Sei sulla pagina 1di 54

Rep. Prog. Phys., Vol. 44, 1981.

Printed in Great Britain

Experimental studies on

llf noise

F N Hooge, T G M Kleinpenning and L K J Vandamme


Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands

Abstract Experimental studies on l/f noise are reviewed with emphasis on experiments that may be decisive in finding the correct theoretical model for this type of noise. The experimental results are confronted with two theories: McWhorters surface state theory and Clarke and Vosss theory of local temperature fluctuations. The applicability of either theory turns out to be very limited. The validity of an empirical relation is investigated. Its application to electronic devices proves rather successful. Experiments show that l/f noise obeying the empirical relation, which we shall call a: noise, is a fluctuation in that part of the mobility that is due to lattice scattering. This review was received in December 1980.

0034-4885/81/050479+54 $06.50 31

01981 The Institute of Physics

480

F N Hooge, T G M Kleinpenning and L K J Vandamme

Contents 1. Introduction 1.1. Spectrum and correlation function 1.2. llfnoise 2. llfnoise due to fluctuations in the conductance 2.1. The experiment of Voss and Clarke 2.2. AC excitation of l/f noise 3. Configurations for measurement 3.1. Two-electrode arrangements 3.2. Four-electrode arrangements 3.3. Point contacts 4. Results for 01 4.1. Accuracy of C1,f 4.2. Some conditions for reliable a values 4.3. 01 values 5. Surface states and l/f noise 5.1. The McWhorter model 5.2. Discontinuous films 5.3. The McWhorter model and a noise 6. Temperature fluctuations 7. Fluctuations in number or in mobility? 7.1. l/f noise in thermal voltage 7.2. l/fnoise in Hall voltage 7.3. I/fnoise and lattice scattering 7.4. llfnoise of hot electrons 7.5. Empirical llfnoise source term 8. Lattice scattering causes 01 noise 9. Non-ohmic junctions 9.1. Single-injection space-charge-limited-current diode 9.2. p-n diode 9.3. Schottky barrier diode 10. MOS transistors 10.1. Oxide trap model 10.2. a-noise model 10.3. Comparison of the two models 11. Summary and conclusions References Page 481 481 483 486 487 488 490 49 1 49 1 492 494 494 494 496 497 497 498 501 501 505 506 509 511 511 513 514 517 519 521 522 523 524 526 528 530 530

Experimental studies on llf noise

48 1

1. Introduction
Generally speaking, electrical noise is a well-understood phenomenon. Its theory is well developed in the more abstract fields of thermodynamics and statistical mechanics as well as in the field of models that show what the electrons are doing in specific cases. Much experimental work on electronic devices supports these theories and models. In most cases the theory helps us to decide whether such devices can be improved by lowering the noise. For instance, generation-recombination noise can be reduced-at least in principle-by using cleaner preparation techniques to avoid trapping centres. On the other hand, thermal noise cannot be reduced by whatever method of preparation is used, based as it is on fundamental physical laws, and it is therefore unavoidable. In this well-ordered field of noise research, with agreement between experiment and theory almost everywhere, there is one notorious exception: I/f noise. This type of noise manifests itself as fluctuations in electrical conductance. It often causes trouble at low frequencies, let us say below 1 kHz, since its spectral density is inversely proportional to the frequency. The noise has been named l/f after this spectrum. In this review we restrict ourselves to l/f noise in electrical conductance, although l/f fluctuations are met with in geological, physiological and musical phenomena (Mandelbrot 1977). Some people have speculated on the common origin of all l/f phenomena. We shall not follow them but concentrate on conductance noise. The electrical llfnoise occurs abundantly in contacts, films, resistors and in nearly all semiconductor devices. In spite of such omnipresence and half a century of research on llfnoise, there is no accepted theory explaining such noise. Many models have been proposed, but each has its own specific experimental facts with which it is at variance. There is not even agreement as to whether all observed electrical l/f noise belongs to the same physical phenomenon. Perhaps there are several types of Ifnoise, requiring different theories to explain all the experimental facts. The main controversy is whether the noise is caused by generation-recombination processes between the conduction band and traps at the surface, or by some as yet not understood fluctuation in the lattice scattering occurring in the bulk. The first possibility is supported by a simple and very reasonable theory, the second by experimental facts. In consequence the simple question as to whether l/f noise is avoidable is not to be answered unanimously at present. Yes, avoid traps!, will be the answer of those who believe in a generation-recombination mechanism. No, not for a given number of electrons, say others who think it is fundamental to the scattering of electrons. In this phase of l/f research the emphasis must lie on simple experiments that directly decide whether certain theoretical proposals are right or wrong. This explains why experimental is the first word in the title of this review. But in order to avoid an unsystematic presentation of unconnected experimental facts we have grouped the facts around theoretical problems. We shall first sketch the application of noise theory to the other types of noise. Then we shall see where this leads to in the case of Ilfnoise.
1.I . Spectrum and correlation function When a quantity X shows noise we may write X ( t )=( X ) A X ( t )

482

F N Hooge, T G M Kleinpenning and L K J Vandamme

where ( ( A X ( t ) } z ) is constant for stationary noise. The symbol ( ) means averaging over a long enough time interval. Now A X ( t ) may be written in a Fourier series : A X ( t ) =Eat exp ( j 2 5 ~ h t+at* ) exp (- j 2nht) (1.2) where the coefficients ai are fluctuating amplitudes. The noise component measured at a frequencyh is then (AX,) =O (1 * 3) ( ( AXZ)~) = 2(aiaz *).

(1.4)

If one measures in a final bandwidth A A the Fourier components inside the bandwidth add up quadratically. The noise measured in a unit bandwidth is called the spectral density. This spectral density S x ( f ) is closely related to the autocorrelation function rpx(t), rpx(t) describes how, on the average, a deviation AX(t0) will decay. The autocorrelation function is defined by

px(t)=(AX(to) AX(to+ t ) ) . (1 5 ) The relation between qx(t)and Sx(f) is given by the Wiener-Khintchine relations:
Sx(f)=4
rpx(t)=

1;

q x ( t ) COS 2rff dt
COS

(1.6) (1.7)

1," S x ( f )

2 7ft d.6

This theory can be found in any introductory book on noise, e.g. MacDonald (1962), van der Ziel (1959, 1970) or Wax (1954). The Wiener-Khintchine relations (1 -6) and (1.7) are all we need here. They are used when making a physical model for an observed noise. The differential equation describing the transitions in the proposed physical model gives the correlation function. Its Fourier transform should be in agreement with the observed noise spectrum, with its shape as well as with magnitude. The simplest case, also the one most often met, is the exponential correlation function, corresponding to the Lorentzian spectrum. If, on the average, a deviation AX decays according to -dAX/dt = AXIT then the correlation function is
cpx(t)= <AX(to>AX(to+
t)> = <(AX(to)}2)exp (-

(1.8) (1.9)

t/~ =)( ( A X ) 2 ) exp ( - t/7).

The Wiener-Khintchine theorem then gives (1.10) This type of spectrum is called Lorentzian. Naturally one finds

J;
since

Sx(f)d f = < ( A X ) 2 )
1

(1.11) (1.12)

dx =tan-1 x.

( ( A X ) z ) is usually known from a very simple consideration. For instance, in the case of number fluctuations ( ( A N ) 2 )=N . When the fluctuation corresponds to the energy in a degree of freedom, the energy of ((AX)2) will be 3kT.

Experimental studies on llf noise

483

The linear equation (1 . 8) is a very general one. It does not say anything about what the electrons are doing, but only that, whatever they do, they do it independently. So, with the slight deviations that we are dealing with, this equation will appear quite often. In most cases we then find the Lorentzian spectrum according to equation (1.10). At low frequencies ( f ~1) g the spectrum is white, and at high frequencies ( ~ T 1) Bit is l/f2. There are three well-understood types of noise that are frequently seen in noise studies. These are thermal noise, shot noise and generation-recombination noise. The third one is also of theoretical interest for llfnoise models. (a) Thermal noise originates from the thermal motion of the charge carriers. In equilibrium situations this motion has an average energy of 4kT. The relaxation time is extremely fast, T N 10-12 s, so that at all attainable frequencies we measure a white spectrum with Sv =4kTR (1.13) or SI = 4kTG. (1.14)
(6) Shot noise is also a white iioise in the usual frequency range. It is found when a current of discrete particles leaves a cathode or passes a potential barrier. Its spectral density is given by SI = 2eI. (1.15)

This spectrum is white sincefn 4 1 at all attainable frequencies because of the very short transit time T I of the electrons. (c) Generation-recombination noise is of particular interest to us, since some l/f noise theories are variations on this theme. Consider a semiconductor with a number of identical trap levels. A fraction of them will be occupied by electrons. Since there is a continuing trapping and detrapping between the traps and the conduction band (or valence band) the number of trapped electrons, and therefore also the number of free electrons, will fluctuate. The transitions between traps and band are described by equation (1 . 8). The generation-recombination spectrum of the conduction electrons is given by
S,(f) = ((An')) 47 - -1 -t(277fT ) 2 '

(1.16)

1.2. l/f noise

After this short presentation of some elements of noise theory we can now approach the problem of llfnoise. This review will discuss its physical origin at length. But in order to keep the introduction short we now simply accept the experimental fact that the conductance of a semiconductor fluctuates with a l/fspectrum. The conductance fluctuations of an ohmic sample can be measured as voltage fluctuations when a constant current is passed through the sample or as current fluctuations when the voltage drop across the sample is kept constant: (1.17)
Clif is a number which is a measure of the relative noise of the sample. It is independent of the measuring conditions, such as current or voltage. It enables us also to directly compare measurements that were made in different frequency ranges and reported in the literature. It is a well-established experimental fact that, in many cases, spectra have been

484

F N Hooge, T G M Kleinpenning and L K J Vundamme

measured with Scocf-l.0. Often the exponent is - 1.0 2 0.1 over six or more decades of frequency. Spectra can be measured in the range 0.1 Hz to 100 kHz without any difficulty. In a few cases measurements have been extended down to 10-6 Hz. It must be realised that the spectrum cannot be exactly l/f in the whole range 0 <f< CO. In the first place, S- oc, when f - 0. Also, the integral of S-, CO at both frequency limits f= 0 and f= CO. Therefore, if the noise is stationary the spectrum must flatten below a certain low frequency and it must become steeper above a certain high frequency. Neither of these frequency limits has ever been observed. Two solutions are possible: ( U ) the noise is not stationary, or (b) we must measure at still lower frequencies. Regarding the last point it must be realised that it is difficult to obtain conclusive experimental results. Since 10-6 Hz corresponds to measuring times of the order of a month, extending the measuring time to a century adds only some 20 % to JSdf when the spectrum was originally measured over 12 decades from Hz to 106 Hz. An exact llfspectrum also makes it impossible to find the correlation function by Fourier transformation of the spectrum. It is true that, because of these mathematical arguments, exact llf noise cannot exist. But the importance of this statement should not be over-estimated. All difficulties occur at f = O or f= CO. Because of the logarithmic dependencies there are no practical difficulties at very low or very high frequencies. So a spectrum experimentally indistinguishable from llf can exist over many decades. When the spectrum is about l/f in a limited frequency range the correlation function goes approximately with In t in a limited time interval. This corresponds to a relaxation of AX according to the differential equation dAX --=A dt exp (BAX) with BAXBl. (1 * 18)

For symmetry reasons this suggests something like -dAXsinhCAX CT dt with (1.19)

as the equation analogous to equation (1 .8). No physical model has been inspired by these purely mathematical considerations. The usual way to make a model is to consider the llfspectrum as the summation of a large number of Lorentzian spectra. If the Lorentzian spectra have relaxation times between 71 and 7 2 and if their statistical weights are proportional to 7-1
g ( r ) dr = -___ - d r In 7 2 / 7 1 7

71

<7 < 7 2

(1.20)

then a llfspectrum is found in the frequency range 7 2 - 1 to 71-1

(1.21) The equation can be approximated as follows : (1.22) 1/2n72 <f< 1/2%-71 1/2n71<f 1 Sx(f)= <(W2> ___'

In 7 2 / 7 1

(1.23) (1.24)

Experimental studies on llf noise

485

We thus find a l/f spectrum over a wide frequency range. The integral of the approximated Sx(f) over this range is ((AX)2). The white and the l l f 2 branch hardly contribute to the total integral fromf=O tof= 00 if T z B ~ . Here we must make a remark that is important when comparing model calculations with experimentally observed spectra. The approximation (1 .23) in the lifregion shows that the noise is very insensitive with regard to T I and 7 2 . Only In 7 2 / 7 1 appears in the equation. Therefore one need not know 7 1 and 7 2 to calculate the spectrum. Even the vaguest idea about their ratio is sufficient. If the model has serious errors in the absolute values of 71and 7 2 they will not show up in the calculated spectrum. Numerical agreement between the calculated and observed spectrum-which is always measured in a limited frequency range including neither 1/71 nor l/.rz-can not be considered as a definite proof of the correctness of the model. If the model does not give absolute values for 71and 7 2 one usually takes In 72/71 N 10 (see, for example, 45.2). If it is desired to express the 7 distribution in the correlation function then we must take

