resonant ultrasound spectroscopy - theory and application

8/11/2019 Resonant Ultrasound Spectroscopy - Theory and Application

1/17

CWP-438P

Resonant ultrasound spectroscopy:

Theory and application

Brian Zadler1, Jerome H.L. Le Rousseau1, John A. Scales1 & Martin L. Smith21Physical Acoustics Laboratory and Center for Wave Phenomena, Department of Geophysics,

Colorado School of Mines, Golden, Colorado2New England Research, White River Junction, Vermont

ABSTRACT

Resonant Ultrasound Spectroscopy (RUS) uses normal modes of elastic bodiesto infer material properties such as elastic moduli and Q. In principle, the com-plete elastic tensor can be inferred from a single measurement. For centimeter-sized samples RUS fills an experimental gap between low-frequency stress-strainmethods (quasi-static up to a few kHz) and ultrasonic time-delay methods (hun-dreds of kHz to GHz). Synchronous detection methods are used to measure theresonance spectra of homogeneous rock samples. These spectra are then fit in-teractively with a model to extract the normal-mode frequencies and Q factors.Inversion is performed by fitting the the normal-mode frequencies. We have de-veloped a forward model to compute resonant frequencies and displacements,based on an approximation that reduces the calculation to a generalized eigen-value equation that can be solved numerically. The results of this calculation,normal mode frequencies and displacements, depend both on properties of themedium and its geometry. For the inversion we start with a high symmetry

model (isotropic) and successively lower the symmetry of the model until areasonable fit with the data is achieved. This ensures that inferred anisotropicparameters are statistically significant.

1 INTRODUCTION

1.1 Laboratory Measurements of Elastic

Properties

All mechanical methods for measuring the elastic prop-erties of laboratory specimens are divided into threetypes: quasi-static, resonance, and time-of-flight (if weare willing to accept modest intellectual casualties).Those types reflect the ratio of the wavelength of themechanical signals used to the size of the specimen un-

der test.Quasi-static methods, such as cyclic loading

(@warning Citation batzle on page 1 undefined), sub-ject the sample to deformations that are very slow com-pared to any of its natural mechanical resonances, sothat the sample is close to mechanical equilibrium atall times during the test. By measuring both the ap-plied stresses as well as the strains induced we hopeto infer the samples elastic compliances; if we think ofstress, strain, and compliance as complex functions of

frequency, we can see that the phase of the compliance(which follows from the phase shift between stress andstrain) tells us about sample anelasticity. Cyclic loadingmeasurements can be made from frequencies of the or-der of 102 Hz (above which the typical apparatus tendsto become animated) down to zero frequency and havethe geophysically-attractive property that they can bemade at frequencies which overlap those of exploration(at least) seismology. The most serious limitations ofthis technique are the difficulty of accurately account-

ing for the complex mechanical behavior of the mea-surement apparatus and of making accurate, low-noisestrain measurements at low frequencies.

Resonance techniques measure the frequencies ofthe specimens elastic resonances, or free oscillations.These frequencies reflect the size, shape, and elasticcomposition of the sample; each corresponds to a partic-ular bundle of bouncing, interconverting traveling waveswhich conspire to exactly repeat at intervals of 1/f,where f is the resonance frequency. Given a sufficient


2/17


3/17


4/17

4 B. Zadler et al.

torsional axisymmetric pure shear motions consistingof rigid rotations of rings of material around the axis.The frequencies of these modes depends wholly uponthe samples shear velocity.

extensional axisymmetric mixtures of compressionand shear motions. At low frequencies, when wave-

lengths are long compared to the samples diameter,these modes are essentially an axial compression cou-pled to a radial expansion, their frequencies are thenalmostwholly dependent on the samples bar velocity,

VE =

E

and the modes are called bar modes. At

higher frequencies the modes become more complex andit would be more appropriate to call them simply non-torsional.

non-axisymmetricor flexural all modes for whichn= 0 in sin(n) (above) are in this class. The axisym-metric modes represent energy traveling purely up anddown the cylinder axis; the non-axisymmetric modestravel along paths that are tilted with respect to that

axis.Although these classes break down when axisymmetryis destroyed (as is almost always the case in a real exper-iment), they are still very useful in nearly-symmetriccircumstances, and even in not-very-symmetric ones.

The flexural modes occur in pairs, called doublets,both members of which have the same resonance fre-quency. This obscure distinction is important when themodes are perturbed by small, non-symmetric changesin the samples composition or shape: perhaps a smallcrack develops in one side of a specimen. In that circum-stance the resonance frequencies are slightly shifted bythe perturbation. For an axisymmetric mode this shiftsimply moves the modes peak a little bit in frequency.

The two members of a doublet, however, may experiencedifferent shifts and what appeared to be a single peak inthe unperturbed sample may become two distinct peaksin the perturbed sample.

3.2 Mode Shapes

Before we get into the exact mechanics of calculating thedisplacement field for a generic object, it is beneficial topicture the possible deformations for a common rightcylinder. Figure 1 shows the particle motion of the sam-ple surface for several types of modes at two extremepoints of the cycle. (The deformation is greatly exag-

gerated: under normal circumstances in the laboratory,particle motions are of the same order as atomic diame-ters.) The top row of figures shows the mode shapes forthe two lowest extensional modes. The motion in bothcases is clearly axisymmetric, and has both axial and ra-dial components. The bottom row of figures shows thetwo lowest torsional modes which are also axisymmetric,but here the particle motion is entirely azimuthal, thatis, in the local direction of the coordinate. The middlefigures show the shape of the first flexural mode, which

is also the lowest-frequency mode of the system. Themodes bending shape is clearly non-axisymmetric.

