Physica B 162 (1990) 344-352 North-Holland THE RELATION BETWEEN l/f NOISE AND NUMBER OF ELECTRONS F.N. HOOGE Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands Received 5 February 1990 The validity of the empirical relation for l/f noise, S,IR* = alNf, is studied. An alternative version with a factor l/A instead of the factor l/N is considered. N is the number of electrons and A the number of atoms. Experimental and theoretical arguments for the relation in its original form are put forward. 1. Introduction The electrical conductance of semiconductors and metals shows spontaneous fluctuations around the equilibrium value. One type of noise, which is always present, has a spectral power density that is inversely proportional to the frequency, hence its name, l/f noise. Twenty years ago [l] an empirical relation was proposed for l/f noise in homogeneous samples, that is s Rz- L-Y R2 Nf' where S, is the spectral power density of the fluctuations in the resistivity R, f the frequency and N the total number of electrons (or holes). The dimensionless parameter (Yturned out to be of the order of 10e3. For a long time it was thought that (Ycould be a constant [2]. The introduction of N was not based on a theoretical model. At that time we did not know what the electrons were doing while producing l/f noise. Even today that is a problem. But we may follow the naive reasoning that, whatever the electrons do, they do it independently. Then, for purely statistical reasons, we may expect the factor l/N. Considering relation (1) as the relative noise of an ensemble of N electrons, (Y is nothing else but a measure of the noise at a frequency of 1 Hz and normalized to one electron. There has always been severe scattering in 0921-4526/90/$03.50 0 the reported a-values. This was not too alarming so long as the deviating a-values could be blamed on faulty contacts, inhomogeneous material or ill-understood devices. But when silicon samples made by modern I.C. technology were studied, cu-values of quite different orders of magnitude were found. Modern technology means two things: (i) higher quality of the lattice, because of epitaxial growth techniques, and (ii) very small, sometimes even submicron, dimensions. Measurements on noise in samples of different size, made on the same wafer, showed that small dimensions brought no extra complications [3]. Measurements on semiconductor or metal samples whose degree of crystal lattice perfectness was changed showed a very systematic variation in (Y.The lattice was damaged by radiation or mechanical means and then partly repaired by annealing [4-61. The results of the many applications of relation (1) to metals [5-61 can be summarized in two statements. (i) The value of Q in metals decreases systematically with the degree of perfectness of the crystal lattice. (ii) The range of a-values is narrower than in semiconductors. The results of noise investigations in metals are often presented in plots where the vertical axis is in units of 10e3. All in all, the success of relation (1) is greater in metals than in semiconductors. Yet, in many publications on noise in metals, the 1990 - Elsevier Science Publishers B.V. (North-Holland) F.N. Hooge authors consider the fact that (Yis not a constant in the strict sense as proof of the absolute failure of relation (1). There is one well-understood reason why (Yis low in some cases [2]. There is good evidence that l/f noise is a fluctuation in mobility. These mobility fluctuations are in the contribution of the lattice scattering. The observed mobility /L is given by the relation 1 1 -=-++ CL Pl,tt 1 PX * When there are two scattering mechanisms: lattice scattering and another one, labelled x, relation (1) becomes s s R2 p2 R=l! a = Ff. If the noise is in the lattice scattering only, s l/&L= SllPllatt3 345 I 1 If noise and number of electrons (4) (5) from which it follows that (6) To summarize this part: LX,that looked like a constant at the beginning, is no constant, but a parameter. It can be lowered by the presence of scattering mechanisms other than lattice scattering. alatt is of the order of 10e3 to 10m4in normal materials. But, in perfect crystal lattices, (Ycan be very low, 10m6or lower. There is no explanation for this dependence on the lattice quality. 