JOURNAL DE PHYSIQUE Colloque C10, suppl6ment au n012, Tome 46, dgcembre 1985

page C10-31

DECONVOLUTION ANALYSIS TO DETERMINE RELAXATION TIME SPECTRA OF INTERNAL FRICTION PEAKS

J.R.

COST

Los Alamos National Laboratory, Los Alamos, N.M. 87545, U.S.A.

RCsumh- O n d t c r i t une nouvelle m t t h o d e d'analyse des pics d e f r o t t e m e n t i n t i r i e u r e n fonction d e la temperature. Elle p e r m e t d'obtenir u n e a p p r o x i m a t i o n d u spectre d e s t e m p s d e relaxation responsables d u pic. O n m o n t r e q u e c e t t e m i t h o d e , a p p e l i e analyse directe d u spectre (DSA), s'applique I d i f f i r e n t s types d e spectres: (i) elle fournit des a p p r o x i m a t i o n s d e certaines formes d e spectre q u i reproduisent la position, l'amplitude, la largeur e t l a f o r m e avec u n e bonne p r i c i s i o n , q u i est d e l'ordre d e 10% (ii) elle ne c o n d u i t p a s I d e s a p p r o x i m a t i o n s a l i a t o i r e s d e la forme spectrale des pics. Abstract - A new m e t h o d for analysis o f a n i n t e r n a l friction vs t e m p e r a t u r e peak t o o b t a i n a n approximation of t h e s p e c t r u m of relaxation times responsible for t h e peak is described. T h i s method, referred t o a s direct spectrum analysis (DSA), i s shown t o p r o v i d e a n accurate estimate o f t h e d i s t r i b u t i o n of relaxation times. T h e m e t h o d is validated for various spectra, a n d i t is shown that: (i) I t provides a p p r o x i m a t i o n s t o known i n p u t spectra which replicate t h e position, a m p l i t u d e , width a n d shape with good accuracy (typically 10%). (ii) I t d o e s n o t yield a p p r o x i m a t i o n s which have false spectral peaks.
I - Introduction T h e problem of analyzing i n t e r n a l friction vs t e m p e r a t u r e peaks t o d e t e r m i n e t h e distribution o f relaxation t i m e s responsible for t h e peak has been a difficult o n e for m a n y years. T h e t r a d i t i o n a l method of analysis has been t o a s s u m e various forms for t h e relaxation t i m e spectrum a n d then choose t h e best f o r m based u p o n goodness-of-fit t o t h e data. T h e problem with t h i s m e t h o d i s t h a t none of t h e trial forms for t h e spectrum m a y be close t o t h e t r u e distribution. I t would be preferrable t o directly analyze t h e d a t a t o obtain a n a p p r o x i m a t i o n of t h e spectrum, t h u s avoiding unnecessary assumptions. Such a method, referred t o a s direct spectrum analysis (DSA), has been developed a n d validated for approximating relaxation t i m e spectra for processes with first-order kinetics / l / . With t h i s method, nonlinear regression least-squares is used t o unfold the integral e q u a t i o n for exponential decay with a d i s t r i b u t i o n of relaxation t i m e s a s t h e t i m e c o n s t a n t s within t h e integral. I n a later paper 121, t h e s a m e m e t h o d was applied t o t h e Debye function a s t h e kernel. T h i s allowed analysis o f i n t e r n a l friction (or dielectric relaxation) peaks when measured a s a function of frequency. I n t h e present paper these m e t h o d s a r e extended t o t h e measurement of i n t e r n a l friction vs temperature peaks.