4 DISPLACEMENT CHARACTERIZATION

Next we turn to the problem of calculating the res-onances for an arbitrary elastic body. More rigorousstudies of body resonances have been performed in thepast (@warning Citation strutt on page 4 undefined;@warning Citation tromp on page 4 undefined; @warn-ing Citation vi91 on page 4 undefined; @warning Cita-tion mi97 on page 4 undefined). Consider an isolatedpurely elastic body (no attenuation) with a volume V.Its kinetic energy Ek is (@warning Citation ak80 onpage 4 undefined; @warning Citation mi97 on page 4undefined)

Ek =

V

12

tuitui dV ,

where ui, 1 i 3, is the displacement field in a La-grangian system of coordinates for R3 and is the den-sity. The summation convention for repeated indices isapplied. The potential energy Ep is given by the strainenergy (@warning Citation ak80 on page 4 undefined)

Ep =

V

12 Cijkljuiluk dV ,

where (Cijkl) is the elastic stiffness tensor.The LagrangianL of the system can then be written

as

L= Ek Ep =

V

12

tuitui dV

V

12 Cijkljuiluk dV .

We want to characterize displacement fields thatnaturally occur in terms of normal modes in the absenceof driving force. Such displacement fields may have beeninitiated by a driving force but we assume this force isnot present anymore and the transient response is neg-ligible. Applying Hamiltons principle we get two equa-tions

ui j(Cijklluk) = 0 , 1 i 3 ,

which is the elastic wave equation (with no right-handside) and

(Cijklluk nj)|S= 0, 1 i 3 , (1)

which is the boundary condition at the free surface S.In the frequency domain, the homogeneous elastic

wave equation becomes

2ui+ j(Cijklluk) = 0, 1 i 3 . (2)

The system of equations (2) and (1) has solutions onlyfor a set of frequencies corresponding to the resonantfrequencies of our sample. As we shall see later, they


5/17

Resonant ultrasound spectroscopy 5

first extensional overtone

lowest torsional mode

lowest extensional mode

first torsional overtone

20.49 kHz

36.85 kHz69.94 kHz

45.33 kHz22.66 kHz

lowest flexural mode

overall lowest elastic mode

Figure 1. Surface particle displacements of several characteristic modes of the generic soft rock model

depend both on the shape and the elastic properties ofour body.

If we now define a new function T as

T =

V

12

2uiui 12 Cijkljuiluk

dV ,

we then have (@warning Citation mi97 on page 5 un-

defined)

T = T

=

V

2uiui Cijkllukjui

dV = 0 .(3)

If we choose a basis we can decompose any componentui, 1 i 3, of our displacement field into this ba-


6/17

6 B. Zadler et al.

sis. With such a decomposition equation (3) can thenbe viewed as a generalized eigenvalue problem (see alsonext section). Solving this eigenvalue problem, we find aset of frequencies (eigenvalues) for which the system hasa non-zero solution. These are the normal modes. Thedisplacement field can then be decomposed in a Fourier

series scheme into the basis of the normal modes.

5 NUMERICAL APPROACH OF THE

GENERALIZED EIGENPROBLEM

To perform numerical approximations of the normalmodes we follow closely Migliori and Sarrao (@warn-ing Citation mi97 on page 6 undefined). We restrictourselves to polynomials of degree less than or equalto N, i.e., the subspace generated by the monomials = x

l ym zn for which

l+ m+ n < N .

We can then write for 1 i 3 and finiteui =

ai, ,

with = xl ym zn.

We expect to obtain an approximation good enoughfor our purpose with N large enough (see Sec. 10). Thisleads to

N+33

= (N+ 1)(N+2)(N+ 3)/6 choices ofl,

mand n for each component ofu. Hence, the dimensionof the problem is nowR = 3 (N+1)(N+2)(N+ 3)/6.Having restricted ourselves to this finite dimensionalsubspace, we can write T in the approximate form

T =

V

12

2 ai,ai,dV

V

12

Cijklai,ak,jl dV ,

which can be written in a matrix form

T = 12

2 at E a 12

at a .

Both matrices Eand are of dimension R. They havethe forms

Eik = ik

V

dV , (4)

and

ik =Cijkl

V

jldV , (5)

where we have assumed that (Cijkl) and are constant.This is the major assumption of this section.

To search for an extremum of T, i.e., an approxi-mate normal mode, we set the differential ofT, evalu-ated at a, to zero

daT(a) = 2Ea a= 0, (6)

which is a generalized eigenvalue problem. The ma-trices E and have remarkable properties that we

can exploit: they are real symmetric. Furthermore,E is positive definite. System (6) can be further-more simplified for its numerical evaluation. In theorthorhombic and higher symmetry case, we can ordermatrices E and in such a way to make them blockdiagonal. There are then 8 diagonal blocks. Instead of

(numerically) solving for a large eigenvalue problem,we now solve for 8 smaller ones. See A for further details.

To solve this generalized eigenvalue system we usestandard computer libraries such as Lapack(@warningCitation laug on page 6 undefined). The eigenvaluesand eigenvectors are both dependent on the elastic prop-erties and the geometry of the specified body accordingto equation (4) and equation (5). The geometry depen-dency is found in the general value of the following typeof integrals:

IV =

V

xlymzndV . (7)

If we can compute integral (7) we can compute both Eand and can numerically solve for the (approximate)normal modes. Here we consider three different shapes:rectangular parallelepipeds, cylinders, spheroids. Thedimensions of those shapes are described in Figure A1.In all those cases, we use the center of symmetry of theshape as the coordinate origin.

The calculations of the integral (7) for spheres,cylinders and rectangular parallelepipeds are carried outin Appendix B.

6 THE INVERSE PROBLEM

So far, we have dealt with a forward modeling problem,i.e., knowing the characteristics of an elastic bodywe compute its normal modes. We would like tosolve the inverse problem: starting from the measuredeigenfrequencies, can we infer the elastic parameters?Here, our model could have different variables: we caninvert for the dimensions of the body (we still have toassume a type of shape though). We can also invertfor the elastic moduli or the crystallographic axis. Weshall now focus on the latter topic.

6.1 Objective function and non-linear

optimization algorithms

To estimate how good a given model is, i.e., how well itcan predict the data, we make use of an the followingobjective function

F =i

wi(f(p)i f

(m)i )

2 , (8)

where f(p) are the computed frequencies and f(m) arethe measured ones. We minimize the difference between


7/17


predicted and measured frequencies in a least squaresense. The weights wi characterize the confidence wehave in the measurements. We minimize the objectivefunction, up to some tolerance, over the space of allfeasible models. In a neighborhood of the an extremumthis is a quadratic minimization problem. However,

given the presence of noise we say a model fits thedata if F is less than some tolerance, which can bequantified via a 2 type criterion.