2. The factor l/N When we interpret noise data with the help of relation (l), we express the experimental findings as S,R-*f. This measured quantity is then written as cw/N. So long as the substitution of N led to constant or nearly constant values of (Y,we had the safe feeling that LYhad a physical meaning. Obviously, it was right to use the normalizing factor l! N. But since it appeared that cy may vary by more than a factor of 1000, one may well ask whether it makes sense to decompose S,Rm2f into two factors, (Yand l/N. The widely scattering values of (Ycan hardly be used as an argument to choose l/N as a parameter that brings order into experimental data. Therefore, without further discussion of the numerical value of LY, we now concentrate on the essential problem: Is there a factor l/N in the relation for l/f noise? The objections raised against the factor l/N are twofold: (i) statistical arguments, (ii) arguments of a more physical nature based on the long characteristic times needed for the low-frequency part of the spectrum. (i) There has been a long discussion on statistical arguments against the factor l/N. This discussion has been definitely settled 7 years ago [7-lo]. Making a theoretical model for l/f noise means making the calculations for an electron i which is representative of the electrons in the sample. One thus discusses the behaviour of this “average” electron, finds pi and Api, and no factor N appears in the analysis. But, in the experiments, one always observes a group of N electrons. The experimental findings on N electrons (e.g. fiN) can be expressed as properties of one electron. The essential steps are (7) (8) (G”)’ = $ = (slA*,)’ =$ i $ (Api)” . (APi)* (9) (ii) The more fundamental objection is that a factor l/N cannot be reconciled with the long times required for the low frequencies. Arguments of this type have often been brought 346 F. N. Hooge I 1 If noise and number of electrons forward and have been summarized by Weissman in a review paper [5], in which he concludes that it is absurd even to try a relation containing a factor l/N. Our assumption behind the factor l/N is that we add the uncorrelated effects of N independent electrons. If we could follow one individual electron we would find that it stays in the sample only for a short time. When it moves by Brownian motion it will stay in a sample measuring 1 cm in length for 0.1 s, since its diffusion coefficient D is of the order of lo-’ m2/s. This time interval of 0.1 s creates problems when we compare it to the low frequencies at which l/f noise is observed. The l/f spectrum has been observed even at 10m6Hz [ll]. We consider the l/f spectrum as the summation of Lorentzian spectra r( 1 + 0 2r2)-1 with r-values between 7, and T* and statistical weight g(r) dr = 7-l dT. The l/f spectrum is then found between the frequencies ril and ~1’) as follows from (IO) Now the problem is: how can the electrons that are present for only such short times contribute to fluctuations with those very long time constants? These arguments have been countered by pointing out that the slow processes are in the crystal lattice. We have good evidence that the l/f noise is in the lattice scattering [2]. Somehow, the cross-sections for the electrons fluctuate slowly with a l/f spectrum. There are always N electrons to probe the cross-sections of the lattice modes. We do not follow individual electrons; but the group of - on the average - N electrons that is continuously present, measures the slowly varying lattice properties. Individual electrons may move in and out; the lattice is permanent. However, it is not that easy to reformulate these verbal arguments in mathematical language. At the Budapest Conference [12] a crude model was presented that suggested that the normalizing factor should not be N but the number of atoms A. At the end of this paper an improved model will be presented that advocates again the use of N. Nevertheless, it is worthwhile trying A as the normalizing factor and comparing the results with those of normalization on N. 3. The factor l/A The choice of A as the relevant parameter can be justified by several arguments. (1) The experiments with samples of different sizes and the same dopant demonstrated not so much the dependence on N, but rather on the volume of the sample. Volume can be expressed by the number of atoms A. (2) Since the noise is in the lattice mobility, the fluctuating entitities are the modes of the longitudinal acoustic-lattice vibrations [2,13]. Their number equals the number of atoms. (3) With metals, it is difficult to estimate the effective number of valence electrons. Here one has always used A as an approximation of N. In order to avoid old (Y’Sbased on N and new (Y’S based on A we now introduce a quantity y, honouring the tradition followed in noise research on metals. The definitions of (Yand y are: S 2 = & Y A __=-= (Y N [eq. (l)] 3 5 x 1O22cm-3 n = -+ , (11) (12) The numerical value 5 x 1O22cm-3 is sufficiently accurate for most semiconductors and metals. This value is used for constructing the framework of fig. 1. In fig. 1 we have plotted the experimental results obtained in homogeneous materials. They were published since 1985. We included