In this type of theory the l/f noise results from the summation of a number of linear processes, all obeying equation (1.8). The amplitude distribution is then Gaussian. In order to avoid the somewhat artificial set of linear relaxation processes one might try to develop non-linear theories. In such theories the amplitude distribution need not necessarily be Gaussian. Therefore it is worthwhile investigating whether l/f noise has a Gaussian amplitude distribution. All experimental evidence suggests that if the distribution is non-Gaussian at all, it is so with only minor deviations from the Gaussian character. Very accurate measurements have been done by Voss (1978) who found l/f noise with a perfect Gaussian amplitude distribution. In his measurements such large amplitudes were measured that their relative probability was 10-7 of the probability at the top of the distribution curve. In some cases Voss found slight deviations from the Gaussian curve. Two explanations are possible. (i) Gaussian and non-Gaussian llfnoise exist. This would mean that there are several types of l/f noise. (ii) Vosss non-Gaussian l/f noise could be contaminated with some other noise. Burst noise, especially, is highly effective in spoiling a Gaussian distribution. In this type of noise the current jumps at random between two constant, nearly equal levels. This brings us back to the problem Is there only one type of l/fnoise? There is strong experimental evidence that at least the greater part of the llfnoise spectra measured have the same physical origin. In 1969 Hooge proposed an empirical relation for the l / f noise in homogeneous samples : (1.26) in which N is the total number of charge carriers and a is a dimensionless constant with a value of about 2 x 10-3. Cl/f was introduced into relation (1.17) to normalise results measured at different currents or at different frequencies. Further normalisation is possible assuming that each electron has its own independent contribution to the noise. Then the absolute noise of the sample must be divided by the number of electrons, and for the relative noise we find Clif= a/N. (1 .27)

486

F N Hooge, T G M Kleinpenning and L K J Vandamme

The surprising result was that 01 turned out to be a constant. There was no theoretical model behind this result, only the idea that whatever the electrons do, they do it independently. In this way of thinking N in the denominator is a trivial factor. It certainly does not mean that the llffluctuations must necessarily be fluctuations in number. The validity of relation (1 .26), the accuracy of a, etc, will be discussed extensively in this review. But we have introduced a here already to create some order in the vast amount of llfdata. Noise intensities will be expressed as a: values according to equation (1.26). Application of equation (1 .26) to homogeneous samples often gives a values of about 2 x 10-3. In inhomogeneous situations equation (1 -26)can be applied to volume elements. Integration procedures-treated in $3--make it possible to express experimental results in 01 values. Complicated electronic devices can also be treated in this way (see $99 and 10). We then find expressions for the current or voltage fluctuations from which a can be determined. In all cases where direct or indirect application of equation (1 .26) yields 01 values of about 2 x 10-3 we are obviously dealing with the same type of llfnoise. This l/f noise will be called a-type llf noise. In $8 we report on experimental results that suggest that 01 noise is a fluctuation in the lattice scattering of the electrons. In $5 another type of l/fnoise will be treated, i.e. McWhorters generation-recombination type. So at the end of this introduction we do not claim a priori that all Ilfnoise must be 01 type. We shall use the empirical relation (1.26) to order experimental data. If, in the end, such an ordering procedure proves to be successful, the summarising of the properties of a noise is easy. This gives then the conditions that have to be met by a theoretical model still to be made.
2. l/f noise due to fluctuations in the conductance

It has been demonstrated only lately that it is the conductance that fluctuates with a l/f spectrum. Before the decisive experiment by Voss and Clarke (1976a, b) this idea was generally assumed to be correct. Some authors were already studying the next question, whether l/f conductance fluctuations were caused by fluctuations in the number of the free charge carriers or by fluctuations in their mobilities. The general acceptance of conductance fluctuations was based on the following simple observations. When a constant current is flowing through an ohmic sample, I/fnoise is found in the voltage across the sample with a spectral density proportional to 1 2 ; when the voltage is kept constant the llfcurrent noise is proportional to V2. However, there were some worrying experimental facts. When such simple experiments were performed with AC current with frequency W O noise was sometimes found in frequency regions around 2w0 and higher harmonics and, most disturbing, at very low frequencies near 0 Hz. These extra noises were more than proportional to the normal llfnoise. Therefore it was thought that the extra noises and the llfnoise itself were generated by non-linear mechanisms. Is the llfnoise a kind of turbulence in the current? Such questions bring us to the more general problem. Does the current generate the noise or does the current only serve to measure the already existing conductance noise ? We shall now first present Voss and Clarkes demonstration of the conductance fluctuations. Knowing then that the llfnoise is in the conductance we shall discuss the various AC current experiments. We shall try to remove the paradoxes, showing that these AC experiments do not provide serious arguments against a conductance model for the l/f noise.

Experimental studies on llf noise

487

2.1. The experiment of Voss and Clarke


When a resistor R is in equilibrium at a temperature T we always find voltage fluctuations across the resistor S V ( f ) =4kTR. (2.1) This is called Nyquist noise, Johnson noise or thermal noise. Relation (2.1) has been very well established, both theoretically and experimentally. If desired equation (2.1) can be interpreted as the Brownian movement of the electrons with kT kinetic energy per degree of freedom. If a current flows through the resistor, there is no longer equilibrium and noise additional to (2.1) may be found, e.g. l/f noise. But in equilibrium we have Nyquist noise and nothing else. The Voss and Clarke experiment could be described as the measurement of l/f noise in the Nyquist noise. When R and T are constants SV is found according to (2.1): no l/f noise is observed. If R or T fluctuates with a l/f spectrum then-also according to (2.1)-we find lif noise in S v ( f ) if the measuring time is not infinitely long. If the measuring time is infinitely long-which should be the case if we really intend to measure the average value of (AV)Z-then of course T and R in (2.1) are also values averaged over an infinitely long time, so that no noise other than Nyquist noise will be observed in (AV)z. By saying that there is l/f noise in the Nyquist noise, we do not mean to correct relation (2.1) (which would be against thermodynamics) but that l/f noise in R will show up in the thermal noise. When studying 1, noise in the thermal noise we should be careful not to confuse noise from different sources. Since we measure the product RT, S v ( f ) will be a direct measure of temperature fluctuations. Temperature fluctuations will also influence SV( f ) since R depends on T. But under well-chosen conditions it will be possible to measure equilibrium fluctuations in R as fluctuations in V. The measurement is performed as follows (see figure 1). The equilibrium voltage fluctuations V(t) of the resistor R are amplified. The white thermal noise passes a bandpass filter with lower frequency limitfi and higher frequency limit fh. The signal is then squared and averaged over a time 7, which results in a signal P ( t ) . Then a Fourier analysis of P ( t ) is made, which gives a l/f spectrum. We express V ( t )as a Fourier series

where ( Af )-I is the length of the truncated time sample. According to (2.1) and (1.4)

% -U-%w
Iv

Pit)

488

F N Hooge, T G M Kleinpenning and L K J Vandamme

The squaring of (2.2) leads to


V 2 ( t ) = X (AnAn*+An*An)+ZZAd,*
?k
TL

exp [2nj(fn-fm) tl

n b

+ XZ An*Ant exp [ - 2nj (fn -fm) t I n m

(mfn).

(2.4)

Here we have omitted terms with sum frequencies because, in the end, they will make no contribution due to the following averaging process over the relatively long time T. The signal V ( t )is averaged according to
1 P(t)=7

V ( t )d t

with

nlAf$1/T$ A$
The contribution of the first term on the right-hand side of (2.4) is called P I . For T+ 0 0 we find

=4 k T R ( f h -fi).
If
7

(2.6)

is large but not infinite Pl(t) shows fluctuations because R shows fluctuations:

1 SP~=CI - / [4kTRCfh-fi>Iz. ~

(2.8)

The second and third term on the right-hand side of (2.4) give rise to white spectral contributions Pz. Straightforward calculations show that the lif part PI is greater than the white part Pa at frequencies below f c :
fc=

c 1lf ( f h -fi)

1 : c l / ff h .

(2.91

Relation (2.9) shows why it is so difficult to measure Ilfnoise in thermal noise. One needs very noisy samples. It is rare to find Cllf values higher than 10-7. The highfrequency limit fh must be less than the reciprocal RC time of the measuring equipment. A corner frequencyf, less than 1 H z makes the measurement of the l / f part of the spectrum difficult. This method of measuring l/f noise in samples in equilibrium was designed by Voss and Clarke, who performed successful measurements on semiconducting InSb films and on metal films of Nb. Similar measurements on carbon samples were later done by Beck and Spruit (1978) who then obtained relation (2.9), which also agreed with the earlier measurements of Voss and Clarke.
2.2.
AC

excitation of l / f noise

Since Voss and Clarke's experiment we know that lifnoise is a fluctuation in the resistance. Our discussion here of the AC effects is based on the knowledge we have now. Part of the AC effects had been observed before the Voss and Clarke experiment was done. Some authors gave interpretations of their experimental results that differ there-

Experimental studies on lif noise

489

fore from modern interpretations. We can accept all experimental results, but not all interpretations originally given. Ohms law states directly what will happen when an AC current with frequency f I flows through a sample in which the resistance shows l/f fluctuations : SI = P C l / f / f . (2.10) An alternating voltage will be found atf l , accompanied at both sides by noise according to
S d f i rfi A f ) = t VAC^ CiiflAf.

(2.11)

For obvious reasons this noise is called l / Af noise. (In this connection Af does not mean bandwidth.) There is perfect agreement between the calculated and observed ratio of the magnitude of the l/Af noise and the magnitude of the llfnoise when measured with DC current (Lorteije and Hoppenbrouwers 1971).

6
f-

Figure 2. Noise excited by an alternating current. The broken curve scales are logarithmic.

is the observed spectrum. The

What is much more puzzling is that in some samples a l/f noise also appeared at excitation. We then have the situation shown in figure 2. The observed spectrum 2: can be interpreted as a l/f spectrum and two l/Af spectra, the lower frequency branch showing spectrum folding. The noise intensity of the l/f spectrum is proportional to VAC^ with 2 < n < 4 . Roughly speaking, the intensity is several per cent of the intensity of the l / Af spectrum. Montagnon discovered Ac-generated low-frequency noise in 1954. It has further been studied by Sutcliffe (1972) who established that the spectrum was 1 1 Many authors thought that this Ac-generated l / f spectrum disagreed with resistance fluctuations as the source of 1 , noise. In 1977 van Helvoort and Beck proposed a model in a paper that also summarised the older literature. This model in its simplest form is a resistor with its resistance depending on the direction of the current. An applied
AC

490

F N Hooge, T G M Kleinpenning and L K J Vandanzme

sinusoidal voltage will lead to a superposition of a sinusoidal current on a DC component ZDC, proportional to VAC,ZDC will give lif noise, proportional to ID$, and hence to VAC^. In a somewhat more complicated form we take
R = Ro exp (a V ) .

(2.12)

If aV is not too large series expansion yields

The second term between the brackets yields


a Vo2 -_

R O

cos2 2xft =

a vo2
~

Ro

(+

+3 cos 4 T f l ) .
(2.13)

Thus there is a constant term

The l/fnoise, being proportional to Z D C ~ ,will now show a VAC4 dependence. If we consider one asymmetrical resistor, it must be strongly asymmetric in order to produce a large ZDC that can account for the observed Ilfnoise. This is the case for a gold contact on p-type germanium. Another possibility is given by grainy samples, for instance carbon resistors. Contacts between the grains, if rectifying, will be 50% of the exp ( a V ) type and 50% of the exp (- a V ) type. The I-V characteristic will be symmetrical and fairly linear. But the noise contributions of the individual contacts add up. Thus this seemingly ohmic resistor produces strong llfnoise on AC excitation. Such models can also explain the appearance of l/f noise around the higher harmonics of the excitation frequency. Jones and Francis (1975a, b) studied correlation effects at different frequencies in the Ilfbands and between l/f and l/Af bands. They, too, concluded from these measurements that l / f noise is a resistance fluctuation.