An efficient way to minimize the objective functionis to apply a conjugate gradient method (@warning Ci-tation fl80 on page 7 undefined). We use a non-linearform of conjugate-gradient which amounts to makingrepeated quadratic approximations. One advantage ofsuch a method is that it does not require the knowl-edge of the Hessian of the objective function. A se-quence of search directions, hi, is constructed and linesearches are performed in order to minimize the ob-jective function along such directions. The procedure

is as follows: assume you have a search direction hiand a model mi. Let gi = F(mi). By minimizingthe objective function, F, along the hi direction, i.e.,minimizing j() = F(mi + hi), we find a new modelmi+1 = mi +ihi. Then the new search direction isgiven by:

hi+1 = gi+1+ ihi , (9)

with gi+1= F(mi+1) and

i =gi+1

tgi+1gitgi

. (10)

The method is initiated by taking the steepest de-scent direction as the starting search direction, i.e.,

h0 = g0 = F(m0). In the case of a true quadraticform such an algorithm yields convergence in a finitenumber of steps, in exact arithmetic. Formula (10) wasderived from that particular case. For further referenceswe refer to Fletcher (@warning Citation fl80 on page7 undefined) and Press et al.(@warning Citation pr86on page 7 undefined).

We can obtain an exact expression for the gradientof the objective function. The partial derivative of theobjective function with respect to a particular parame-ter p is

pF = i

wi2pf(p)i (f

(p)i f

(m)i ),

and differentiating equation (6) with respect to p , com-posing by at on the left and using and Esymmetriesyield (@warning Citation mi97 on page 7 undefined)

p(2) = at (p

2pE) a ,

where we made use of the normalization atEa = 1.Hence, knowing the eigenvector a as well as the eigen-frequency allows us to compute the gradient of theobjective function. Knowledge of the gradient is a key

parameter for convergence and efficiency. Knowledgeof the Hessian would allow more efficient optimizationmethods, e.g., Newton methods, but would requiresome significant additional computational effort whilecomputing the eigenvectors (@warning Citation mi97on page 7 undefined).

To perform the non-linear conjugate gradientalgorithm we use the C++ COOOL library (CWPObject Oriented Optimization library) (@warning Ci-tation de96-a on page 7 undefined; @warning Citationde96-b on page 7 undefined). A code does the forwardmodeling and computes both the objective functionvalues and its gradient using the measured resonantfrequencies. Using this result, COOOL can perform theline search along the direction chosen and find the beststep to use. Starting from this point, with the value ofthe new gradient a new direction is chosen as describedin equation (9) and a new line search is performed.

7 EXPERIMENT DESIGN

In typical RUS measurements, we try to measure all ofthe resonances below some upper limit. Having a com-plete set of resonances assures us that we have extractedall of the available information and significantly sim-plifies the inverse calculation. Most important, we canconfidently match each observed mode with a computedone, a step that is essential to interpreting the data.As Qs decrease, however, it becomes increasingly dif-ficult, and eventually impossible, to determine all (oreven most) resonances.

Figure 2 shows the theoretical resonance responsefor the generic soft rock sample (defined in Section 3) us-ing edge-contacting pinducers, for three assumed valuesof compressional and shear Q. The three cases repre-sent, by geologic standards, very high, moderate, andlow (but still plausible) Q regimes. Its clear that wecannot depend upon observing a complete mode cata-log in the presence of even moderate attenuation.

7.1 Sample Shape and Mode Sensitivities

Figure 3 shows the relative sensitivity of each mode tochanges in compressional velocity for our generic soft

rock defined in Section 3. The quantity shown is

RP = af

af+ bf, (11)

whereaf andbfare related to the sensitivities of themode toVP andVS, respectively.RP is a measure of therelative importance of the samples compressional andshear speeds in determining the given modes frequency.Both af and bfare non-negative (see (@warning Ci-tation strutt on page 7 undefined)) and so 0 RP 1.


8/17

8 B. Zadler et al.

0 20 40 60 80 100

kHz

QP= 500, Q

S= 250

QP= 50, Q

S= 25

QP= 10, Q

S= 5

Figure 2. The theoretical resonance response of a genericsoft rock sample for high, moderate, and low values of com-

pressional and shear Q.

20 30 40 50 60 70 80 90

kHz

0

0.2

0.4

0.6

0.8

1

RP

sensitive only to VP

sensitive only to VS

extensionaltorsionalnon-axisymmetric

Figure 3.Relative sensitivity to compressional and shear ve-locity, RP, defined in equation (11), for the lowest-frequency

modes of the generic soft rock model

IfRP= 0 for some particular mode, then the reso-nance frequency of that mode is completely insensitiveto changes in the samples VP. The sample motion ispurely shear and the frequency of the mode is exactlyproportional to the samples shear velocity. The axisym-metric torsional modes are the only modes for which the

above are exactly true, although there are many modesfor which RP 1.

If, on the other hand, RP = 1 for some mode,then the resonance frequency of that mode is completelyinsensitive to changes in the samples VS . The elasticcylinder has no modes for which this is true, although itprobably becomes asymptotically true for certain fami-lies of modes as frequency becomes infinite. It turns outthat we have an interest in looking for modes at reason-able frequencies for which RPis as large as possible.

0 50 100 150 200 250 3000

0.2

0.4

0.6

0.8

1

extensionaltorsionalL = 1L = 2L = 3L = 4

length/dia ~ 2.5

Figure 4. RP defined in equation (11) for the generic softrock model with an inset drawing showing the samples pro-

portions. The principal plot is the same as that in Figure 3but covering a broader frequency range.

The values in figure 3 are all remarkably close to 0.None of these modes is more than slightly sensitive toVPand we cant expect to get robust estimates ofVP.At a given measurement error level, the errors in VP willbe twenty or more times those in VS.