all published data without discrimination. Some older data are included, mainly from systematic investigations of the dependence of (Y on volume or on N. The values of (Y derived from the data of electronic devices are not included. Their derivation is much too indirect, requiring too complicated reasoning and too much interpretation. An earlier version of fig. 1 F. N. Hooge nN= 13 Id" I 1 if noise and number of electrons < lo** 1oz3 1O24 II n IO6 12 IO4 IO3 IO2 16' Fig. 1. a- and y-values of semiconductors and metals according to the relations (1) and (11): (0) Si; (0) others; (am) amorphous. was published in the proceedings of the Budapest Noise Conference [14], where also the references to the original publications were given. Two groups of materials are worth mentioning. (i) Semiconductors. We must keep in mind that a-values may be reduced by a factor of lo3 347 or so in high-quality silicon. When the quality of the lattice has been improved by annealing in a series of experiments, then we use the a-value of the lowest-quality material of this series. Several semiconductors other than silicon are represented in the plot. The quality of their crystal lattice is that of silicon before the I.C. era. Here we see a return to (Y2: 10e3 that was normal for silicon long ago. (ii) Metals. In old publications, a-values were calculated using the number of valence electrons per atom. Those old (Y’Swill now be called y’s in our new notation. The number of electrons that effectively take part in the conduction is smaller than the number of valence electrons. It is the group of electrons that are within a few kT of the Fermi level. For a calculation of CY,we must use the effective number of free electrons. Calculations on simple band models show that the reduction factor is about 20 [15]. Therefore we use n=1/20x5x1022cm-3=2.5X 1021cme3. If one believes that the 1 lf noise in metals and in semiconductors has the same origin then fig. 1 provides a strong argument for preferring (YlN to y/A. There are, however, good theoretical models for 1 lf noise in metals [5,6] that do not apply to semiconductors. That is a good reason to restrict ourselves to semiconductors. But even then, comparing the spread in (Yand in y, there is no need to reject N as the normalizing factor. 4. Experimental evidence for l/N The problem with fig. 1 is that the noise data were collected from samples with different degrees of perfectness, and therefore varying values of CLWe need experiments where N is varied in the same part of the sample, so that cr keeps the same numerical value at different values of N. There are two groups of publications that describe experiments where N has been varied in the same volume. (i) Injection of electrons in forward diodes, (ii) photoconductivity. F. N. Hooge 348 I 1 If noise and number of electrons (i) Forward diodes, Kleinpenning [ 161. In general one must prefer homogeneous samples to complicated devices, but the analysis of this simple device is so straightforward that we run no great risk by accepting it as good evidence. Now let us first consider a simplified model that shows all the essential features. We consider a p+-n junction in forward bias. The current Z injects holes into the N region. There they recombine in recombination time r. A concentration of holes builds up in the N region. The total number of holes P is given by z P -=_ 4 7’ (13) The diode shows l/f noise and shot noise a sI = 2 (14) Ff z + 2qz. The frequency-independent shot noise and the 1 lf noise are equal at a crossing point f, that lies in the region 100 Hz to 1 kHz. The crossing point f, is given by fx_az2 1 _L?Lx_a. P 2qz Zrlq 2qz (15) 27 ’ f, is independent of 1. This was confirmed by the experiments in which the diode current Z was varied by a factor of 1000. If a/P were replaced by y/A then f, would go with I. Kleinpenning [16] gave a detailed analysis of the forward diode. His results for different situations are presented in table 1. The a-values in the last column follow from his expressions for the corner frequency and his own or previous measurements by others. Here the presence of the factor 1 lN was demonstrated by the con- stancy of the corner frequency. There are other investigations on diodes that indirectly show that l/N is the correct normalizing factor; however, without explicitly giving the constant corner frequency. All studies lead to normal a-values 10-4-10-3. For Si diodes, see [17-201, for Cd,Hg,_,Te diodes, see [21,22]. (ii) Photoconductivity, Hofman and Zijlstra [23] An elegant way of varying the number of electrons in a given volume is provided by the “presistent photoconductivity effect”. Deep centres are present in Al,Ga,_,As. These, socalled, DX centres have a small capture crosssection. Therefore