3. Configurations for measurement


Conductance fluctuations are easily probed by passing a constant current through a pair of driver electrodes and measuring the voltage fluctuations across a pair of sensor electrodes. In such a four-electrode configuration sensors and drivers are separate. The noise SV across the sensors is proportional to the square of the current through the drivers, but only in very special cases will S V = I ~ S Ewhere , SR is the spectral power density in the fluctuations of the resistance between the sensors. Four-electrode arrangements are used to minimise the influence of the noise generated at the current contacts. When the sensors are placed so that the average voltage across them is zero, then the so-called transverse noise will be measured. In this case, the measured transverse noise does not change when the current source Z at the drivers is replaced by a voltage source passing an average current I through the drivers. The transverse noise arrangement is very convenient when using DC pre-amplifiers to measure low-frequency spectra. If one can make good noiseless contacts then two-electrode arrangements are attractive for measuring noise in low conductance samples by applying a constant voltage across the sample and measuring the current spectrum SI. Two-electrode configurations have the

Experimental studies on llf noise

49 1

advantage of an unambiguous representation of the noise S v / V 2= S I / I ~ = S R / R ~ SG/G2 = = CI/j/f = m(nvefff)-l. For homogeneous samples submitted to uniform fields the effective volume veff equals the volume of the sample. In non-uniform fields Deft is smaller than the sample volume. In order to interpret experimentally obtained values of SV in terms of cy, general formulae are needed for the calculation of the resistance R and the resistance fluctuations SR for different electrode configurations. The measured noise in two- and four-electrode arrangements always consists of the thermal noise and a conduction noise term. The thermal noise equals 4kTR with R the resistance between the sensors. The conduction noise is proportional to the square of the bias current I. We shall now consider only noise due to conductivity fluctuations.
3.1. Two-electrode arrangements

The general formula for the resistance between an electrode pair on a sample is based on the relation for the total dissipated energy in the sample with a bias current I. The total dissipated energy is given by

RIz=S![ pJ2 dv

(3.11

and the integral must be carried out over the whole conductor except the electrodes because ideal contacts are assumed. Some local increase in resistivity leads to (i) an increase in the dissipated energy, and (ii) a small change in the local current density J. Now the correlation function for the conductance must be introduced for the calculation of ~ v ( T the ) , correlation function for the voltage. Butterweck (1975) simplified the calculation of ~ v ( Tby ) assuming 6 p uncorrelated in space except at distances shorter than a correlation length. The correlation length is small in comparison with contact dimensions. J2 does not change substantially on displacement over a correlation length. For the resistivity fluctuation apzlnf is introduced so that (1.26) will result in homogeneous situations. Here p is the resistivity and n is the carrier density. The general relation for the power density in the voltage for a two-electrode arrangement then becomes

When the resistivity is homogeneous and the current density is not, an effective volume may be introduced by Clif= a / n v e f f . Using equations (3.1) and (3.2) the effective volume for a two-electrode configuration is written as

The integrals must be taken over the whole sample except the electrodes. The effective volume is at most as large as the sample volume. Any deviation of the homogeneous field in uniform samples will lead to an increase in R and SV. Then Sv/V2 is no longer determined by the total number of free charge carriers in the sample. The noise would seem to be concentrated in an effective volume with higher current density. 3.2. Four-electrode arrangements Vandamme and van Bokhoven (1977) have given a general relation for the noise voltage

492

F N Hooge, T G M Kleinpenning and L K J Vandamme

between arbitrarily shaped and placed sensors when a constant current is applied to arbitrarily shaped and placed drivers :

The integral must be taken over the whole sample except the electrodes. Relation (3.4) is quite similar to equation (3.2) except that J4 is replaced by ( J g (2 which is the square of the scalar product of the current density J caused by the current I through the drivers and adjoint current density j. In a thought experiment the adjoint current is the current that flows when the current source has been switched from the drivers to the sensors. A proof for equation (3.4) was given by van Bokhoven (1978). This general relation was inspired by the sensitivity theorem in electrical networks. For practical applications of the equations one has to solve boundary value problems in order to find the required J and J. In general this problem will be too difficult to solve analytically. Then the application of an electrical network giving a discretisation of the Laplace equation leads to a tractable problem. The sensitivity of a change in voltage across a sensor pair due to a slight change in a resistor in the kth branch is given by 6 Vi8R = ik &/I,where the current and adjoint current in the kth branch are denoted by ik and fk. A number of numerical results are given by Vandamme and van Bokhoven (1977) and Vandamme and de Kuijper (1979). The adjoint current approach of equation (3 -4) provides a possibility of finding the areas of high and low contribution to the noise. The noise measured across small sensors stems from the neighbourhood around the sensors where J j is large. Quite often one can make some qualitative statements about SV by sketching the current lines J and j, led by physical intuition. Regions of high current density contribute heavily in equation (3.4) if at least the densities J and J" are not perpendicular. When the sensors and drivers coincide, then J = I and the four-terminal equation (3.4) reduces to the two-terminal equation (3.2). If we assume the sample to be homogeneous also in the statistical properties of the conductivity fluctuations, then p and ap2inf can be put in front of the integrals in the above equations. 3.3. Point contacts

If one can make reliable contacts, point contacts are eminently suitable for noise measurements because of their very small effective volume. A good approximation to the equations of the noise of a point contact can easily be derived by a simple treatment of the contact between two large spheres (Bell 1960, Hooge and Hoppenbrouwers 1969a) (see figure 3). The contact area is a circle with radius a. The equipotential surfaces are assumed to be spherical and concentric around the centre of the contact area. We consider a shell between the equipotentials at distances x and x + d x from the centre. The shell is homogeneous, also in the current density, so that equation (1.26) applies:

For the resistance in one contact member we find

Experimental studies on llf noise


cec----

493

/*/ /

--

/ /

Conductor

Figure 3. Simplified model of a point contact between two conducting spheres. The broken curves represent hemispherical equipotential surfaces.

The noise in one contact member is

By analogy to equation (1 .26) this is expressed as

For a complete contact with two contact members we find from SR = 2SR, and R = 2R1

This result also follows directly from equations (3.1) and (3.2) when J ( x )= 1/2nx2 is substituted. The equipotentials close to the contact area are hemiellipsoidals rather than hemispheres. Vandamme (1976a) calculated the noise using the two-electrode equation (3.2) and the more realistic equipotentials. His results give a firmer theoretical basis to the result obtained from the simple model of figure 3. The two models give different results for SE as a function of a and for R as a function of a. However, approximation (3.9) giving the relation between SR and R is surprisingly good. In experimental work one often uses the contact between two crossed cylindrical bars. The resistance can then easily be varied by a factor of 20. When there are no complications one finds that Clif is proportional to R3. This permits the calculation of 01 by means of equation (3.9). In this way a: of metals has been determined, which can scarcely be done in another way. With semiconductors there is still another possibility, the use of an ohmic contact between a large semiconductor sample and a thin metal wire. The metal wire, a good conductor, contributes neither to the resistance nor to the noise. Then a: can be calculated from equation (3.8).

494
4. Results for a
4.1. Accuracy o f

F N Hooge, T G M Kleinpenning and L K J Vundumme

C1/j

When l/f noise is measured it is usually accompanied by other noise which we shall call here background noise. The spectral density of the l / f noise will be denoted by S, that of the background noise by S b . When S + S b is of the same order of magnitude as S b the accuracy in S will be low. The background noise S b consists of the pre-amplifier input noise 4kTRe and the thermal noise 4kTR of the sample where Re is the equivalent noise resistance. When the filtered and squared noise is passed through a low-pass filter to give a running-time average we estimate that the detection limit for reliable a: values is S/sb>o.5. When the background noise is dominated by the pre-amplifier noise a cross-correlation measuring set-up as described by Storm (1978) is preferable. Then the filtered and multiplied noise of the two channels with inputs in parallel is passed through an integrator. The estimate for the detection limit is S / S b 2 0.1 for samples with a resistance greater than the equivalent noise resistance of the system. The equivalent noise resistance of a cross-correlation set-up can be 100 times lower than the Re of the pre-amplifiers used in each channel. Using equation (1 .26) for homogeneous samples submitted to uniform fields we find
S
EPd

UE29P

S<=4kT(1+ Re/R)fn=4kT(1 +Re/R)f


where Pd is the dissipated power density and E is the applied electric field in the homogeneous sample. For thin metal films, Pd is at most 10s W ~ m - ~ The . dissipated power density in the sample is limited by an excessively high temperature rise of the sample. Using point contact arrangements, larger values of the dissipated power density are permitted and thus higher values of S / S b . An equation similar to (4.1) is obtained using the relation for point contacts (3.9)

where EOequals the applied voltage V across the contact divided by the contact radius. Especially for metals and low-resistivity semiconductors simple point contacts are the best configuration for obtaining sufficiently high values of s / s b . Excessive heating is avoided in homogeneous samples by keeping E < lo2 V cm-l, for point contacts this value can be as high as EO= V/a< lo4 V cm-l. From equations (4.1) and (4.2) it follows that it is preferable for the pre-amplifier noise to be smaller than the thermal noise of the sample. In general the accuracy of Cllf is better than 20%. Quite a different question is whether accurate Cllf values will lead to reliable a values. In the calculation of a from Clif we always use a model with many parameters. It is often doubtful whether the model describes the sample well enough or whether all parameters are sufficiently well known, e.g. homogeneity of doping, dimensions of contacts, surface layers, etc.

4.2. Some conditions for reliable a: values


Non-homogeneous samples, noisy contacts, near-intrinsic materials and surface charges on the sample are some of the pitfalls in the determination of 01 from experimental Cllfvalues. Very thin evaporated metal films (about 10 nm) often show island structures. The a values from gold films presented by Bergero et a1 (1968) were several orders of magnitude greater than 2 x 10-3. The thin contacts in such a granular structure cause

Experimental studies on lif noise

495

local increases in the electric field, leading to non-uniform fields on a microscopic scale. In these cases the experimental results can not be interpreted by equation (1 .26). Another condition for obtaining reliable 01 values is that the samples are provided with low-noise ohmic contacts with a low specific contact resistivity ( ! 2 cm2). If one were to calculate an a: value for samples with poor contacts the result would have nothing to do with the true a: of the material. For near-intrinsic materials both electron and hole mobility fluctuations must be taken into account. Then a: can be calculated from the following equation :

where an and a , are the conductivities of electrons and holes, respectively, n and p are the free carrier concentrations and U is the volume of the sample. Cl/f is inversely proportional to the total number of carriers that dominate the conductivity. This can be the concentration of minority carriers as observed in MnO by Kleinpenning (1976a). Surface treatments have a strong influence on the l/f noise (Leuenberger 1967, Hanafi and van der Ziel 1978). This is still not fully understood. It may be caused by accumulation layers and surface charges in the oxide on the sample. It always results in low 01 values. With point contacts serious difficulties may arise from oxide f i l m between the contact members. Such a film may determine the resistance and the noise of the contact. If it does then the contact is called film-dominated ; otherwise it is constriction-dominated in which case equation (3.9) applies. If, in the model of figure 3, a film with constant resistivity Pfilm and constant thickness t is present, two situations are possible (Vandamme 1974a).
Constriction-dominated contact
Upbulk% tpfilm

Film-dominated contact
upbulk< tpfilm

Rcca-l CiinK R3

Rcca-2
Ciif cc R
~

Reliable a: values can be found from crossed-bar experiments only if Clif remains proportional to R3 during a variation in R by at least a factor of 10. It is also possible that a multispot contact is formed. The mechanical contact area then contains many small conducting spots. If there are k such spots equation (3.9) becomes

If we had just used equation (3.9) the result would have been a much too high 01 value. On the other hand, with the help of equation (4.4) the number of spots and their average diameter can be determined when 01=2 x is used (Ortmans and Vandamme 1976). Clarke and Voss (1974) investigated Bi films in order to see whether the noise was inversely proportional to the number of atoms or to the number of free carriers. In Bi there is a great difference between the number of atoms and free carriers. Their results gave no support for relation (1.26). However, they overlooked the fact that the hole concentration at room temperature is about 3 x 1019 cm-3 instead of 1017 cm-3 which is the value at 77 K. Vandamme and Kedzia (1979) obtained from crossed-bar experiments a: values for Bi of 3 x 10-3, using for the hole concentration 1.5 x 1019 cm-3 as followed
32

496

F N Hooge, T G M Kleinpenning and L K J Vandamme

from the Hall effect measurements. Vandammes data on manganin and bismuth show that (i) a:-type llfnoise is not a temperature-induced noise, and (ii) relation (1.26) holds for Bi and nianganin. One may wonder whether a: differs in the solid and liquid phase. Therefore Kedzia and Vandamme (1978) investigated the lif noise in liquid and solid gallium. From the experimental results they concluded that a for solid, liquid and undercooled gallium is about 2 x 10-3 and that a does not greatly depend on the temperature in the range 77-600 K. Stroeken and Kleinpenning (1976) found for mercury at room temperature an a value of about 2 x 10-3. May and Aniagyei (1974) did not succeed in measuring an 01 lower than 20 on a tungsten wire. Vandammes (1976b) results in incandescent tungsten filaments were in agreement with their results. However, his experiments on tungsten crossed bars resulted in a: values of about 10-3. The large amount of noise in thin wires is due to multispot contacts at the clamping contacts of the filament. Stroeken and Kleinpenning (1976) demonstrated that Si and Ge at 300 and 77 K show a: values of about 2 x even if the samples were plastically deformed. Although the mobilities decrease with increasing deformation 01 is practically independent of the deformation. From point contact experiments it was found that a: has no appreciable pressure dependence in the range from 0 to 25 kbar (Vandamme 1976b).

4.3. a: values
The survey made by Hooge (1969) of the data published on llfnoise did not provide absolute evidence that relation (1 .26) is correct for compound semiconductors. Therefore further study was made of 111-V compounds by Vandamme (1974b). The averages of the experimentally obtained results are summarised in the table below.
Concentration from Hall measurements (cm-3) n-type InSb p-type InSb p-type GaSb n-type GaAs p-type GaAs n-type GaP
1.6 x 1 .6 x 1.2 x 1. 5 x

cc
1 . 3 10-3 ~ 3 . 4 10-3 ~ 7 x 10-3 2 x 10-3 3x10-3 6x10-3 3.4~10-3 9x10-3

T W
77 300 77

lOI4
1016 1OI6 1017

77
300

2.3 x 1016 2.3 x 1016 2.9 x loL6

300
300 300

It follows from these experiments that (i) a: values for 111-V compounds are approximately 2 x 10-3 and (ii) n does not greatly depend on temperature. Studies of point contacts of ten metals showed that their a: values were about 10-3 (Hoppenbrouwers and Hooge 1970). The temperature coefficient of the resistance and the total number of atoms in a resistor play a central part in the temperature-noise model of Clarke and Voss (1974) (see $6). They found experimental evidence for their theoretical prediction that manganin with a negligibily low temperature coefficient will show no llfnoise. However, from manganin point contacts Vandamme (1976b) obtained a: values of 0.7 x 10-3 using an electron concentration of 3.5 x 1022 cm-3 in the calculation. The conclusion from all these experiments is that equation (1.26) describes the l/f noise for many materials, such as metals, solid and liquid, and semiconductors, n- and p-type. The numerical value of a is about 2 x 10-3. Because of the many complications that may occur it is not surprising that the spreading in a: values is so high: about a factor of three from the average. This

Experimental studies on Ilf noise

497

inaccuracy makes it impossible to determine whether a really is a constant. a may differ for different materials. 01 may be slightly temperature-dependent. All this is true as long as lattice scattering prevails. In $8 we shall see that other scattering mechanisms reduce CY. values considerably.