Figure 4 shows RPout to about 300 kHz. The firstmode substantially sensitive to VPis a flexural mode atabout 150 kHz, for which RP > 0.4. Note that if wewanted to use this mode in a measurement we wouldhave to extend the upper frequency range of our mea-surements. Going to higher frequencies may bring ad-ditional problems: the number of resonances increasesrapidly with frequency and mode identification prob-lems are likely to increase, particularly in the presence

of substantial attenuation.We can do better without going to higher frequency,

however, by changing the sample geometry. Figure 5showsRPfor a model identical to the one used above ex-ceptthat it has twice the diameter. In this case the firstP-sensitive mode is a flexural mode at about 80 kHz,about a factor of two lower in frequency than for theoriginal sample geometry.

Figure 6 shows RP for a model greater by yet an-other factor of two in diameter. In this case the firstP-sensitive mode is at about 60 kHz. But notice thatthe numberof modes in any particular frequency bandis increasing rapidly as we make the sample wider. As wementioned above, the increased crowding quickly com-plicates mode identification problems, particular in thepresence of dissipation.

7.2 Excitation and Observation

In this section well look at a few numerical resultsthat show how transducer location plays a central rolein determining which resonances are observed. Figure2 showed that moderate amounts of attenuation could


9/17


0 50 100 150 200 250 3000

0.2

0.4

0.6

0.8

1

extensionaltorsionalL = 1L = 2L = 3L = 4L > 4

length/dia ~ 1.25

Figure 5. RP defined in equation (11) for the generic softrock model with an inset drawing showing the samples pro-

portions. Similar to Figure 4 (preceding), but computed fora model with the same length and twicethe diameter as thegeneric soft rock model used for the preceding figure.

0 50 100 150 200 250 3000

0.2

0.4

0.6

0.8

1

extensionaltorsionalL = 1L = 2L = 3L = 4L > 4

length/dia ~ 0.62

Figure 6. RP defined in equation (11) for the generic soft

rock model with an inset drawing showing the samples pro-portions. Similar to Figure 4 (preceding), but computed for

a model with the same length and fourtimes the diameter

as the generic soft rock model used in figure 4.

make it impossible to discern all of the resonances in aparticular frequency interval. A possible tool for solvingthis problem would be the use of transducer arrange-ments that favor particular modes.

Figure 7 shows the computed response for thegeneric soft rock sample (Section 3) for two transducer

arrangements for moderate values ofQ. The importantpoint to note is that there are modes that appear promi-nently in the edgespectrum that are missing from theendspectrum. The great simplicity of the latter curveis one of the strengths of the resonant bar technique.

Figure 8 reproduces the two curves of figure 7 inthe top two plots and adds two more plots showing theeffects of making the sample, respectively, twice andfour times greater in radius. As the sample becomeswider, the spectrum undergoes two important changes.

0 20 40 60 80 1000

5e-08

1e-07

1.5e-07

2e-07

2.5e-07

3e-07

endedge

QP= 100, Q

S= 50

Figure 7. The theoretical resonance response of our generic

soft rock sample for the edge and end mounted transducerconfigurations.

20 40 60 80 100

Figure 8. The theoretical resonance response of a generic

soft rock sample for several transducer arrangements andsample shapes

First, bar-wave approximation breaks down, as we cansee by the loss between the second and third plots ofthe simple regularly-spaced peaks of the narrow sam-ple; this change is good because some of the changedmode types are more sensitive toVPthan the bar modeswere. Hand-in-hand with that, however, is much greater

spectrum complexity, as seen most dramatically in thebottom curve. In the presence of attenuation, spectralcomplexity makes it much harder to interpret the ob-served spectrum in terms of specific modes.

At this point, we think these results tells us that de-veloping the ability to get VP and VS in a single exper-iment will probably involve a balance between sampleaspect ratio and somewhat more sophisticated trans-ducers. We do not yet have a global solution to theresonance experiment design problem.


10/17

10 B. Zadler et al.

46000 47000 48000 49000 50000

frequency (Hz)

0

0.2

0.4

0.60.8

1

amplitude

0

100

200

phase

Figure 9.Amplitude and phase of an isolated mode in gran-ite.

7.3 Data acquisition

A function generator (Stanford DS345) sends a 10 Volt(peak-peak) swept sine wave to the source transducer.

This signal is detected synchronously with a DSP lock-in amplifier (Stanford SR850), which digitizes the inputand reference signal with 18 bit precision, directly mea-suring the in-phase and quadrature components. Am-plitude and phase are calculated from these. Figure 9shows a typical amplitude/phase response for an iso-lated mode measured in a granite core. We see the phase shift associated with the arctangent function asthe frequency passes through the resonance.

To limit loss of energy and control the humidityin the bench top measurements we place the sample ina vacuum chamber held at approximately 160 mbar atroom temperature, 23 C. In the theoretical derivationsof section 4 we did not implement boundary and radi-

ation conditions. With the use of the vacuum chamberin the experiments we can stick to those assumptions.The vacuum chamber also yields a quieter environmentfor the experiment. For porous media it is well knownthat even a single monolayer of water adsorbed onto thepore space of the sample can have a dramatic influenceon the moduli.

8 RESONANT ULTRASOUND

SPECTROSCOPY VS. RESONANT BAR

Before moving into the experimental observations ofthis paper, it will be useful to mention the differences

between RUS and resonant bar techniques. The mostobvious difference is sample shape. For resonant barmeasurements, the sample is a bar with a length-to-diameter ratio of 10 or more. For RUS, the sample canbe any shape that can be modeled, though the mostcommon are spheres, cylinders and parallelepipeds. Thescale of the objects can also different. The full frequencyrange available for resonant bar techniques is similarto RUS (5 kHz - 200 kHz), however with a sufficientlylong resonant bar of rock (approximately 30 cm), normal

mode frequencies in the seismic frequency range (0-3000Hz) can b e measured. A specific mode catalog is excitedand studied for the sample (@warning Citation lucet91on page 10 undefined) (torsional and/or extensional asdescribed in Section 3.1). In RUS, all modes are allowedto be driven and measured for a more complete catalog

which can give a measure of anisotropy. This can beboth good and bad. Experimentally, it may beneficialto mount a sample on its edges to generate more modesand reduce losses due to system coupling. (For rectan-gular samples, mounting at a corner of three sides elim-inates the possibility of accidental degeneracy.) On theother hand, in-situmeasurements are more difficult forRUS since they involve strong coupling to a pressurizedmedium.