the optically excited electrons remain in the conduction band for long times. A homogeneous electron excitation is realized since the absorption is weak. These experiments resulted in normal a-values: (Y= 3.4 x 10m4 at a temperature of 79 K. 5. A model for mobility fluctuations Finally we present a simple model for mobility fluctuations. It is based on the treatment of lattice scattering as given in many textbooks. We follow here Chapter XI of Spenke’s book [24] that treats lattice scattering in great detail. As a new element we introduce fluctuations in this formalism of average mobility. This results in a proportionality factor l/N in the relative noise. In order to keep the expressions as simple as possible, we write explicitly only the quantities E, k, (kT), f and L3 = volume of the crystal with dimensions L x L x L. Other quantities and numerical constants will be included in the constants B, . . . B, and C, . . . C,, , since our only interest is whether we end up with 1 lN or 1 /A. Table 1 Type of p’-n diode Simplified (15) Long, g-r region Long, low injection Long, high injection Short f, Lyl27 al37 al87 al47 (aD/2w2) In(kJpw) (Yin units of 10m3 0.3 1.2 0.9 0.3 F, N. Hooge I 1 If noise and number Boltzmann’s constant will be denoted by k, which will not lead to confusion since it always occurs together with T; we shall write kT in parenthesis as (kT). We chose E = 0 at the bottom of the conduction band, therefore we have for the Fermi level E, < 0. The subscript e applies to electrons, p to phonons. The probability that an electron is scattered by one longitudinal acoustic phonon is P,: P, = B, (kT) L (16) , independent of k,, k, or the angle between them. Since the electron has approximately the same energy after the scattering, as before, the final k,-values of the electron lie on a sphere with radius kk. See fig. 2. The phonons that can scatter the electron will be called here relevant phonons. The relevant phonons have k-vectors that obey k; + k, = ka . (17) There are as many relevant phonons as there are allowed k-states on the sphere. The spacing there is 21r/L. Thus, the number of relevant phonons is given by =$ k2dk. 349 Here we have counted the number of discrete states in a shell with thickness dk, = 27rrlL. The probability P that an electron is scattered by acoustic phonons is (kT) = B,(kT)ka P=iP,=RBIT L dk, 1 = B3(kT)E;12 dE, = dE, , r(E) T(E) dE, = B,(kT)-1E,“2 (19) dE, . (20) An electric field E, in the x-direction gives a current i = ix. We use spherical coordinates; 13is the angle with the z-axis and 4 the angle with the x-axis in the xy-plane. One electron gives a contribution to the current 4 i=-vv,, L (21) j=+u,. (22) There are dN, electrons in a volume element dV, of k-space: dN, =f(k) 3 R = 4nkt dk, of electrons dV, $ . (18) k' The spacing of the two k-state. f(k) The volume rent density j= 3 is 27~/L; there is a factor 2 because possible spin orientations in each is the Fermi occupancy probability. element dV, contributes to the cur- v,f(k) dl/, -$ . (23) All electrons with the same Ikl contribute: 97 j= 2% Y$ If u, f(k)k%in 8 dtI dr$ dk . -$ , e=o r#J=o (24) Fig. 2. Relevant phonons. Such a phonon scatters an electron from kc into one of the states kf on the sphere with radius /CL. 2rm u, = v sin 8 cos t$ = - h k sin 13cos 4 , (25) F. N. Hooge I 1 if noise and number of electrons C, -$ k sin20 cos 4 0=0 +=o x f(k)k* -$ d4 do dk . (26) The occupancy f(k) of a state deviates little from the equilibrium Fermi factor f,(k) because of the influence of the field e,. Using the Boltzmann approximation, we obtain f(k) = f,(k) df(k) + q&x do =.6(k) + C, & =i,tk)+ ‘2 -(kT) UT(k) kT(k) fo k sin 13cos Q,T(k) . (27) The term f, does not contribute to the integral giving the current, but substituting the second term into eq. (26) yields B=O +=a c jk = k2+V sin38 cos’r$ d9 d+ 3L3(kT) x [$ fo(k)k2 dk] . 1 (28) The latter factor between square brackets is dN,, the number of electrons with the same value of lkl: dN, = L’ n= fo(k)k2 dk , and the fluctuations therein. This is the essential difference from the oversimplified model of ref. [12]. There we did not consider individual electrons. In that model it is as if an electron is smeared out over AIN states. According to the present model this gives an extra factor A lN in the noise, since the AIN states fluctuate together, because fractions of an electron cannot be scattered. The noise becomes proportional to (A/N)’ instead of to A lN. The final result 1 /A will therefore become l/N, as in the present treatment. We try to find an expression for Sjlj2. Therefore we calculate j from eq. (30) and Sj from eq. (31). It is no surprise that the first calculation leads only to well-known results. It is the second one that is important. Relation (30) together with (20) leads to jk = C&) = & (kTt\/E (k\)eeEIkTE2 da, ewkT C6 f(W4dk (32) r j = C, eEFikT e-EIkT 0 d (k?) = c, eWkT Comparing (33) (33) with (34) and (35) gives ( I_L ): (34) (29) (35) (p) - (T) - (kT)-3’2 @jk12 = [C, T3$!)]