5. Surface states and &noise

5.1. The Mc Whorter model


One of the major difficulties in understanding llfnoise is the inability of simple physical models to give a l/f spectrum in a natural way. A simple model that does give a l/f spectrum was proposed by McWhorter (1955,1957). The McWhorter model has remained the most accepted model for llfnoise in MOST. Trapping of charge carriers in traps located at a distance from the semiconductor-oxide interface is considered as the noise source. The McWhorter school of thought believes in (i) the surface origin of l/f noise, and (ii) fluctuation of number instead of fluctuation in the mobility of the free charge carriers. Surface states in a broad sense are defined as any electronic state which is localised at the surface, i.e. its wavefunction has a maximum amplitude at or near the surface and vanishes at a sufficiently great distance from the surface. In this respect states in the oxide near the interface or at the interface and in the accumulation or inversion-layer region of the silicon can be classified as surface states. The terms slow and fast surface states were used in early investigations of states at the Si/SiOz interface. The fast states are conjectured to interact either directly, or through tunnelling, with the silicon bulk and therefore must be located at or at least very near the interface. Charge exchange between the silicon and the slow states requires a very long period of time. The orientation dependence of interface states correlates with the density of the available bonds per unit area on the corresponding crystal orientations. Surface states are thus often attributed to some unsaturated or dangling bonds at the silicon surface. Some physical models for interface states predict the absence of surface states on an oxidised silicon surface. Other models assume a strong correlation between interface states and oxide charges or between surface states and disorder or misfits. There is no single method by which the distribution of interface states over the energy gap can be measured throughout the band gap. A complete picture can only be had by combining results obtained with two or more measuring techniques. The peaks in the interface state density near the band edges have often been questioned by, for example, Boudry (1973) and Declerck et a1 (1973). The bias dependence of the noise in MOST is often explained by the energy distribution of the interface states in the gap. Only traps within a few kT of the Fermi level, which is bias-dependent, are effective in generating noise. This explanation for the bias dependence of the noise from Pai (1978), for example, seems doubtful, considering the apparent peaks in the surface states. In a letter entitled Evidence of the surface origin of the llfnoise, ah and Wielsher (1966) presented correlations between the noise, the lossy part of the gate impedance due to carrier recombination and the interface states of the MOST. However, when we analyse their results between 20 Hz and 100 kHz at a constant gate voltage we must conclude that the observed noise is not I/$ For example, at V G = -9 V their spectrum between 20 Hz and 100 kHz is proportional to l/f2. This misleading situation was the start for much experimental and theoretical work on the relation between l/fand the surface states (see, for example, Hsu et al 1968, Klaassen 1971, Broux et a1 1975). Since l/f spectra were observed in a frequency range of ten decades or more, the summation

498

F N Hooge, T G M Kleinpenning and L K J Vandamme

approach of equation (1.21) requires that the 1/r distribution of time constants holds good over an extremely large range and that the physical processes at different are independent. The values of the limiting frequencies are not that important in view of the discussion given after relation (1.23). This wide range of time constants can arise from a thermally activated process or a tunnelling process. In MOST it is usually assumed that the time constant dispersion is caused by quantummechanical tunnelling of carriers from the Si/SiOz interface to traps located inside the oxide, This McWhorter model has been used in theoretical calculations, for example by Christensson et aZ(1968), Leventhal(1968), Berz (1970) and Hsu (1970). They all got more or less the same expressions for the noise. Fu and Sah (1972) suggested that direct tunnelling of free carriers, either from the conduction band or the valence band into the oxide traps, is an unlikely mechanism. Instead, they proposed a two-step process in which the free carriers first communicate with the fast surface states at the interface and then tunnel into the oxide traps elastically. The addition of an intermediate state gives essentially the same expression for the noise. The dominant low-frequency noise comes from the fluctuation in occupancy of the oxide traps. If the electrons in an element of the interface do not interact with one specific oxide trap but with several ones having different T , then the result is a simple generation-recombination spectrum with an effective r which is close to the smallest r since

Only if each element of the interface has its dominant trap with a specific r are the traps independent. The noise is then a summation of generation-recombination spectra, each with its own r . The simplest form of the McWhorter model with a continuous set of with statistical weight l / r is as follows. It is assumed first that in the oxide layer on the semiconductor there are electron traps with a constant concentration through the whole layer. The second assumption is that the probability of penetration into the oxide layer decreases exponentially with distance from the interface. In this case the r are determined by quantum-mechanical tunnelling or overlap of the wavefunctions. The probability of penetration is proportional to 1 / r and to exp (-x/X). Therefore r becomes
T = 70

exp (x/h)

( 5 -2)

where h is a characteristic decay length of the wavefunction with an order of magnitude of 1 A. The traps far from the interface have large r , yielding the low-frequency part of the spectrum. The distribution function g(T) becomes

which is required to obtain a l/f spectrum.


5.2. Discontinuousfilms

Discontinuous metal films show extremely high l/f noise. The high noise intensity is connected with the special type of electrical conduction in such films, which is by tunnelling. An extensive study of discontinuous films, their conductance and their noise has been published in two papers by Celasco et al(l978). Section 5.2 is a summary of these papers, especially with regard to the llfnoise.

Experimental studies on llf noise

499

The films consist of metal islands evaporated on to an insulator. The main conduction mechanism is the tunnelling of electrons from island to island. The strongest experimental support for this model is that the conductivity decreased very steeply from about infinity to very low values during the evaporation process when a critical thickness was reached as in a percolation process. The tunnelling process 1 in figure 4 gives the conduction. The potential barrier height is Ec - EF,the distance between the conduction band of the insulator and the Fermi level. The position of the conduction band is determined by the electron affinities of the metal and the insulator and also by the occupancy of the trapping centres. The origin of the noise is in the fluctuation of the occupancy of the traps. The mean number of trapped electrons nt is, of course, given by Fermi statistics, but there will be a kind of generation-recombination noise because of tunnelling processes between metal and traps, such as process 2 in figure 4. The fluctuation Ant in the number

fF

lnsuiator

Figure 4. Model of a discontinuous metal film.

of trapped electrons changes the charge Q of the insulating layer and thereby the potential energy Ec. This barrier fluctuation causes the tunnelling to fluctuate. The noise originating in process 2 causes the noise measured in process 1. Since this noise is McWhorters generation-recombination noise its spectrum is l/f, Not only does the shape of the spectrum agree with the model but the intensity, too, agrees well with reasonable numerical estimates for the concentration of the trapping centres. Apart from the two tunnelling processes there are thermionic transitions from a metal island into the conduction band of the insulator and from there either to the traps or to another metal island. There will be generation-recombination noise generated by the transitions between conduction band and traps. In the original papers it is theoretically and experimentally demonstrated that the thermionic processes are of no importance below temperatures of 400 K. In this summary we shall therefore neglect thermionic processes from the start.

500

I:N Hooge, T G M Kleinpenning and L K J Vandamme

The conductance of the structure of figure 4 follows from the current equation given by Simmons (1964): ~(Ec ] -EF e v ) exp [ -A(Ec -EF Joc(Ec - EF) exp [ -A(Ec - E F ) ~ /where Vis the applied voltage over the barrier and

+ eV)1/2]

(5.4)

A = 4~(2m)1/2 h-ld

(5 ' 5 )

where d is the thickness of the barrier. In this part we shall omit the non-essential proportionality constants which are complicated products of fundamental constants and dimensions of barriers and islands. The conductance G follows from dJ/dV with eV<(Ec-EI;.), i.e. GKA(Ec -E ~ ) l / 2 exp [ - A(Ec (5.6) dG/dEcE - [ A- (Ec - EF)-"~] exp [ - A(Ec - EF)li2]

(5.7)

It is assumed that the traps are homogeneously distributed in the insulator. The relaxation time of a tunnelling process between the metal and a trap at a distance x from the metal is

This is thus a McWhorter model leading to llfnoise in the number of trapped electrons according to equation (1 .21) : (5.10) The relative noise in the conductance now follows from (5.8) and (5.10) since

SE^ c cSQc cStzt.

(5.11)

Quantitative agreement is obtained assuming 1015 traps cm-2, which means that all atoms at the outer surface of the insulator act as traps. The authors took I n 7 2 / 7 1 = 16, which could be criticised, but this value is not critical at all in view of the discussions on equation (1 .23). At higher temperatures, T > 400 K, the thermionic processes become important. Electrons are then excited into the conduction band from where they can fall into the traps. Traps that were difficult to reach by tunnelling now become more involved in fast generation-recombination processes. The larger 7 disappear from the 7 distribution, which makes the spectrum flatter at the lower frequencies. There is good agreement between theory and experiment in this respect. The McWhorter model which explains the llfnoise in discontinuous films so well was originally proposed for the surface noise at the semiconductor-oxide interface, e.g. in MOS transistors. But now it is very doubtful whether MOST exhibit McWhorter noise. There are strong arguments in favour of 01 noise (see $10). Such difficulties in the interpretation of the lif noise do not exist with discontinuous films. Indeed, the discontinuous film is the McWhorter noise source par excellence.

Experimental studies on Ilf noise

50 1

5.3. The Mc Whorter model and a noise


The strongest objection to the McWhorter model as a possible explanation of 01 noise is that a noise is a bulk effect. a noise is experimentally found to be inversely proportional to the total number of free charge carriers. This excludes surface effects as the source of a noise. The noise density is not related to the geometrical shape and the surface area, but depends only on the volume of the bulk. A clear demonstration of the bulk hypothesis is given by the l/f noise experiments on continuous gold films by Hooge and Hoppenbrouwers (1969b). Here the l/f noise density of films with different thicknesses depends purely on the volume of the films. There are a number of l/fnoise experiments where the results are interpreted as a bulk effect. Several examples are presented in this review (993, 7 and 9): llfnoise in solid-state injection diodes, in p-n diodes, in thermo-EMF, in Hall EMF, in two- and fourprobe configurations, and so on. In all these cases the llfnoise is calculated on the basis of formula (1 .26). The agreement between experimental and calculated results is striking. There are inore experiments in which the McWhorter model fails to explain the results. These experiments are described in 97, which discusses the problem of whether the fluctuations are in the concentration or in the mobility. The l/f noise has been investigated in a number of devices, where the conduction occurs close to the surface. Here the llfnoise density depends on the quality of the surface. Some people concluded therefore that McWhorters model prevails. Other people came to the conclusion that the condition of the surface affects the free-carrier density profile, the electric-field strength profile and the scattering probability of the carriers. Since the l/f noise depends on these quantities it is possible that the surface indirectly affects the llfnoise. Examples are given in 98 (thin Bi films) and in $10 (MOST transistor). Another illustration was given by Kleinpenning (198Qa). He discussed the llfnoise of a short p+-n diode in terms of equation (1 -26). The magnitude of the ILfnoise is determined by the concentration of minority carriers in the base, which is related to the contact recombination velocity of the injected minority carriers at the collecting contact. So the contact recombination velocity indirectly determines the I / noise density. Van der Ziel (1979) suggested in a review paper that grain boundaries in semiconductors are a source of l/f noise. However, a high intensity of l/f noise due to grain boundaries can also be interpreted as a volume effect. Around the point of contact of the grains the electric current shows constriction effects, which lead to high electric-field strengths and to high noise levels (93).

6. Temperature fluctuations
Hooge and Hoppenbrouwers (1969b) measured the l/f noise in continuous gold films. The results obey the empirical relation (1.26) with 01=2.4 x 10-3. They concluded that the magnitude of llfnoise in gold is of the same order as in semiconductors with the same number of free carriers. Voss and Clarke (1976b) studied resistance fluctuations in thin metal films. They found general agreement with equation (1 ,26) with two exceptions, viz. bismuth and manganin. These exceptions are discussed in this review in 994 and 8. Voss and Clarke interpreted the observed resistance fluctuations in terms of equilibrium fluctuations in temperature that modulate the film resistance R. Their reasoning is as follows. According to the statistical mechanics of the equilibrium state the average of

502

F N Hooge, T G M Kleinpenning and L K J Vandarnme

the energy fluctuations 6E of a system in contact with a heat bath is given by

((8E)z)=kT2Cv

(6.1)

where k is the Boltzmann constant, Tis the temperature and Cv is the heat capacity of the system. Since energy fluctuations are connected with temperature fluctuations, 6E= Cv8T, we obtain ((6 T)2) =kT2/Cv. (6.2) The mean square resistance fluctuation of a metal film in contact with a heat bath is then given by {(6 R)) = (dRjd T) ( ( 6 T)) = y2R2k/ Cv (6.3) where y = d In R/d In T. For most metals at room temperature ycz 1, NE N A and Cv=3 NAk where NA is the number of atoms in the film. So the mean square of the relative resistance fluctuations is ((SR/R)2)- ~ / ~ N 1/3N. AE (6 4)

A vital problem is now the frequency dependence of the spectral density of the fluctuations in R. The spectral density has to be of the magnitude given by equation (1.26) with a l/f dependence over many decades of frequencies. The temperature fluctuation model is encouraging with respect to the magnitude. This is based on equations (1 . I l), (1.26) and(6.4). As to the power spectrum, a large number of papers have been published on this subject. The main difficulty here is to obtain an explicit ljfregion over many decades of frequencies. At best a l/for a l/f-like spectrum can be obtained for a very restricted frequency range. We shall present some results based on the Langevin diffusion equation of temperature fluctuations given by
1 V.F(P, t ) a T(P t , - DV~T(Y, t ) = ~at

cv

where T(r, t ) is the temperature at spot Y at time t, D is the thermal diffusivity, cv is the specific heat, and F(v, t ) is the random Langevin source function associated with the transport fluctuations in the system. The quantity F(v, t ) is uncorrelated in space and time. Applying equation (6.5) to a system yields the power spectrum for the temperature fluctuations. The frequency dependence of this spectrum depends both on the geometrical dimensions of the system and on the boundary conditions governing the heat flow between the system and surrounding space. In this review we restrict ourselves to metal films with length 1, width iv and thickness t , I > w > t. We shall present spectra for different boundary conditions. For a completely isolated film there is no heat flow between the film and the surroundings, so ((6R)Z) =O. Let the ends of the film (x= 0, x = Z) be connected with supply wires, so that the system is thermally homogeneous in the x direction, and let the heat flow between film and the rest of the surroundings be negligible. In this case Kleinpenning (1976b), Liu (1977) and Ketchen and Clarke (1978) found S~(j)xf-1/2forf<fi= D / d 2 and S ~ ( f ) a z f - 3 / forf>fi. 2 Normalisation of the spectrum is achieved by using

s;

df=((W2).