I1 - Method
T h i s method has some m i n i m a l requirements of the data. First, i t is necessary t h a t t h e internal friction vs t e m p e r a t u r e d a t a h a v e had background subtracted. I t is i m p o r t a n t t h a t this subtraction be d o n e with t h e highest possible accuracy because t h e DSA m e t h o d t e n d s t o find c o n t r i b u t i o n s t o t h e relaxation t i m e spectrum for a n y i n t e r n a l friction which m a y be present o v e r a n d above t h a t for a given peak. Second, i t i s necessary t h a t there b e roughly 20 o r m o r e d a t a points; a s h a s previously been shown /1,2/, t h e resolution capabilities of t h e method increase with b o t h n u m b e r a n d accuracy of t h e d a t a points. T o discuss t h e method we will use t h e n o t a t i o n a n d development of Nowick a n d Berry / 3 / . We consider d y n a m i c experiments i n which stress i s applied periodically a t frequency o s o t h a t for a n anelastic solid t h e r e will be a phase lag of t h e strain behind t h e stress. T h e angle o f t h i s phase lag is cp, t h e loss angle, a n d the compliance is described by two d y n a m i c response functions, a real p a r t , J , ,

Article published online by EDP Sciences and available at http://dx.doi.org/10.1051/jphyscol:19851006

(210-32

JOURNAL DE PHYSIQUE

and an imaginary part, J2. The internal friction is given by tan ( P = J ~ / J.IFor a standard anelastic solid /3/ responding to periodic stress, these two dynamic response functions are given by

where GJ=J,-J., A=6J/J, is the strength of the anelastic relaxation and J, and J. are the relaxed and the unrelaxed compliances, respectively. Because the internal friction is measured as a function of temperature, certain corrections for temperature-dependent parameters are required. Nowick and Berry again provide an excellent discussion of these temperature corrections 131. First, since we want our approximation o f the relaxation time spectrum a t some constant temperature, Tdi*,, we must correct w and A to this temperature. For implementation of these two corrections, the reader is referred to Weller, et a1 /4/. Correction of A to Tdirtinvolves a Curie-Weiss temperature dependence. Thus, i t is required that a value be chosen for the critical temperature, T,. Since the value may not be known, i t is important t o d o the analysis using upper and lower limiting estimates of this parameter. Analysis involves making two approximations both of which are valid for the usual experimental condition that tan cp<<l. First, we consider J l J. so that tan cp*J2/J.. Second, we consider that the T which is measured (the relaxation time a t constant stress) is equal to the average T (the one in the above equations) /3/. We now proceed to obtain our approximation of N ( l n ~ ) ,the spectrum of relaxation times at Tdlrt.We first speci'fy the upper and lower T limits for the spectrum. For the first attempt a wide range is chosen; with later tries this is adjusted keeping the range sufficient to show the tail regions of the spectrum. We next divide the range of 1og10.rinto n bins of equal width and designate the relaxation time of the ith bin as T , , the midpoint value (10g~o.rscale) of the bin. The number of bins chosen depends upon the resolution desired. I t is typically 10<n<100, hut with the constraint n<=m, where m is the number of data points. The crux of the DSA method is to make a sum approximation for the integral (see Ref. 2) which contains the function N(ln T). This gives

where A, is the relaxation strength i n the ith bin. Eq. (3) is the same as given by Nowick and Berry 131 for multiple relaxations with discrete relaxation times. Thus, using this equation is the same as considering the relaxation to be made up of n different discrete relaxations all of which are equally spaced in log T. The .ri in Eq. 3 will exhibit a temperature dependence according to the Arrhenius relation, T=ro exp(Q/kT), where 70 is the pre-exponential factor, Q is the activation energy and k i s Boltzmann's constant. For a given measurement of the internal friction a t some temperature other than Tdlot,the .ri need to be corrected to their values at Tdist. This is done using either of two models for the distribution of relaxation times as discussed by Nowick and Berry 131. The first model considers that the distribution is only due to a distribution of activation energies and that TO is constant. The second model is just the opposite of the first, i.e., there is only distribution of To's and Q is a constant. We will use the first of these models in the discussion which follows, although the second has been used and found to provide good spectral estimates. After choosing n and the 7 limits for the spectrum, we continue as follows: (i) Choose Tdirt; typically this will be the temperature of the peak. (ii) Input the experimental value for w. (iii) Provide a n estimate of T o . (iv) Estimate the total relaxation strength, A. Typically, this is estimated as twice the ; : , . T o start the iteration process we set the Ai, the relaxation strength in the ith bin, peak height, Q all to the same value, Ai=A/n. The nonlinear regression least-squares method is now ready to be used. I t will iterate adjusting the Ai to minimize the least-square differences between the calculated and the experimental points. We terminate the calculation by either choosing a desired number of iterations or by a statistical evaluation of convergence.
111. - VALIDATION