9 IDENTIFICATION AND ESTIMATION

OF NORMAL MODE FREQUENCIES

9.1 Mode identification vs. mode estimation

Before delving further, it is necessary to make a distinc-tion between mode estimationand mode identification.Mode estimation is the process of extracting a set ofparameters{pi} from a given resonance peak such thatthe peak could be reconstructed at a later time. Thisis done interactively by fitting a model to a peak orset of peaks. The equation should include at least thepeak frequency, Q, amplitude and some combination ofbackground terms. In our case, the Breit-Wigner modelis used (@warning Citation breit-wigner on page 10undefined). This does not associate the observed res-onance peak with the appropriate member(s) of the

eigenfrequency spectrum of the forward model. We termthat mode identification, which is also an interactiveprocess. Identifying modes should not be a problem fornicely homogeneous, well-made samples. However, forsamples with heterogeneity or low Q, poorly cut sam-ples or those with spectra containing a high eigenfre-quency density, mode identification can be difficult orimpossible.

9.2 Normal mode estimation

The eigenfrequency of a mode is not exactly at the max-imum of the corresponding peak in the spectrum. Anyone mode is assumed to have Lorentzian shape, and

multiple overlapping modes are considered to be a su-perposition of Lorentzians. An example of this shapecan be seen in Figure 9.

The model we used to fit the amplitudes of one ormultiple modes is due to Briet and Wigner (@warningCitation breit-wigner on page 10 undefined):

A(f) = B0+ B1(f f0) +Nn

Cn+ Dn(f f0)

(f fn)2 + 142n

(12)


11/17


where the displacement amplitudeAis a function of fre-quency. The parameters B0 andB1 describe a constantand linear background, respectively. The shape of eachmode is then constructed via 4 parameters: amplitudeCn, skewness Dn, eigenfrequency fn and full width athalf max n. Therefore, a simultaneous fit of n reso-

nance peaks will have 4 n+ 2 parameters. A non-linearleast squares fit of the data is then performed. Sincethe fit is sensitive to the initial parameters, they needto be chosen carefully and accurately. This reduces thechance of divergence in the fit and greatly reduces thecomputation time.

In an ideal resonance spectrum, the observed modeswould be spaced apart such that each mode appears iso-lated. In addition, each mode would be fit one at a time.Unfortunately, this is not always possible due to degen-eracy and the overlapping modes. Every mode in themeasured spectrum effects every other mode to someextent, the size of the effect being directly proportionalto the frequency spacing between adjacent modes. This

is mostly a problem when one peak dominates its neigh-boring peak in amplitude to the point where you see asmall hill on the side of a mountain. The fitting proce-dure will tend to favor the larger peak, so the estimatesof the parameters of the smaller peak will have a higheruncertainty.

We put a quantitative number on this effect byputting less (or zero) weight on the small peak duringthe inversion process if were unsure of its existence inthe forward model. One way to be fairly certain the peakis a true resonance is the following method.If there area number of prominent peaks (at least 4) before the onein question, the inversion can be performed on the setof prominent peaks with the questionable peak ending

the set. Giving a weight of one to all peaks except thelast will result in a fit to the first 4 high confidencepeaks, and the last zero-weight peak will be effectivelyguessed at by the inversion. If the frequency from theinversion seems reasonably close to the observed (butquestionable) peak, then it may be a true resonance.This is an iterative game that can be played under theright conditions, but to this authors knowledge, thereis no fool-proof solution for this problem. However, ifrepeated measurments are possible, the less prominentpeaks may become greatly enhanced in other data sets,giving rise to an increased confidence level in their trueexistence.

9.3 Normal mode identification

There are a number of reasons why a measured spec-trum cannot be mapped in a reasonable fashion to amodeled spectrum. It may be that the proposed modelis not able to reflect the observed measurement. For ex-ample, trying to fit an isotropic model to a sample withcubic symmetry. It may be that during modeidentifica-tion, a number of modes were identified incorrectly due

to a low quality factor. It may also be that the sampleitself is poorly cut or heterogeneous to the point that itcannot be modeled as homogeneous, in which case eithera heterogeneous model or a new sample is needed. In allbut the last case, a little more care and more numericaleffort can fix the problem.

As mentioned earlier, the lower the Q of the speci-men, the more difficult it is to identify consecutive peaksin any particular spectrum. This is often the case withsoft rocks. Figure 2 showed this effect best. In high Qmaterials the peaks are relatively narrow and splittingof degenerate peaks is easier to detect than in lower Qmaterials. However, if a rock is inhomogeneous, the de-generate peaks may split enough so that both (or all)are observable. In this case, the inhomogeneity saves youduring identification, but will likely come back to biteyou in the inversion uncertainty.

Another source of difficulty in eigenfrequency iden-tification comes from the physical coupling of the sam-ple to the apparatus. Slight differences in the posi-

tioning of the sample between the transducers alongwith the force applied by the transducers contribute tothis uncertainty. This error can be estimated by com-pletely mounting and unmounting the same sample sev-eral times. This was performed on a cylinder of Elbertongranite in Section 10.2. Although remounting the sam-ple multiple times to estimate error is quite feasible forbenchtop measurements, it is too time consuming formeasurements involving complex sample preparation.

9.4 Resonance Inversion

Missing a normal mode during measurement can be fa-

tal to the inversion if a shift between the measured andcomputed frequencies appears. The optimization mayconverge to the wrong model. To avoid this problem,we use sets of peaks we are confident in to performthe inversion. We can also use the weight (inverse datastandard deviation) wi in the objective function (cf.equation (8)). Lower weight would be associated withmeasured frequencies in which we have low confidence;although we will directly estimate the variance in themeasured eigenfrequencies later.

The first iterations of the inversion are done withlow order polynomials (N=5-7). These approximationscan rapidly predict the first few mode frequencies ac-curately but fail to accurately predict higher modes.