(6.6)

The result is plotted as curve D in figure 5.

Experimental studies on llf noise

503

Figure 5. Calculated noise spectra of relative resistance fluctuations in metal films using NA= N , a = 2 x 10-3, y = 1 and ( l / ~ ) 103. ~ = Curve A : according to equation (1.26). Curve B: according to Voss and Clarke's semi-empirical formula, equation (6.7). Here, f o = f i = D / n P . Curve C : according to equation (6.8) with fo= 1/2VT. Curve D: one-dimensional system, thermally homogeneous in the x direction, length I, fo= D / n P .

Let the film be embedded in a medium, so that film and medium can be considered as a thermally homogeneous system. In this situation Voss and Clarke (1976b) found
f

<fi = D / d 2;
<f <ft = D/nt2;

ST( f ) =constant
S T ( f ) x I n (Uf)
ST(f .cf- l I 2
ST(f ) x f -312.

fi<f<fw=D/rw2:
fw

ft<f:

Here we do not find an extended region of l/f behaviour. According to Voss and Clarke, the experimental system of metal film on substrate does not present a uniform medium for heat conduction. For such a complex system they assume that there is a l/f-like region in the spectrum at intermediate frequencies f withfi<f<fw. Here the high- and lowfrequency limits are taken to be ST(f ) Kf -312 for f >fw and ST (f) = constant for f <fi. With the normalisation condition, equation (6.6), for the l/f-like region of metals where Cy= 3 N ~ they k find

Curve B in figure 5 represents this semi-empirical Voss-Clarke formula. It should be noted that Voss and Clarke's temperature fluctuations model has neither a pure bulk nor

504

F N Hooge, T G M Kleinpenning and L K J Vandamme

a pure surface character. The integral JtS(f)df is determined by the bulk volume, but the frequency behaviour of S(f)is determined by the geometrical shape. For metal films with good thermal contact with the substrate where the heat exchange between film and substrate is dominant, we find

where .r=RthCv is the heat relaxation time, and Rth is the thermal resistance between film and substrate. Equation (6.8) holds for T < I ~ / ~ D The . result is plotted in figure 5. Let us compare the experimental results from the gold films as obtained by Hooge and Hoppenbrouwers (1969b) and Kleinpenning (1976b) with equation (6,7). For these films we have I= 800 pm, w = 10 pm, y = 0.5 and D = 1.2 cmz s-1, so that equation (6.7) predicts SR(f)/R2= 7 x lO-3/fN in the frequency range fi = 60 Hz and f w = 400 kHz. Experimentally, it is found that S~(f)/R2=2.4x lO-3/fN in the 10 Hz to 30 kHz range, which agrees with the semi-empirical Voss-Clarke formula equation (6.7). However, the heat relaxation time of the Au films was experimentally found to be about T = 0.3 ps. Since T <Z2/2D N 2.7 ms, the spectrum of the temperature-induced resistance fluctuations is given by equation (6.8) and should have the Lorentzian shape. It should be noted that for frequencies betweenfi and fw the ratio ,B between resistance fluctuations due to temperature fluctuations and the l/f fluctuations is

,B =yZN{3[3+ 2 In (Z/w)] a N ~ } - l .

(6.9)

For metal films y N 1, taking N / N AN 1 and In (I/w)N 4 yields ,B N 10. So equations (1 .26) and (6.7) give about the same order of magnitude. Therefore Voss and Clarke's temperature fluctuation model is compatible with equation (1.26), at least for metal films in a restricted frequency range. Nevertheless, there are many serious objections to the suggestion that l/fnoise is a result of equilibrium temperature fluctuations. The main objections are : (i) In view of the ratio Z/w of the investigated films, the model cannot give a l/f spectrum over many decades of frequencies. (ii) The model underestimates the experimentally obtained magnitude of l/f noise in J cm at 300 K semiconductors by a factor sometimes greater than 1010. For n-Si of lo4 ! we have NINAN 10-10 and a N 10-3. (iii) Kleinpenning (1974) has found from l/f noise experiments on the thermO-EMF of intrinsic germanium that the fluctuations in the conductivity of electrons and holes are uncorrelated, (6 0%6up)=0 (see $7). For temperature-induced conductivity fluctuations in intrinsic semiconductors 6an/un= 6up/up applies, so that the fluctuations are completely correlated. Kleinpenning (1976b) also found for intrinsic germanium that the temperature dependence of the relative fluctuations in the resistance was given by exp (Eg/2kT), with Eg the band gap, which agrees with equation (1.26). According to equation (6,7), the temperature dependence should be given by y2= (Eg/2kT)2ccT-2 which is contrary to the experimental results. (iv) Equation (6.7) holds good only for frequencies around the reciprocal thermal time constant T of the system. This time constant is related to the dimensions of the film, but also to the heat resistance between the film and substrate. Gold films with a very low heat relaxation time (0.3 ps) still give llfnoise at low frequencies ( 10 Hz) with magnitude in accordance with equation (1.26). It was found by Hooge (1972) that metal point contacts with radii r < 1 pm show l/f noise at low frequencies ( 100 Hz) according to equation (1.26), while their thermal time constants are of the order of T N ~ Z / < D10-8 s.
N N

Experimental studies on l l f noise

505

(v) For temperature-induced resistance fluctuations there should be a correlation between the noise in different parts of the films at low frequencies (fiN D j d 2 ) . Such spatial correlation effects are observed by Voss and Clarke (1976b) and Clarke and Hsiang (1976) on thin metal films with poor thermal contact between film and substrate and in a frequency region aroundfi. However, in the semiconductor n-GaAs no such correlation effects could be observed by Kleinpenning (1976b). (vi) Eberhard and Horn (1978) have investigated the temperature dependence of I/' noise in metal films on fused silicon and sapphire substrates. In contrast with both equation (1.26) and Voss and Clarke's formula they found a rather strong temperature and substrate dependence of SB(f)/R2. Neither of the two models can adequately describe the data presented by Eberhard and Horn. Possibly the results may be ascribed to a combination of temperature-induced resistance fluctuations on the one side and a noise on the other. (vii) Van Vliet and Mehta (198 1) have treated the problem of temperature fluctuations in a theoretical way. They have shown that no l/fnoise will arise from heat diffusion in a system, unless very specific noise sources are proposed whose physical origin is untenable. In a discussion of temperature fluctuations Kleinpenning (1976b) concluded that there is a real danger of confusing 01 noise and temperature-induced resistance fluctuations. In semiconductors the temperature-induced fluctuations are always negligible with respect to a noise. In metals the temperature-induced fluctuations are generally greater than a noise at frequencies near the thermal relaxation time and are negligible with respect to the a noise at frequencies far away from this relaxation time (see figure 5). We conclude that a noise in semiconductors cannot be caused by temperature fluctuations. Further, it is most improbable that temperature fluctuations are the origin of l/fnoise in metals. They give only a l/f-like contribution at frequencies around the reciprocal thermal time constant.

7. Fluctuations in number or in mobility?


The conductivity u of an extrinsic semiconductor is given by
u =q pn

(7.11

where q is the elementary charge, p is the mobility and n is the free-carrier density. From equation (7.1) it is obvious that conductivity fluctuations can be caused by mobility fluctuations Sp or by density fluctuations Sn. So we have the possibilities Su=(do/dn) Sn=qp8n
6 o = (d old p) 6 p =q n8p.

(7 * 2) (7.3)

We shall discuss both types of fluctuation for a semiconductor in thermal equilibrium. The assumption of mobility fluctuations as the origin of l/f noise was introduced by Hooge (1972). He suggested that the mobility of a free charge carrier fluctuated as
Spip."= . i f .

(7 4)

For homogeneous samples with N free carriers, equation (7.4) becomes

s,/p2= 4 f N

(7.5)

506

F N Hooge, T G M Kleinpenning and L K J Vandamme

assuming that the mobilities of the free charge carriers fluctuate independently of each other. To discriminate between number fluctuations SN and mobility fluctuations Sp, we have to study fluctuations in a quantity X ( p , N ) which depends on p and N in different ways. For number fluctuations we have SX,=(dX/dN) and for mobility fluctuations SX,=(dX/dp) Sp
Sxp = pd X/d P>

SN

SxN= (NdXldN)2 S N I N ~

(7 * 6)

<

sfilp2*

(7.7)

So we have to investigate the l/f noise of a quantity Xwith 1 NdX/dN( # I pdX/dpL(. The conductance is certainly not such a quantity; here G- p N and thus NdG/dN= pdGldp. Quantities obeying the inequality are, for example, the thermal voltage and the Hall voltage of extrinsic semiconductors, the conductivity at high electric-field strengths (hot carriers) and the conductivity of heavily doped semiconductors. In this section we shall treat successively the a noise in these quantities.
7.1. l/f noise in thermal voltage

The treatment of this subject is based on a paper by Kleinpenning (1974). For an n-type non-degenerate semiconductor the thermal voltage Vth is proportional to the energy distance between the Fermi level and the mean conduction level in the conduction band (see figure 6) Vth=(EF+ &IC) ATIqT (7 * 8 ) where CO &G(E) de G(E)de)-' (7.9) and (7. IO) E F =kT In (Nc/N).

&&=SO

( 1 :

- - - - - - - - -t

Figure 6. Energy diagram of a semiconductor.

Experimental studies on llf noise

507

Here Nc is the effective number of states in the conduction band, N is the number of free electrons, A T is the applied temperature differences, G(E)de is the conductance of electrons with energy between E and &+de, and G=J," G ( E ) dE is the conductance. For lattice scattering we have ( E k ) = 2 kT. Conductance fluctuations 6 G = J 8G(e) de lead to fluctuations in Ek 6 & k = J r (E-(&) % ( E ) de/(G). (7.11) If the conductance fluctuations are due to number fluctuations, we have
~ E F = -(kT/(N))

6N

SQ = (kT)2SN(N)-2.

(7 * 12)

The important result is that number fluctuations lead to 6 ~ = atf<TC-l, 0 the reciprocal collision time of electrons. A fluctuation 6N will be redistributed over the conduction band according to the Fermi distribution within a short time ET^). Then 6N(s)ccN(e) and thus SG(E)EG(E).Using this result in equation (7.11) yields S E ~ = O . Up to here we see that number fluctuations lead to B E F # 0 and 6 ~k = 0, and mobility fluctuations to ~ E F = Oand possibly 6 ~ k # O . The next question is to find the relation between mobility fluctuations and 6% The contribution of the free carriers is assumed to be independent. We can divide the conduction band into a large number of sub-bands, so that in the sub-band i the contribution to the llfnoise is
SG,/Gi2 = a:/fNi

(7.13)
N

where the subscript i refers to the sub-band i. Furthermore, the fluctuations in different ) sub-bands are uncorrelated. It should be noted that carriers scatter rapidly ( T ~ from one sub-band to another, but this will only lead to a generation-recombination-like noise component at very high frequencies (- 1012 Hz). From equations (7.11) and (7.13) we find S t k = x ( E ( - ( E ~ c~Gi~(fNiG')-~. ))~
i

(7.14)

Finally, using equations (7.8), (7.12) and (7.14), we obtain Svth=P(kAT/q)2S G / G ~ (7.15) where P= 1 for number fluctuations, and /3=2.75 for mobility fluctuations if lattice scattering prevails. Thus we obtain different results for 6 p and 6N. But, owing to inaccuracy in measurements, it is difficult to discriminate between number and mobility fluctuations. To distinguish between the possibilities 6 p and SN, we have studied the noise in thermal voltage with an additional external voltage (see figure 7). For the open-circuit situation we have
(

v =i0/m - (e) A T
68=

6 V = - (Io/G') 6G - A TB 8

(7.16)

where 8 = - ( e k + E F ) / ~ isT the Seebeck coefficient. With equations (7.11) and (7.12) the fluctuations in 8 are found to be

--

qTG o

sm

(E-(E~))

~G(E de+k6N/qN. )

(7.17)

Now we calculate y=Sv(Va, AT)/Sv(O, A T ) as a function of the external voltage Va=lo/(G) at fixed A T , both for number and mobility fluctuations. The results are plotted in figure 7. Experimental data obtained from n-Si are represented by dots. The conclusion is that a: noise is a fluctuation in the mobility.