The validation method is one of generating an artificial set of internal friction data from a known relaxation time spectrum, adding random (Gaussian) error with a desired standard deviation, and then applying the DSA method to the data t o test how well the input spectrum is recovered. In the examples which follow, the input distribution of relaxation times was either a single lognormal or a

combination of several lognormal (Gaussian in log,,r) distributions. Fig. 1 shows a n example of such a computer-generated i n t e r n a l friction peak plotted vs reciprocal temperature. T h e internal friction is normalized a s t a n q / A . The d a t a a t t h e various temperatures have been calculated from t h e Arrhenius relation with a n activation energy, Q=1.4 eV a n d preexponential, ~ ~ = 1 0 S. - 'T~ h e 50 d a t a points have been generated for w=1.0 s from a single lognormal distribution centered a t r,=1.0 s and with width P=1.0. R a n d o m fractional error with standard deviation o=0.01 has been added t o the d a t a . Fig. 2 shows t h e histogram for t h e s t a r t of t h e nonlinear regression analysis of t h e d a t a i n Fig. 1. Each of t h e 40 bins has equal spectral a m p l i t u d e such t h a t t h e area of t h e histogram is unity (on a In r scale). The spectral limits f o r t h e calculation h a v e been chosen t o cover three decades with t h e center a t roughly r = I / o . Fig. 3 shows o n t h e s a m e axes t h e histogram a p p r o x i m a t i o n of the i n p u t spectrum for Tdlst=500 K after 500 i t e r a t i o n s using DSA a n d assuming t o be 10-l4 S. At the t o p of t h e figure the scale is for the equivalent activation energy. I t may be n o t e d t h a t t h e histogram a p p r o x i m a t i o n agrees well with t h e original i n p u t distribution which is shown a s a solid curve i n t h e figure. A lognormal fit t o t h e histogram gives r,=1.02 s a n d P=1.03 i n good agreement with these parameters for t h e i n p u t spectrum. Validation was also done for i n p u t spectra with a single lognormal d i s t r i b u t i o n , but for values of P ranging from 0.25 t o 3.0 a n d r a n d o m fractional e r r o r with o=0.01. As with t h e result i n Fig. 3, t h e DSA method was found t o p r o v i d e a good a p p r o x i m a t i o n o f t h e i n p u t spectra o v e r this range of p. T h e capability of t h e method t o give good spectral approximations for symmetrical d i s t r i b u t i o n s with widths ranging f r o m t h a t of a nearly discrete relaxation (P=0.25) t o d i s t r i b u t i o n s covering u p t o five decades i n r i s encouraging. T o examine t h e recovery of asymmetrical spectra, d a t a were generated from multiple, but overlapping, lognormal spectra of varying widths. These spectra gave overlapping i n t e r n a l friction peaks which would be expected t o be very difficult t o deconvolute. An example of such a complex peak i s shown in Fig. 4 a s a broad i n t e r n a l friction peak made u p a s t h e s u m o f four separate peaks shown a s dashed curves. Again, 50 d a t a points h a v e been used for o=1.0. Fractional e r r o r with s t a n d a r d deviation a=0.01 h a s been added t o simulate experiment. T h e f o u r separate peaks a r e eachgenerated from a lognormal distribution. These vary i n fractional strength a n d width a n d are centered roughly a factor of four a p a r t a s shown i n Fig. 5.
'

