaings, C. Bacon, B. Hosten, M. V. Predoi, Journ. of Acoust. Society of Am.
115, pp. 1125-1133 (2004).
10. J. D. De Basabe and M. K. Sen, Geophysics 72, pp. T81-T95 (2007).
11. M. Drozdz, "Efficient Finite Element modelling of ultrasound waves in elastic media",
Ph.D. Theses, Mechanical Engineering, Imperial College (2008).
12. M. Drozdz, L. Moreau, M. Castaings, M. J. S. Lowe, P. Cawley, “Efficient numerical
modelling of absorbing regions for boundaries of guided waves problems”, in Review
of Progress in QNDE, 25, edited by D. O. Thompson and D. E. Chimenti, AIP
Conference Proceedings, American Institute of Physics, Melville, NY (2006), pp. 126-
133.
13. Comsol Multiphysics, User’s Guide, Version 3.4 by – COMSOL AB 2007,
http://www.comsol.com/
14. N. Ryden and M. J. S. Lowe, Journal of the Acoustical Society of America 116, pp.
2902-2913 (2004).

1488