508

F N Hooge, T G A 4 Kleinpenning and L K J Vandamme


I
I

qV,/kT

Figure 7 . llfnoise of thermo-EMF with applied voltage. Dots: experimental results of n-Si. Broken curve, calculated results for number fluctuations, full curve for mobility fluctuations.

At the same time Kleinpenning (1974) investigated the llfnoise in the thermal voltage of intrinsic non-degenerate semiconductors. Here the thermal voltage is (7.18) with Gn and Gp the electron and hole conductances, and
Vthn= ( EF

+ ~ k )A T/qT

v t h P = - (Eg - EF

+ Ek')

A T/qT.

(7.19)

Here Eg is the band gap and Ek' is the mean conduction level in the valence band (see figure 6). Let us define a quantity /3 according to equation (7.15), where G=Gn+ Gp. Now the question is, what is the magnitude of p for number and mobility fluctuations, respectively ? For number fluctuations we have the product NP is constant, since a shifting of the Fermi level will affect both the number of electrons N and holes P. Consequently SN/N= - 8P/P. Hence we find

6 Vthn= 6 VthP= - ( kA T / q ) SN/N


and

GGnlGn=

- 6Gp/Gp=6N/N

6G/G= [(G' - Gp)/G]SNjN.

Applying these results to equations (7.15), (7.18) and (7.19) we obtain for lattice scattering ( E k = ek' = 2kT)

'= (

2GnGp(Eg/kT+4) - G2 G(Gn-Gp)

) =(

2b(Eg/kT+4)- (b 1)' b2-1

(7 * 20)

where b = Gn/Gp= pn/ppis the mobility ratio. For mobility fluctuations the free charge carrier mobilities are assumed to fluctuate

Experimental studies on llf noise

509

independently. Consequently (6pn 6,up)=0 and (6Gn 6Gp)=0. Now we obtain for / I approximately P=(Eg/kT+4)2 2b2(b+ 1)-2(b2+ 1)-1. (7.21)

P E 150.

Experimental results obtained from intrinsic germanium at room temperature yields Using b=2 and Eg/kT=26 and equation (7.20) we obtain P 2 : 1400. Equation (7.21) yields 160. So the mobility fluctuation hypothesis is here also supported by the experimental results.

P=

7.2. Ilfnoise in Hall voltage

An additional support for the mobility fluctuation hypothesis is obtained from investigations on llf noise in the Hall voltage. Recently Vaes and Kleinpenning (1977) and
1 I

Figure 8. llfnoise in Hall voltage. Dots: experimental results of n-Ge. Broken line, calculated result for number fluctuations, full line for mobility fluctuations.

Kleinpenning (1980b) have investigated the influence of magnetic induction B on the llfvoltage noise density between two small circular Hall electrodes in terms of the ratio y(B)=S(B)/S(O)(see figure 8). Here S ( B ) and S(0) are the llfvoltage noise densities with and without magnetic induction when the applied voltage across the line-shaped outer electrodes is constant, and the frequency fixed. The ratio y was calculated for two cases: l/f fluctuations are due to mobility fluctuations (yp) or to free-carrier density fluctuations ( y ~ ) . In these calculations an n-type non-degenerate homogeneous isotrope semiconductor with spherical constant-energy surfaces has been considered. There is a direct current density Jx in the x direction and a magnetic induction B in the z direction. The current densities in the x and the y directions ( x _ L y l z )at point r at time t are given by
Jx@, t ) = A ( r , t ) E&, t)-D(r, t ) E&, t )
Jdy,

(7.22) (7.23)

t>=D(r,t > E&, t)+A(r, t ) E&, t )

510 with

F N Hooge, T G M Kleinpenning and L K J Vandamme

div J(r, t ) = 0 where

(7.24) (7.25) (7.26)

U(&,

r, t)=qp(&,r, t ) (2~/3kT) n(E, r, t ) .

(7.27)

Here U(&, r, t ) is the conductivity, p ( ~r, , t ) is the mobility and n(e, r, t ) is the carrier density with energy E at point r at time t. E(r, t ) is the electric-field strength. To obtain the voltage fluctuations between two points in the xy plane, fluctuations around the steady state have to be considered. These fluctuations are assumed to be caused by either mobility or density fluctuations. So the quantities A and D fluctuate and as a result the quantities J and E do also. The relation between the fluctuations in A , D, J and E= can be found from equations (7.22)-(7.24). The result is -grad
A div grad 8rp(r, t ) = E grad 8A(r, t ) [ EA grad 6D(r, t)lz.

(7.28)

Making a Fourier transform in the frequency-wavevector representation, equation (7.28) yields (7.29) rp(k, w ) is the Fourier transform of Srp(r, t ) . The fluctuation in the open-circuit voltage between two point electrodes at r1 and YZ for a two-dimensional conductor is t(r1, re, w)=!rp(k, U ) [exp (jk.rl)-exp(jk-rz)] dk. (7.30)

In order to find the spectral density of v(r1, rz, U ) the properties of the a-noise sources have to be known. As the calculations of the ratios y p and Y N are rather complicated, we shall restrict ourselves to giving the final results. A detailed description is given in the paper by Vaes and Kleinpenning (1977). If the l/f noise can be ascribed to mobility fluctuations, according to equations (7.4) and (7.5) the result of y p for lattice scattering is

where A ( B ) and D(B) are the quantities of equations (7.25) and (7.26) at magnetic induction B, and U(&), p(e) and n(E) are the averages of the conductivity O ( E , Y, t ) , mobility p(e, r, t ) and free-carrier density n(E, r, t ) , respectively. At low magnetic inductions ( ~ H B 1) < equation (7.31) can be approximated by yp(B)=1+(1+8/3n) ~ H ~ B ~ with p~ the Hall mobility. If the l/f noise is due to free-carrier density fluctuations, then the ratio (7.32)
YN

is

(7 * 33)

Experimental studies on Ilf noise


At low magnetic inductions this reduces to yN(B)=l +2pH2B2.

511

(7.34)

From equations (7.32) and (7.34) it follows that at low B the difference between ylL(B) and ~ N ( Bis ) too slight to enable us to discriminate between mobility and number fluctuations. Using equations (7.31) and (7.33) the ratios y p and y~ at high B can be calculated numerically. In figure 8 the results of ya (full line) and Y N (broken line) have been plotted for 0 < ~ H <B 1. Experimental data obtained from n-type germanium at 2 cm and p~ = 3300 cm2 V-l s-1 are also plotted in figure 8. Here the 300 K with p = 20 ! dots represent the mean values of all measurements and the vertical lines indicate the standard deviation. Since the experimental data follow the line for mobility fluctuations, the conclusion is that these results give additional support to the claim that a noise is caused by mobility fluctuations.
7.3. l l f noise and lattice scattering

There is still another piece of evidence in favour of mobility fluctuations. This is the case of l/f noise in the conductance of heavily doped semiconductors where the mobility is determined by a mixture of lattice scattering and impurity scattering. Hooge and Vandamine (1978) have experimentally found the relation
SG(f

>/G2 =(aifN>(P/p1att)2

(7.35)

pimp-' and platt and p i m p are the mobilities that would be found where p-I= if only lattice or impurity scattering were present. The result will be treated in detail in $8. But here we shall use the numerical results for the discussion of the 6p-6N problem. By analogy with equations (7.6) and (7.7) we write

G = X(p~att, N ) K [Pimpplatt(pimp + p~att)-'] N= p N


so that mobility fluctuations lead to

(7.36)

(7.37) and number fluctuations to


S G / G= ~

(Xd N )

gT

sN/N2=h ( N ) a/fN.

(7.38)

If the mobility does not depend on the number of free carriers then we have a reduction factor h(N)= 1. If, according to Hilsum (1974), the relation between p and N is given by p = platt [1+ ( N / N O > ~ / ~then ] - ~ ,we find ,< h ( N ),< 1. The experimental results show that a noise is due to mobility fluctuations since reduction factors down to 10-3 have been found (see figure 9).

7.4. l/f noise o f hot electrons The investigation of 01 noise of hot electrons in n-type silicon by Kleinpenning (1981a) enables us also to discriminate between mobility and number fluctuations. The results will be presented here briefly. Consider an n-type semiconductor sample with ohmic contacts. The contact spacing is L and the cross section is A . The point is to calculate the l/f noise density at high electric-field strengths where the carrier mobility is field-depend33

512

F N Hooge, T G M Kleinpenning and L K J Vandamme

Figure 9.

CL as a function of p at 300 K. CLO and po are the values found in samples not affected by impurity scattering or surface scattering. Circles: experimental, p-type Ge. Squares : experimental, n-type GaAs. Dots: experimental, thin Bi films. Full line: theory, equation (8.12). Broken line: theory from equations (8.13) and (8.14).

ent. We make two calculations assuming mobility fluctuations and number fluctuations respectively. Then the calculated results will be compared with experimental data. Assuming mobility fluctuations the current density J ( t ) at time t is
J(t> =~p[E(x t) , , p o k t)l t>

(7.39)

where E(x, t ) is the electric-field strength at point x (O<x<L) at time t, and p is the mobility, which is a function of the low-field mobility po and E. At low fields we have p=po, and at high fields pE=vs, the saturation velocity. To calculate the open-circuit voltage fluctuations due to low-field mobility fluctuations 8p0, equation (7.39) has to be linearised around the steady state. For the situation where a constant current flows through the sample we have 6 J ( t )= 0 = (dJ/dE) SE(x, t ) (dJ/dpo) S,UO(X, t ) and

(7.40)

At low fields, where p = po, equation (7.41) becomes (7.42) Equation (7.42) yields SV = 01 P / f N .

Experimental studies on llf noise Combining this result with equations (7.41)and (7.42)yields
(po dJldpo)2 a.V2 "=

513

(E dJ/dE)2

flv'

(7.43)

According to Jacoboni and Reggiani (1979) the relation between p, po and E can be approximated by

(7.44)
with ,i3 of the order of unity. With equations (7.39)and (7.44)we find po d J/dpo =Ed J/ dE, so that equation (7.43)reduces to

Sv = a.V2/fN.

(7.45)

If we assumed that l/f fluctuations were caused by number fluctuations, we would have the following equation : (7.46) J(t) =q pCL[E(x,t>ln(x, t ) E(x, t).
For the open-circuit fluctuations the following applies : AJ(t)=O=(dJ/dE) SE(x, t)+(dJ/dn) 6n(x, t ) and thus

(7.47) (7.48)

Following the same procedure as in the paragraph above we find

(7.49)
where p(E) is given by equation (7.44)and pd(E) is the differential mobility defined by d[p(E) E l P . Hence, number fluctuations would lead to noise voltage densities that are larger by a factor p2(E)/pd2(E) compared to the noise densities from mobility fluctuations. Therefore the llfnoise study of hot electrons also gives a possibility of discriminating between 6 p and 6N. Since the experiments have to be performed at high electric-field strengths, Joule heating of the samples creates serious problems. To avoid this heating problem, samples with a hemispherical geometry were investigated. On the other hand, the noise calculations for such a geometry are complicated. The details of these calculations are omitted here. The noise is measured as a function of the voltage across the sample. The experimental data agree well with the calculated curve for mobility fluctuations and are far away from the curve for number fluctuations. As an illustration let us compare the results at the voltage where the current is one-half of what it would have been if the sample were ohmic. The noise calculated for number fluctuations is then a hundred times higher than that calculated for mobility fluctuations. The experimental results lie within a factor of two around the mobility curve. The conclusion is that a. noise of hot carriers can be described in terms of low-field mobility fluctuations characterised by equation (1 .26).
7.5. Empirical

llf noise source term

In order to describe a. noise in systems where the electric-field strength and the freecarrier density are non-homogeneous, we have to transform equation (1.26)into a l/f

514

F N Hooge, T G M Kleinpenning and L K J Vandamnze

noise source term. Here we shall derive such a source term which describes the l/f noise as a result of fluctuations in the lattice scattering. From experimental data it can be concluded that l/f fluctuations in the conductivity are spatially uncorrelated. In principle there will be a physical limit with respect to the distance at which the fluctuations are spatially correlated. In practice this distance turns out to be short compared with all characteristic dimensions of conductors and their electric contacts. Hence, the cross-correlation spectral density in the conductivity U can be written as

So(r, r , f ) = [cxu~(r)ifut(r)] 6(r-r)

(7.50)

where n(r) is the free-carrier density at spot P, u(r) is the conductivity and 6 is the Dirac delta function. From l/noise in thermo-EMF it is concluded that the l/f fluctuations in the conduction band are energetically uncorrelated. So the cross-correlation spectral density of the l/f fluctuations in the conductivity of non-degenerate semiconductors is
auZ(r, E ) So(r, r, E, E, f ) = ~6(v - r) 6(E- E) fa@,E )

(7.51)

where n(r, E ) is the density of free carriers with energy E at spot r, and u(r, E ) is the conductivity of carriers with energy E. The delta functions indicate that the fluctuations at r and r f r are uncorrelated, and so are the fluctuations at the energies E and E#E. Consequently, the electron conductivity and the hole conductivity fluctuate uncorrelatedly, which is confirmed by experiments on thermal voltage of intrinsic germanium. Kleinpenning (1978) obtained
S&, r , f ) = / / w S,(r, r, E, E , f ) dEdE=
0

(3n/8) au2(r)

6(r -r)

(7.52)

for lattice scattering. The factor 3n/8 is the result of the averaging procedure. This factor can be avoided by using (8/3n-)a in (7.51) for the description of the noise in a sub-band.