05,

,
INTERNAL FRICTION PEAK INPUT LOGNORMAL OlSTRlBUTlON 8-10 Tm<1O S m ~ 5 DATA 0 POINTS FOR FRACTIONAL ERROR=O.OI Y .IOI-l

g
LL

3 t

:
-

STARTING HISTOGRAY n~108111S

OP\
'

G-

; ;

r,.003t

4 W a
LL B

a:I.~~v
.10.14s
-

$F

Z 0.3 -

2 0 5 -

3 A
a

:
AREA. 112

303

10"

100 RELAXATION T I M E . r ( s )

1 0 '

Fig. 1 - I n t e r n a l friction v s temperature d a t a generated from a lognormal distribution t o check the capability of t h e DSA m e t h o d t o recover t h e i n p u t spectrum. Fig. 2 Starting histogram for calculation of t h e DSA approximation to t h e i n p u t spectrum. T h e vertical dashed lines a r e t h e spectral limits. Fig. 3 - I n p u t relaxation t i m e spectrum a n d o u t put histogram approximation after DSA of d a t a i n Fig. 1.

1 . 0 ~ ~

:
j :

I3

; lWUT SPECIRIM.N<," 71
SINGLE LOGNORMAL B.IO

-,...
1.4

1.5

l6

WP"T nlSTOGRAM.~-.l

4081NS r.:0031 5 0 0 ITERATIONS

Td,,,

500K

1"'

C10-34

JOURNAL DE PHYSIQUE

The results of the calculation using the DSA method o n the data in Fig. 4 are also shown in Fig. 5. Here they may be compared with the input spectrum (solid curve) which defines the internal friction peak. The historgram approximation of the input spectrum is done a t Td,,,=500 K for 40 bins of T. The only assumption used for the calculation was t h a t .ro=10-14 S. I t may be seen that the DSA approximation shows very satisfactory agreement with t h e spectrum which was used t o generate the data. I t is particularly satisfying t h a t the analysis is able t o reproduce the four peaks which make up the input spectrum. Similar validations using i n p u t spectra of two a n d three peaks also showed this good resolving power. Many different kinds of spectra have been validated using the method described here. Importantly, these validations never gave false approximations t o the i n p u t spectrum, i.e., in no case d i d the results of the DSA show extra peaks o r miss peaks which were of reasonable magnitude ( - 10% of the total spectrum). The validation described in Fig. 5 demonstrates t h a t the amplitude, shape, width a n d location of the peaks are all reconstructed with acceptable accuracy. Only a minor difficulty is evident; this is f o r the location of the third peak. The histogram shows this peak t o be a t a relaxation time which is roughly 40% too high. This tendency was observed in other analyses, particularly when the i n p u t peaks were close together. Typically, the method resolves input peaks which are a factor of two or more apart, but the peaks tend t o be displaced away from each other a s shown i n Fig. 5 .

Fig. 4 - Internal friction data generated from 4peak input spectrum.

Fig. 5 - Input relaxation time spectrum and output histogram approximation after DSA of the data in Fig. 4. The a's are the fractional contribution of each lognormal spectrum.

V CONCLUSIONS

(a) A new method is described for analysis of an internal friction vs temperature peak t o obtain a n approximation of the spectrum of relaxation times responsible for the peak.
(b) I t is shown that this method provides approximations to known input spectra which replicate t h e position, amplitude, width a n d shape with good accuracy (typically 10%). (c) Validation of t h i s method for a wide variety of spectra shows that i t does not yield approximations which have false spectral peaks.

REFERENCES
/ l / Cost, J. R., J. Appl. Phys., 54(1983) 2137. /2/ Cost, J. R., Nontraditional Methods in Diffusion, ed. b y Murch, G., TMS-AIME, New York (1984). /3/ Nowick, A. S. and Berry, B. S., Anelastic Relaxations in CrystallineSolids, Academic Press, New York (1972). /4/ Weller, M., Zhang, J. X., Li, G. Y., Ke, T. S. a n d Diehl, J., Acta Met., 29 (1981) 1055. / 5 / Cost, J. R., proceedings of this conference.