* [dNkl. (kt) . (36) The noise follows from (31): (31) The noise is in the scattering, expressed here by AT(k). We do not consider the number fluctuations, i.e. the fast fluctuations in the occupancies of the states. Separation of the factor dN, in (30) and (31) is the basis of our treatment of average mobility f(W6 dk , sjk = C, x eEFIkTy (37) (kT)3’2 emEIkT( g)“’ d(&) , (3% F. N. Hooge Sj = C, e E,IkT w)3’2 -yT-7 I 1 If noise and number of electrons Ias (E) 0 X emE,xT(~)5’2d(j$). (39) Sj is proportional to 1 lf; this does not necessarily mean that each S,(E) is proportional to 1 lf. But if this is so, and if S,(E) can be written as S,(E) = y EX(kT)Y , (40) then eq. (39) can easily be worked out: 351 tion volume [25]. So we might end up with some power of the volume, since R is proportional to volume (eq. (16)). In principle, it could be possible that the factor l/N we just found (eq. (43)) should be multiplied by AP. But there certainly is no risk that N should be multiplied by a factor containing some power of N. So, instead of the factor l/N we could have ApIN. However, all experimental evidence (e.g. fig. 1) shows that we could have either l/N or 1 /A, but not some combination of N and A like ApIN with p deviating appreciably from zero. Therefore, it seems correct to assume that S,(E) does not depend on L3. 6. Conclusions (41) =f C,C,, eEFikT y?(W (42) The magnitude of l/f noise in homogeneous samples can be expressed by the empirical relation s Ai=_ (Y From (33) and (42) we then obtain R2 s, _ c,, eE~‘k=(q3’2-x+y~-3 where N is the total number of electrons. (1) Theory does not provide arguments against the use of the normalizing factor l/N. (2) There are several experimental investigations that directly verify the dependence on 1 IN. (3) (Y is not a constant but a parameter. The values of (Yare of the order of 10m4 to 10e3 in normal materials. Perfect crystal lattices go together with very low a-values. .2 I e2EFlkT - f = C(kT)3-“+y f 1 77. (43) If S,(E) cannot be written as (40) we still find the factor 1 lN, provided S,(E) does not depend on L3. From theory - e.g. eq. (20) following from eq. (16) - and experiments, we know that T(E) does not depend on volume. Is this also true for S,(E)? In the discussion of the relevant phonons (around fig. 2) we considered Pi, the probability of an electron being scattered by one phonon (eq. (16)). We obtained the total probability P by summation over R phonons (eq. (19)). If we do the same for the fluctuations in P, and P of one electron, there is no way for f(k) to come in, and therefore not for N. But AP, could depend on L3 and there could be correlation effects to be taken into account in the summation over the R relevant phonons. There could be a correla- Nf' References PI F.N. Hooge, Phys. Lett. A 29 (1969) 139. 121F.N. Hooge, T.G.M. Kleinpenning and L.K.J. Vandamme, Rep. Prog. Phys. 44 (1981) 479. 131 R.H.M. Clevers, Physica B 154 (1989) 214. [41 L.K.J. Vandamme and S. Oosterhoff, J. Appl. Phys. 59 (1986) 3169. M.B. Weissman, Rev. Mod. Phys. 60 (1988) 537. ;:; N. Giordano, Rev. Solid State Sci. 3 (1989) 27. [71 K.M. van Vliet and R.J.J. Zijlstra, Physica B 111 (1981) 321. 352 F. N. Hooge / 1 If noise and number of electrons [8] A. van der Ziel and K.M. van Vliet, Physica B 113 (1982) 15. [9] F.N. Hooge, Physica B 114 (1982) 391. [lo] A. van der Ziel, C.M. van Vliet, R.J.J. Zijlstra and R.P. Jindal, Physica B 121 (1983) 420. [ll] T.G.M. Kleinpenning and A.H. de Kuijper, J. Appl. Phys. 63 (1988) 43. [12] F.N. Hooge, in ref. [14]. [13] T. Musha, G. BozbCly and M. Shoji, in ref. [14] p. 385. [14] 10th Int. Conf. on Noise in Physical Systems, Budapest (1989), A. Ambrozy, ed. (Akademiai Kiad6, Budapest, 1990). [15] T.G.M. Kleinpenning and J. Bisschop, Physica B 128 (1985) 84. [16] T.G.M. Kleinpenning, Physica B 98 (1980) 289. (171 T.G.M. Kleinpenning, Physica B 94 (1978) 141. [18] T.G.M. Kleinpenning, J. Vat. Sci. Techn. A 3 (1985) 176. [19] T.G.M. Kleinpenning, Physica B 154 (1988) 27. [20] P. Fang, L. He, A.D. van Rheenen, A. van der Ziel and Q. Peng, Solid State El. 32 (1989) 345. [21] X. Wu, J.B. Anderson and A. van der Ziel, IEEE Trans. Electron Devices ED-34 (1987) 1971. [22] L.K.J. Vandamme and B. Orsal, IEEE Trans. Electron Devices ED-35 (1988) 502. [23] F. Hofman and R.J.J. Zijlstra, Solid State Commun. 72 (1989) 1163. [24] E. Spenke, Elektronische Halbleiter (Springer, Berlin, 1965) chap. XI. [25] L.B. Kiss and T.G.M. Kleinpenning, Physica B 145 (1987) 185.
© Copyright 2026 Paperzz