This allows one to obtain a rough idea of the truemodel. Higher order polynomials considerably increasethe computation time but yield a better resolution us-ing information of higher normal modes. Along withthe different iterations, new information becomes newa prioriinformation for the next iterations. One usuallystarts with an isotropic model. If a proper fit cannotbe achieved we lower the symmetry of the model used:cubic, then hexagonal, etc.

The following results show that we can achieve a


12/17


13/17


20 40 60 80

frequency (kHz)

0

0.2

0.4

0.6

0.8

1

normalizedamplitude

Figure 11. Spectrum of a cylindrical granite core (2.27 cm

diameter by 7.15 cm long). A sweep from 20 to 100 kHz witha 5Hz step was applied.

10.2 Granite cylinder

The anisotropy of particular granites due to microfrac-tures has been known for many years and is a good testfor our RUS methods (@warning Citation granite onpage 12 undefined). Previous work suggests that to beable to accurately measure anisotropy in a sample wewould need at least 20 consecutive resonant frequencies(@warning Citation ulrich02 on page 12 undefined).

We studied a 75 mm by 25 mm cylindrical core of El-berton granite and were able to measure and identifythe first 25 normal modes. An example of a measuredspectrum can be seen in Figure 11. In order to obtain theuncertainties in frequency for our 2, the granite sam-ple was completely mounted and unmounted 14 times.The mean values of the eigenfrequencies and their cor-responding uncertainties are shown in Table 2. The dif-ferences in frequency and amplitude can be seen in Fig-ure 13. Frequency uncertainty ranged from 22 Hz to801 Hz with a mean uncertainty of 258 Hz. Althoughthe amplitudes in Figure 13 seemed to vary greatly forspecific modes, the Q values of the modes were fairlystable. In Figure 14, the Qs range from approximately

100 to 200, with an average uncertainty of 41.A first attempt to match the first six normal modes

with an isotropic model was unsuccessful. Larger fre-quencies could not be fitted at all. We then triedanisotropic models with lower and lower symmetries.Finally with the use of an orthorhombic model (with 9independent parameters, c11, c22, c33, c23, c13, c12, c44,c66, c55) we can achieve a fit of 25 consecutively mea-sured resonances with a2 of 37. The resulting2 fromfitting all 25 peaks with each possible symmetry is:

f(obs) (MHz) f(pre) (MHz) (obs) (Hz) freq. #

0.031490 0.031490 50 4

0.032380 0.031680 58 5

0.033620 0.032540 29 6

0.049670 0.050080 31 8

0.052280 0.052230 43 9

0.059500 0.059400 22 10

0.061970 0.061960 27 11

0.064380 0.065220 39 12

0.069760 0.070050 75 13

0.071860 0.071800 801 14

0.072330 0.072310 327 15

0.075700 0.075750 376 16

0.076020 0.076070 411 17

0.078200 0.078300 396 18

0.079200 0.078640 401 19

0.079200 0.079256 406 20

0.081080 0.080978 789 21

0.084800 0.084833 393 22

0.085600 0.085088 575 23

0.086800 0.087480 394 24

0.087500 0.087512 569 25

Table 2.Comparison between the observed (f(obs)) and pre-dicted (f(pre)) normal mode frequencies with uncertainties

for an Elberton granite core.

symmetry # parameters 2

isotropic 2 3486cubic 3 4089

hexagonal 5 1793tetragonal 6 496

orthorhombic 9 37

The comparison between predicted and measured fre-

quencies can be found in Table 2. The inverted modelis c11 = 67.87 1.14GPa, c22 = 81.94 0.70GPa,c33 = 81.83 0.69GPa, c23 = 27.15 1.10GPa,c13 = 28.95 0.86GPa, c12 = 39.76 0.14GPa,c44 = 23.72 0.02GPa, c55 = 29.16 0.01GPa andc66 = 28.670.01GPa. In the plane perpendicular to thesymmetry axis, this yields qPwave speeds of 500321m/s and 455638 m/s in the c22 and c11 directions,respectively. This nearly 10% P-wave anisotropy is con-sistent with the ultrasonic measurements performed by


14/17

14 B. Zadler et al.

Malcolm and Scales (@warning Citation malcolm:02on page 13 undefined) shown in Figure 12.

In the anisotropic symmetry plane the S wavespeeds are 29865 m/s for the vertical polarization,269410 m/s and 29615 m/s for the horizontal polar-ization in the vertical and horizontal directions, respec-

tively. For orthorhombic symmetry, c44and c55decoupleand are not the same as shown by their fit values. Alsonote the difference in values for the other decoupled co-efficients:c12 and c23, c11 andc22. Here our

2 test wasused as a tool to decide the symmetry of the sample.Granite is known to be anisotropic. The way granite isextracted and cut in quarries takes advantage of thatanisotropy. This ensured, fortunately, that for such asample we have the anisotropic symmetry axis coincid-ing with the geometrical symmetry axis of the sample.This makes our assumptions valid in this case. Other-wise, we would have to invert for the crystallographicaxis as well.

It must be mentioned that it is generally accepted

that 5 resonances are needed for each elastic parame-ter being fit. This being the case, that means for or-thorhombic symmetry, we would need 45 (consecutive)observable resonances. For nearly any rock, this is allbut impossible. Of course, going to monoclinic symme-try would reduce our 2 as well, but we have indepen-dent measurements telling us to stop at orthorhombic.Data by Malcolm and Scales (@warning Citation mal-colm:02 on page 14 undefined) in Figure 12 shows thatthere is evidence of anisotropy in the sample (roughly10%), but it is not due simply to inhomogeneity. By ob-serving the surface waves traveling around the core inopposite directions, it can be said that the paths of thecounter-rotating waves are the same, therefore the ob-

served splitting in the arrivals of the two waves at somepoint must be due to inhomogeneity. This was measuredto be 5%. We conclude then, that while we are certain itis not necessary to reduce symmetry below orthorhom-bic, it is accepted that a portion of the error in the 2

is attributed to the sample being inhomogeneous.