8. Lattice scattering causes U noise


After it had been demonstrated that 01 noise is a fluctuation of the mobility, further experiments were done in order to investigate the nature of the mobility fluctuations in more detail. An obvious question is then: does l/f noise depend on the kind of scattering? The simplest picture explaining the finite value of the mobility is that each electron, moving around in a Brownian motion at an average thermal velocity, is accelerated by an applied electric field until it collides with some scattering centre. After the collisions there is no memory of the previous motion. In this way electrons obtain a limited drift velocity proportional to the field E , which leads to the definition of mobility
(U)=(Udrift)E

(8.1)

Simple mechanical considerations lead to


p =qr/m

(8 2)

where 7 is the average time that the electron is accelerated between two collisions. The reciprocal of this collision time is proportional to the probability that the electron will be

Experimental studies on llf noise

515

scattered. Hence, if two independent scattering mechanisms are simultaneously active, we shall, for the effective collision time T, find
l/T=

1/T1+ 1/Q.

As a direct consequence l / p = l / P l + 1/p2. Of the many types of scattering three are important in semiconductors and metals : lattice scattering, impurity scattering and surface scattering. Lattice scattering very often dominates other mechanisms. The electrons are then scattered by phonons of the lattice vibrations, either optical or acoustical. Electrons can also be scattered at electrically charged defects in the lattice. This so-called impurity scattering dominates in highly doped semiconductors, especially at low temperatures. In thin layers nearly inelastic scattering occurs at the surface. This surface scattering dominates when the layers are thinner than A, the free path between two collisions in a thick sample of the same material. If there are two scattering mechanisms, each with its own noise, then the observed noise in the effective p will be

as follows from differentiating equations (8.4). Because of (1 .26) this leads to


al+

);(

2
012.

We shall show that all experiments can be described very well by assuming llfnoise in the lattice scattering and very little or no llfnoise in any other type of scattering. In other words, we shall demonstrate that it is correct to take

Hence (8.10) The first demonstration of the correctness of (8.10) was the measurement of cy. on semiconductors with lattice scattering and impurity scattering (Hooge and Vandamme 1978). The noise was measured on ohmic metal-semiconductor point contacts. Three p-type Ge crystals were used with p=3.0 x 10-2 Q cm, 3.7 x 10-3 Q cm and 4.5 x Q cm. The mobility in these samples is lower than platt, the value found in high-ohmic material because of the additional impurity scattering. A few point contacts of each crystal were investigated. The results are presented in figure 9. These experimental results, together with results from n-type GaAs with p =2.7 x 10-3 Q cm, agree with relation (8. IO) represented by the full line. For a further discussion see Weissman (1980) and Kleinpenning (1981b). A large reduction of the mobility by the influence of impurity scattering will only occur in highly doped material. The carrier concentration will then be high, which makes it necessary to use very small samples. Even then the noise is low. This is why in all previous measurements of a low-doped material was chosen, where lattice scattering always prevailed, resulting in a values close to

516

F N Hooge, T G A 4 Kleinpenning and L K J Vandamme

Another demonstration of lattice scattering as the source of lif noise is given by noise measurements on thin bismuth films (Hooge et a1 1979). The influence of surface scattering is often described by a relation for free path X
l / X = l/Xo+k/t

(8.11)

where ho is the bulk value for the free path found in an infinitely thick sample, t is the thickness of the sample, and k is a constant with an order of magnitude of unity. The surface influence is representable here by an effective free path t/k. Since (8.11) is very similar to (8.4) the result for a will be
01 = (p/po)2 010

(8.12)

by analogy with (8.10). The subscript refers to bulk values. An advantage of equation (8.12) is that it does not depend on k, the numerical value of which has not been agreed upon. Equation (8.12) gives the full line in figure 9. Since equation (8.12) is only a reasonable guess, one may ask whether it is good enough to permit calculations of the noise. Therefore the calculations were also made for a model in which the influence of 0 was taken into account. The angle 0 is the angle between the velocity of the electron and its distance to the surface. Averaging over all values of 0 and all values of the distance to the surface gives a relation for h as a function of t much more complicated than equation (8.11) : x-lexp(-x)dx Differentiating (8.13) gives an expression for the noise: (8.14) From (8.13) and (8.14) follows the dependence of a on X and thereby of a: on p. This dependence has been plotted as the broken line in figure 9. Figure 9 shows that when p N po the value of a depends quadratically on p and that (8.12) is a good approximation. When p<<po the quadratic p dependence changes into a less steep dependence. Of course, it is possible to introduce further refinements in the model. But figure 9 shows that the simple treatment is already enough to demonstrate the main point: surface scattering is free of l/noise. There is more evidence for the correctness of equation (8.10) and the lattice scattering model than the experiments on well-defined simple samples presented in figure 9. This evidence stems from the successful application of (8.10) to more complicated samples such as electronic devices. It is true that for each application, taken on its own, arguments can be put forward to the effect that such an application does not provide definite proof of the correctness of (8.10). The main argument is always: how certain is it that the model used in the calculation of the theoretical results is a true description of the actual device from which the experimental results were obtained? Usually there are so many parameters involved that another choice of numerical estimates could also be made. Nevertheless, many experimental findings inexplicable before can readily be explained by the application of (8. lo). All the cases collected from different devices together provide further support for the lattice scattering model. In lo the application to MOS transistors will be presented. Here we shall discuss Becks (1979) experiments on ionimplanted resistors. Let us, therefore, first consider a rectangular sample with length I,

rXoF. (8.13)

Experimental studies on llf noise

517

width w and thickness t. The carrier concentration strongly depends on x, the distance from the surface (0 < x .c t ) . In the low-concentration regions lattice scattering will prevail, and in the high-concentration regions we shall have impurity scattering or electron-electron scattering. For the noise in the conductance parallel to the surface we find (8.15) as follows from applying (8. IO) and (1 .26) to homogeneous layers with thickness dx. This complicated relation must be used when local values of n give local values of p differing from plait,which in turn leads to local values of 01 differing from 2 x 10-3. If we had no reduction in 01 then the result would be (8.16) which does not differ so much from (1.26) the simple result of the homogeneous case. Beck used silicon resistors where a conducting p-type channel with a Gaussian concentration profile was formed by ion implantation. The profile, and thus the resistance, could be varied by a reverse bias voltage applied to the substrate. According to the simplest relation (1.26) one would expect the relative noise to be inversely proportional to N and hence proportional to R. Experimentally the relative noise is proportional to R2. This cannot be explained by a (8.16)-type relation. Only (8.19, based on 01 reduction and the lattice scattering model, gives the right result. The general picture following from the experiments described in this section is that lattice scattering alone gives llfnoise. It is as if the density of the phonons fluctuates with a I / f spectrum. These phonon fluctuations occur independent of the presence of an electric current through the sample. This gives the solution to the old problem of the spectrum being llfeven at frequencies much lower than the reciprocal of the time that an electron stays in the sample. The fluctuations in different phonon modes are not correlated, otherwise we would have fluctuations in temperature, which is not the case as demonstrated in $6.
9. Non-ohmic junctions

In the previous sections the l/fnoise has been treated in semiconductor and metal structures, where the current-voltage characteristic is ohmic. Up to now we have restricted ourselves to semiconductor structures where the distribution of free carriers is homogeneous. In this section we shall treat the llfnoise in structures with a non-homogeneous free-carrier distribution. The following devices will be discussed. (i) Single-injection space-charge-limited-current solid-state diode (SCLC diode). Such a diode consists of p- or n-type material in which the dope concentration is low in one part and high in the other part. Two electric contacts are applied, one to the high-dope part, the other to the low-dope part. So we have a p+-p junction or an n+-n junction.

518

F N Hooge, T G M Kleinpenning and L K J Vandamme

(ii) p-ii junction diode. Here we have a semiconductor structure in which one part is p-type and the other n-type. (iii) Schottky barrier diode. Here we have a semiconductor material in contact with a metal. The metal-semiconductor junction has rectifying properties with respect to the current-voltage characteristic. It should be noted that metal-semiconductor and semiconductor-semiconductor junctions occur abundantly in modern electronic devices and integrated circuits. Since l/fnoise often causes trouble in electronic devices, the aim is to minimise this noise. For that purpose a deeper understanding of the phenomenon of llfnoise is required. In view of the present state of the theory, the llfnoise of non-ohmic junctions will be discussed mainly in terms of mobility fluctuations. There have been many attempts to interpret the lifnoise in Schottky barrier diodes and p-n diodes in terms of free-carrier fluctuations (McWhorter model). We shall omit a discussion of this problem for two reasons: (i) the conclusion of $7 on Sp or SN, and (ii) the recent discussion by Van der Ziel (1979) in an extensive survey paper. Besides, it should be noted that in the past much experimental research has been done on diodes of poor quality. For instance, the observed current-voltage characteristics do not obey the theoretical predictions owing to edge effects. In this section we restrict ourselves to the llfnoise of p-n and Schottky barrier diodes obeying the ideal I-V relation given by Z=I~[exp (qV/kT)- 11 (9.1)

where the experimentally observed values for the saturation current IO agree with the values found theoretically. In the table given below, a summary is presented of the various types of junctions treated in this review.
Junction Metal-to-metal Metal-to-semiconductor I-Vrelation Ohmic contact; Zcc V Ohmic contact; Zcc V Rectifying contact; equation (9.1) (Schottky barrier diode) Ohmic contact; Zcc V Injecting contact; IKV2 (SCLC diode) Rectifying contact; equation (9.1) (p-n diode) Section(s) 1,4
1, 4 9

Semiconductor-to-semiconductor

1, 4 9 9

To calculate the cy. noise of non-ohmic junctions we need an a-noise source term for the mobility fluctuations. Since the junctions will have a nearly planar geometry, the treatment will be one-dimensional. In this case the cross-correlation spectral density in the conductivity reads (see equation (7.50))

where A is the cross section of the device, and n(x) A is the number of free carriers per unit length at spot x. Since cy. noise is due to mobility fluctuations, the cross-correlation spectral density in the lattice mobility for a one-dimensional case is

Experinzental studies on l / f noise


9.1. Single-injection space-charge-limited-current diode

519

In a review paper Nicolet et a1 (1975) have discussed the noise in single- and doubleinjection solid-state diodes. Although the noise spectra of injection diodes often show a l/f noise component, the discussion was mainly on thermal noise, trapping noise and generation-recombination noise. They did not give a discussion of the l/f noise because of the absence of literature on this subject. Kleinpenning (1978) has discussed the l/f noise in single-injection diodes in terms of mobility fluctuations. Were we shall present his results for a trap-free diode with a plane-parallel geometry. At low voltages the current-voltage characteristic is linear (ohmic regime) and at high voltages it is quadratic (space-charge regime). Consider a p-type semiconductor with two hole-injecting (p+) contacts. The contact spacing is L and the cross section is A . So we have a p+pp+ structure. Diffusion currents are neglected as is usual in injection diodes. Since only low-frequency fluctuations are considered, dielectric relaxation effects can be also neglected. The current density J ( t ) at time t is given by J(t)=qpLp(x, t>p(x,t ) E ( & t ) (9 * 4 ) where q is the elementary charge, pp(x,t ) is the hole mobility at point x ( O < x < L ) at time t , p(x, t ) is the hole density and E ( x , t ) is the electric-field strength. The mobility is assumed to be independent of the electric-field strength, and the current density is independent of the position x. In addition we have Poissons equation

where E is the dielectric constant, pinj is the density of injected holes, and p t h is the density of the thermally generated free holes. We are interested in the AC open-circuit voltage fluctuations due to mobility fluctuations. Therefore, we consider fluctuations around the steady state by linearising equations (9.4) and (9.5). For the situation where a constant current flows through the sample we have

ax

t)=l [ A p ( x , t)-Apth(x, E

t)]

in which p ( x ) = (p(x, t ) ) , E ( x ) = ( E ( x , t ) ) and p p = (pP(x,t ) ) . For diodes operating in the space-charge-limited-current regime we have Pinj %p$pth, so that equation (9.7) can be approximated by

Substituting equation (9.8) into (9.6) we find the differential equation

The distributions of holes p(x) and the electric field E ( x ) are given by (9 * 10) where V = V(0)- V(L) is the mean voltage. The position of the injecting contact is at

520

F N Hooge, T G M Kleinpenning and L K J Vandanzme

x =O. Using equation (9.10) and the boundary conditions E(0)= 0 and AE(0, t ) =O the solution of equation (9.9) is found by the method of variation of parameters

(9.11)
The
AC

open-circuit voltage fluctuations follow from AV(t)=

-Jf

AE(x, t ) dx.

(9 12)
I

Making a Fourier transform from time to frequency we obtain

(9.13)
where V ( f ) ,E(x,f) and p p ( x , f ) are the Fouriers transforms of AV(t), AE(x, t ) and App(x, t ) , respectively. Partial integration of equation (9.13) yields
(9.14)

Hence, the spectral density of the

AC

open-circuit voltage fluctuations is given by


x ' , f ) dx dx'.