11 CONCLUSION

RUS offers a promising technique for characterizing theelastic moduli of rocks. It fills a gap in existing spec-trum of measurement techniques between low-frequencystress-strain and ultrasonic delay-time measurements.

RUS allows one to infer the complete elastic tensor froma single measurement, without having to machine a sam-ple according to the assumed anisotropic symmetry.

To implement RUS, we first derived the forwardmodel to compute the frequencies and shapes of thenormal modes. The forward model is greatly simplifiedwith a polynomial approximation, which allows fornumerical evaluations. Further symmetry assumptionsyield a speed-up of the algorithm.

0.010

0.012

0.014

0.016

0.018

time(ms)

0 60 120 180 240 300angle

Figure 12.Laser ultrasonic measurement of P-waves travel-ing straight across a 5.5 cm diameter granite core. A pulsed

IR laser in the thermoelastic regime is used to excite elasticwaves which are measured with a laser Doppler vibrometer.The source and detector positions are antipodal on a line

through the middle of the sample. The sample is rotated 10degrees between measurements. For more details see (@warn-ing Citation malcolm:02 on page 14 undefined).

58 60 62 64 66

frequency (kHz)

0

0.2

0.4

0.6

normalizedamplitude

Figure 13. A set of 3 resonance peaks from 3 of 14 El-berton granite data sets. Amplitudes varied with each set

due to mounting. The average uncertainty in eigenfrequencyfrom 14 complete mountings and remountings was 258Hz .

This uncertainty does include error from misidentification of

resonance peaks during fitting.

RUS methods have been used extensively in the ma-terials science community to evaluate high-Q crystalsof small sizes. The method then yields an efficient wayto invert for elastic properties. In this paper we showthat this method may be applied to larger objects andalso objects that have small-scale heterogeneities suchas rocks. We have, so far, only studied clean, relatively


15/17


10 20 30 40 50 60 70 80 90

frequency (kHz)

0

100

200

300

Q

Figure 14.Q values for Elberton granite have an increasingtrend over the measured range. Higher frequencies had sligtly

higher uncertainties, or more variation in estimated Q fromspectrum to spectrum.

homogeneous rock samples, using normal modes whosewavelengths are large compared to the granularity ofthe sample. For samples with large-scale heterogeneity,it will presumably be necessary to perform a preliminaryanalysis (e.g., tomography) to characterize the sample.We have also analyzed the sensitivity of the resultingelastic moduli to transducer position, sample aspect ra-tio and frequency of the measurement. We have foundno globally optimal solution to the problem of experi-ment design. In practice it is important to know whichmodes most influence the parameters one is interestedin and how best to excite those modes by appropriate

choice of sample shape and transducer position.

ACKNOWLEDGMENT

We thank Lydia H. Deng for her help with the COOOLlibrary, Mike Batzle for his help in the sample prepara-tion and Gary R. Olhoeft for valuable comments. Thiswork was partially supported by the National ScienceFoundation (EAR-0111804), and the sponsors of theFluids III Consortium at the Center for Rock Abuse,Colorado School of Mines.

APPENDIX A: SYMMETRIES AND BLOCKDIAGONALIZATION

An elastic medium is characterized by its stiffness tensor(Cijkl). With the so-called Voigt notation (@warningCitation th86 on page 15 undefined), one can representthe medium by a 6 6 matrix, (cij) in accordance with

ij or kl: 11 22 33 32 = 23 31 = 13 12 = 21. 1 2 3 4 5 6

If one restrict oneself to orthorhombic or higher sym-metries the non-zero entries of the matrix (cij)are givenby

c11 c12 c13c12 c22 c23

c13 c23 c33 c44c55

c66

.

The corresponding symmetries are isotropic, cubic,hexagonal, tetragonal, and orthorhombic. Note that wealso assume that the symmetry axis are aligned withthe coordinate axis.

We can exploit the vanishing entries and symmetries ofthe resulting stiffness tensor Cijkl while reordering thematricesEand [see equations (4) and (5)]. Reorderingyields simplifications that can be exploited for numericalefficiency. Consider the entry

ik = Cijkl

V

jldV , (A1)

where = xl ym zn and = x

l

ym

zn

.Assume first the i = 1 and k = 2. Then the only

non-zero terms in the sum in equation (A1) are for(j, l) = (1, 2) and (j, l) = (2, 1). In both cases the in-tegrand is of the form

K xl+l1 ym+m

1 zn+n

.

Since we integrate with symmetric limits if such a termis non zero then

l even (odd) = l odd (even),

m even (odd) = m odd (even),

n even (odd) = n even (odd) .

If now i = 1 and k = 3, then the only non-zero termsin the sum in in equation (A1) are for (j, l) = (1, 3) and(j, l) = (3, 1). In both cases the integrand is of the form

K xl+l1 ym+m

zn+n1 ,

and if such a term is non zero then


m even (odd) = m even (odd),

n even (odd) = n odd (even) .

If now i = 1 and k = 1, then the only non-zero termsin the sum in in equation (A1) are for (j, l) = (1, 1),(j, l) = (2, 2), and (j, l) = (3, 3). In these cases to obtainnon-zero terms we require


m even (odd) = m even (odd),

n even (odd) = n odd (even) .

We obtain similar results for the cases (i, k) = (2, 2) and


16/17

16 B. Zadler et al.

(i, k) = (3, 3). There are 8 choices for the parities of l,m, and n. The previous relations give an equivalencerelation on the set

{(i, )| 1 i 3 , = (l, m, n) with l+ m+ n < N},

with 8 equivalence classes. If one takes two representa-tives of two distinct equivalence classes the associatedentry in is zero. Therefore, by grouping the represen-tative of the same classes we decompose the matrix into 8 diagonal blocks. It is straightforward to checkthat this yield a similar decomposition for the matrixEbecause of the Kronecker symbol in equation (4).

With such a block diagonal decomposition, instead ofsolving one large generalized eigenvalue problem, wesolve eight smaller ones. Since the computational com-plexity of the algorithm used is proportional to R3 (Ris the dimension of the problem) the gain in efficiencyis obvious.

A1 Low symmetry

In the case of symmetries lower than orthorhombic ablock diagonalization of the matrices and Edoes notapply. This simplification similarly breaks down whensymmetry axes do not align with coordinates axes (i.e.,sample faces). In such a case call (Rij),i, j= 1, 2, 3, therotation matrix transforming coordinates from the sys-tem associated to the symmetry axes to that associatedto the sample faces.. The stiffness tensor (cijkl) trans-forms according tocijkl =RiiRjjRkkRllcijkl. Sucha transformation breaks the structure of the matrix inthe sample-face coordinate system. This yields a higher

computational complexity for the modeling operation.Note that The rotation matrix involves three ad-

ditional parameters. These can be chosen to be Eulerangle as in (@warning Citation gold on page 16 unde-fined). To determine symmetry axes one has to invertfor these three additional parameters. Note also that thedetermination of symmetry axes is restricted to sampleswith distinct face orientation. In the case of a sphericalsample the additional parameters cannot be invertedfor. In the case of a cylindrical samples with a circularbase the axes of symmetry can only be inverted for up toa rotation around the sample symmetry axis (@warningCitation backus on page 16 undefined).

The inversion procedure we have followed assumes

some a priori knowledge of the symmetry axis for theanisotropy of the sample. Another approach that onecould follow is to invert for the general elastic tensor,i.e. the 21 elastic constants. In such a case inversionrequires knowledge of a large number of resonance fre-quencies. Assuming that one can invert the 21 elasticparameters, one would then want to know what typeof anisotropic symmetry the sample exhibits. In otherwords the question is: Can one guess the anisotropicsymmetry with the elastic tensor given in any orthogo-

nal coordinate system? Backus (page 649) propose onestrategy to answer this question. His approach is basedon the representation of the elastic tensor with harmonictensors. The elastic tensor,Cijklis first decomposed intotwo tensorsC= S+ A, one symmetric Sijkl, one asym-metric, Aijkl. S can be represented with the help of

three harmonic tensors, one of order 4, one of order 2,one of order 0. S can be represented with the help oftwo harmonic tensors, one of order 2, one of order 0. Inthree dimension the action of a harmonic tensor H(q) oforder qcan be written as

H(q)(r) = A

1q

a() r + r2P(q2)(r), (A2)

where r is the position, a(), 1 q is a uniqueset of directions, A is a scalar, r = |r|, and P(q2) is ahomogeneous polynomial of order q 2. Backus showsthat the symmetry properties ofH(q) coincide with thatof the set of directions a()), e.g. H(q) is rotation in-variant with respect to some axis if and only if the setof directions a() is so too. Hence one can decomposethe five Harmonic tensors according to (A2) and obtainfive bouquets of directions. The common symmetry ofthese bouquets is then that of the elastic tensor C.

In our inversion framework, one would thus havethe geometrical mean to obtain the symmetry axis. Be-cause of uncertainties in the data and hence in the in-verted components of the elastic tensor one would onlybe able to make a guess. One can then however trans-form the elastic tensor Cinto the system of guessed sym-metry axis. In such a system some entries of the tensorshould vanish, at least within the uncertainties. Such anobservation can then confirm the symmetry type of the

elastic medium. With an anisotropic symmetry in mind,one could now proceed with the main method discussedin this paper and having fewer parameters to invert for,sharpen the estimates of the elastic constants.

APPENDIX B: EVALUATION OF

POLYNOMIAL INTEGRALS

B1 Rectangular parallelepiped

As far as computing integral (7), the case of a rectan-gular parallelepiped is easiest. We know that if one of l,m or n is odd then IVvanishes, otherwise we have

IV = 8dl+11 d

m+12 d

n+13

(l+ 1)(m+ 1)(n+ 1) .

B2 Cylinder

Again, if either l, mor n is odd then IV is zero. If theyare all even, then IV is equal to

IV = 2dl+11 d

m+12 d

n+13

n+ 1

10

rl+m+2 dr

20

cosl sinm d


17/17


12d

y

x

z

y

x

z

32d

2d2

32d

d1

d1

d2

d2

3d

y

x

z

Figure A1. Dimension associated with the different shapes used in our work: rectangular parallelepipeds, spheroids, cylinders.

after integrating overzand applying the change of vari-ables, x = d1r cos , y = d2r sin . The second integralis the integral of powers of trigonometric functions over

a period and is equal to (@warning Citation gr65 onpage 16 undefined)

2 (l 1)!! (m 1)!!

(l+ m)!! ,

where

(2a)!! = 2a(2a 2)(2a 4)...(2) = 2aa!,

(2a + 1)!! = (2a + 1)(2a 1)(2a 3)...(1), and

0!! = (1)!! = 1.

which gives

IV = 4

dl+11 dm+12 d

n+13

(n+ 1)

(l 1)!! (m 1)!!

(l+ m+ 2)!! .

Note that the form ofIV shows the cylindrical symme-try of the system.

B3 Spheroid

Here also, IV (equation (7)) is only non-zero for l, mand n even. In the spheroid case, we apply the changeof variables,x = d1r sin cos , y = d2r sin sin , z=d3r cos . This yields the integral

IV =dl+11 d

m+12 d

n+13

1

0

rl+m+n+2 dr

0

sinl+m+1 cosn d

2

0

cosl sinm d

= dl+11 d

m+12 d

n+13

l+ m+ n + 3

0

sinl+m+1 cosn d

20

cosl sinm d .

The second integral is equal to (@warning Citationgr65 on page 17 undefined)

2 (l 1)!! (m 1)!!

(l+ m)!! ,

whereas the first is (@warning Citation gr65 on page17 undefined)

2 (l+ m)!! (n 1)!!

(l+ m+ n+ 1)!! .

Therefore, we obtain

IV = 4 dl+11 d

m+12 d

n+13

(l 1)!! (m 1)!! (n 1)!!

(l+ m+ n+ 3)!! .

Note that the form ofIV shows the spheroidal symme-try of the system. Other geometrical shapes and valuesof IV are given in Vissher et al. (@warning Citationvi91 on page 17 undefined).

resonant ultrasound spectroscopy - theory and application

Documents