9v2 I , S V ( ~ ) = (~ d L~ - d~ x )( d L - dx') S ( , x , PPL 0

11

(9.15)

Using equations (9.3), (9.10) and (9.15) the AC open-circuit l/f noise in the spacecharge regime is found to be Sv(f)=olqLV/Sf&A. (9.16) The steady-state current-voltage relation in the space-charge regime is

(9.17) (9.18) (9.19)


With regard to the I-V characteristic, the transition voltage V R between the ohmic and the space-charge regime can be found from equations (9.17) and (9.19)

v~=8qL~pth/9&,

As regards the noise, the transition voltage V N can be found from equations (9.16) and ( 9 . 18) V N qL2pth/5E .
The ratio ,8 between the two voltages is given by

,8 = vR/VN =40/9 4.4.

(9.20)

If the bulk is intrinsic equations (9.16) and (9.17) are unchanged. In the ohmic regime we find with b = p n / p p (9.21)

Experimental studies on llf noise In this case we obtain

521

p=- 40 b2+
~

1 9 b+l'

(9.22)

The results given in equations (9.16), (9.17) and (9.22) are confirmed by experiments on p+ip+ diodes where the bulk is intrinsic. For a typical diode figure 10 shows the llf noise density and the current as a function of the bias voltage. For three diodes the mean of the observed ratio /3 was found to be 6 ; equation (9.22) predicts p-9. The observed 01 values are in agreement with the constant 01= 10-3.
I I

I
0.1

l
V

I
10

Figure 10. Experimental I-V and SV-V characteristics of a SCLC diode.

9.2. p-n diode

In an extensive paper Van der Ziel (1979) has discussed the llfnoise in p-n diodes in terms of fluctuations in the surface recombination velocity. Here the surface recombination velocity is assumed to be modulated by the fluctuating occupancy of traps in the oxide adjacent to the oxide-semiconductor interface. The interpretation is based on the McWhorter model. However, in view of the conclusion of $7 concerning mobility and number fluctuations we shall discuss the llfnoise in p-n diodes in terms of mobility fluctuations. We follow a paper by Kleinpenning (1980a) who has discussed several types of p-n diodes. Here we restrict ourselves to the presentation of llfnoise investigations on a long p+-n diode operating in the diffusion current region. Consider a long p+-n diode where the base extends from x = 0 to x = W. For x < 0 the material is p+-type. The base W is much thicker than the hole diffusion length L , in the n-type base, so that most of the holes diffusing into the base do not reach x = W. Under forward bias voltage V~$kkT/q, the hole and electron currents Z , In and the hole density Pn in the base are given by

Zp(x)= -ql>,Adp/dx =IF exp ( - x/Lp)


Pd-4 =Pn(O) exp ( -x / w .

In(x) = I F - Ip(x)

(9.23) (9.24)

522

F N Hooge, T G M Kleinpenning and L K J Vandainme

To calculate the l / f noise, we divide the base into two parts. In the first part O<x<L, the hole current is dominant, and in the second part L, < x < W, the electron current. According to (9.23) it is obvious that p p fluctuations lead to D, fluctuations and so to fluctuations in I,. On the other hand, we have L, = (DP~,)1/2where T~ is the hole lifetime in the base, so that p, fluctuations lead to fluctuations in the hole density (see equation (9.24)). Since dp/dxccL,-loc DP-li2, we have Ip(x)cc(pp(x))l~z.Fluctuations in the mobility App(x, t ) lead to current fluctuations such as AI,(x,

0 =+(IP(wpP)

ApLp(x9 t).

(9.25)

Since the current is independent of x, we have

AI,(t) z ___ I,(x) Ap,(x, t ) dx. (9.26) 2LPPP 0 Using (9.3) and the approximations I , =IFand p ( x ) =p(O) in the region 0 < x <L,, the current noise is found to be

lLD

SIF=oIIFF2[4fALpp(0)]-1=olgzF(4 fTp)-.

(9.27)

Here we have used the Shockley equation ZF=qAp(o)L,/T,. Since the shot noise is 2 q b , the corner frequencyfc, where the llfnoise equals the shot noise, becomes
fc =01/8~p.

(9.28)

The contribution of the electron current in the base between Lp < x < W to the noise may be neglected at relatively low currents because the electron density in the base is much higher than the hole density p(0). The calculated results are confirmed by experiments on n+-p diodes with Tn = 0.5 ps. Using a= 10-3 there is full agreement between calculated and observed corner frequencies.
9.3. Schottky barrier diode

The lifnoise in Schottky barrier diodes have been studied by a number of investigators. The papers published up to 1978 were discussed by Van der Ziel (1979). Two models are given for the origin of the l / f noise; both have the McWhorter model as their basis. In one model the main source of the l/f noise is supposed to be in the edge current. Here the carriers interact with the surface oxide and the l/f noise density is related to the surface state density. In the other model, introduced by Hsu (1971), there are no surface effects. The l / f noise is interpreted in terms of trapping and multistep tunnelling processes in the depletion region of the barrier. Both processes introduce a random fluctuation in the occupancy of the trap centres. As a consequence, the electric field at the metal-semiconductor interface fluctuates and this leads to barrier height fluctuations and therefore current fluctuations. The main objection to the two models is once more the conclusion of $7: l/fnoise is due to mobility fluctuations. A discussion of the lifnoise in terms of mobility fluctuations has been given by Kleinpenning (1979). For Schottky barrier diodes where the current is completely determined by thermionic emission of carriers, he concluded that 01 noise cannot be present in ideal diodes. Here the I-V characteristic is independent of the mobility of free carriers and is given by I=Io[exp (gV/TkT)- 11 (9.29) where the ideality factor y~ is equal to 1. However, the series resistance of the diode, edge currents and diffusion effects in the depletion region can generate l/f noise. More-

Experimental studies on

llf

noise

523

over, these effects lead to an increase in the ideality factor 7. We shall briefly discuss these deviations from the ideal case. The series resistance is related to the conductivity of the substrate and shows l/f noise according to equation (1.26). The influence of the series resistance on the I-V relation finds expression in a slower increase of the current with increasing voltage, so that the ideality factor becomes larger than 1. The current transport in Schottky barriers is determined not only by thermionic emission of carriers across the top of the barrier but also by diffusion of the carriers in the depletion region. Since the diffusivity of carriers is related to their mobility, mobility fluctuations lead to current fluctuations. This diffusion mechanism leads also to an increase of the ideality factor 7. It has been suggested that l/f noise is generated at the periphery of Schottky barriers. Zettler and Cowley (1969) have made Schottky barriers with a p-n junction guard ring. Such a ring improves the noise properties and the ideality factor. The equivalent circuit of this structure is composed of the main Schottky barrier in parallel with a small p-n junction. Although the I-V relation is determined by the Schottky barrier, there is also a slight current flow in the p-n junction. It is known that p-n junctions show l/f noise. For an ideal Schottky barrier with a p-n junction guard ring we may expect that the I-Vrelation is determined by the current flow in the Schottky barrier and the l/f noise by the current in the p-n junction. Wall (1976) found experimentally that the l/f noise was generated at the edge of the diode. Besides, he found that edge effects lead to an increase in the ideality factor 7. So the llfnoise density increases with increasing 7 . A similar correlation between l/f noise and 7 was experimentally found by Peyriguer et aZ(l973). In view of the mobility fluctuation hypothesis it is expected that the llfnoise density and the deviation of the ideality factor from unity are correlated. This is supported by the results plotted in figure 11. Here the l/f current noise density has been plotted against the ideality factor 7 - 1 for various Schottky barrier diodes. Figure 11 supports the statement that ideal Schottky barriers do not generate 01 noise.

10.

MOS

transistors

In a metal-oxide-silicon transistor (MOST) the current is passed between the so-called source and drain contacts which have an impurity dope of a type different from the substrate. The current is carried by charge carriers with a polarity opposite to that of the majority carriers in the substrate. The inversion layer thickness is about 20 nm and is embedded in between the Si-Si02 interface and a depletion layer which isolates the current path from the substrate. Figure 12 shows a diagram of the Mosr device. The total number of carriers in the channel can be controlled by the voltage on the gate electrode. Two different types of models are used to explain the experimental llfnoise results in MOST. The oxide trap models are based on McWhorters theory of tunnelling into oxide traps, while the bulk model uses the relation (1.26) and the fact that competing scattering mechanisms other than lattice scattering lead to a reduction in llfnoise ($8). In $10.1 the oxide trap model will be treated. The noise sources are in the oxide. Therefore the coordinate x is chosen perpendicular to the oxide and pointing into it. In $10.2 the coordinate x is perpendicular to the oxide and pointing into the substrate in order to treat the bulk model, where the noise sources are in the substrate. The models based on sources at different sides of the Si-Si02 interface will be compared in $10.3. The MOST is commonly used as an amplifier. Therefore the noise in the drain current

524

F N Hooge, T G M Kleinpenning and L K J Vandamme

10

J
N 0

$1 0-

IO-

102

lo-i

11 -1 Figure 11. Spectral density of llfcurrent noise plotted against ideality factor 7- 1 for various Schottky diodes. For the origin of the data see Kleinpenning (1979). T= 300 K, I = 100 PA, f= 10 Hz.

is often expressed as equivalent noise voltage at the gate input. The equivalent input noise S,, is defined by

S I svg - (dI/d V Q ) ~

(10.1)

where SI is the short-circuit spectral density of the drain current at a constant drain voltage VD and (dZ/dVG), the so-called transconductance gm at the same constant drain effective gate voltage VQ) voltage VD. When the MOST is biased at saturation ( V ~ > t h e then gm=(w/l) ~ C O V G The . equivalent input noise voltage is then denoted by Svgs.
10.1. Oxide trap model

Some formulae for l/f noise in MOST will be presented based on McWhorters theory of tunnelling ($5.1) and on the theory of generation-recombination noise. All oxide trap models have in common that the noise is attributed to fluctuations in the number of carriers in the channel, but they differ in details. The fluctuating occupancy of the oxide traps modulates: (i) the number of free carriers in the channel following McWhorter, and (ii) the surface potential, causing fluctuations of the mobile channel charge following Hsu (1970). Christensson et a1 (1968) considered direct tunnelling of free carriers into oxide traps. Fu and Sah (1972) suggest the following reversible transition. A free electron and an empty fast surface state at energy E gives a filled surface state. A filled fast surface state and an empty oxide trap at about the same energy level at x gives a filled trap. This transition is by tunnelling and results in large time constants depending on x ($5.1). The differential equation governing the interaction between the surface states and the oxide traps is of the form of equation (1.8) where AX must be replaced by the average

Experimental studies on l / f noise

525

Figure 12.

MOS

transistor with coordinates x, y and z. The channel length, width and thickness are denoted by I, w and S. S is the source, D is the drain.

number of electrons in traps with energy E and lying at distance x from the interface and where T is the time constant corresponding to x as given by equation (5.2). The corresponding generation-recombination spectrum given in equation ( 1 .16) for the occupancy of those traps with equal T at depth x becomes (10.2) Here the average of the squared fluctuations in the occupancy of those traps, ((Ant)2), is replaced by Nt(E, x ) f l ( l - f l ) . Nt(E, x ) is the density of traps at E and x . ft is the Fermi factor [l +exp ( E - E F ) / ~ T ]for - ~ those traps. This means that the trapping process is assumed to obey binoniinal statistics. The function ft( 1-ft) peaks around the Fermi level. In consequence, only the traps within a few k T of the Fermi level contribute to the fluctuations. We now assume a uniform distribution of traps with respect to x and an arbitrary distribution of traps with respect to the energy E. We integrate over the oxide thickness, keeping in mind the T dependence on depth (McWhorter model). The resulting l / distribution ~ function g(T) is given by equation (5.3). By analogy with equation ( 1 .23) this yields
(10.3)

where T I corresponds with the maximum integration depth in the oxide, and where Nt(E) stands for the number of filled oxide traps over the whole range of x with energy E. Next we integrate with respect to E. Introducing the effective trap density at the Fermi level as N t , e f f = J T zN t ( E ) f t ( l - f t ) d E we find S n t t f ) KNt, e d f (10.4) for the spectrum of the trapping fluctuations of all traps over the whole channel area. If Nt(E) changes slowly near E= EF then Nt, eff = Nt(EF) kT. If we assume that the surface state density Nss is proportional to the effective oxide trap density Nt, eff we have given a theoretical basis for the experimental results that the noise in a MOST is proportional to N,, as has been found, for example, by Broux et a1 (1975). Simple expressions for the noise were derived by Pai (1978), assuming that the fluctuation in occupancy of traps corresponds to the fluctuation in the fast interface states and in the channel charge,

526

F N Hooge, T G M Kleinpenning a n d L K J Vandamme

so SNt= -6N in the channel. His recalculation of Klaassen's (1971) result for the noise gives s v g s = p q vG/2wlcOf (10.5)
where ,8 is independent of VG and proportional to N,, with a value calculated from Klaassen's (1970) experimental results of about 5.5 x lop4 at VG-3 V. Pai (1978) also derived another expression in which Svgs depends only on VGthrough N B s ( E ~i.e. ), (10.6) where N s s ( E ~is ) the value of the interface state density at the Fermi level. COis the oxide capacitance per unit area. 10.2. a-noise model The source of the l/f conductivity fluctuations in this model are the fluctuations in the mobility ($97 and 8). F,elation (1.26) is applied to sublayers in the MOST parallel to the interface in which the concentration profile n(x) and the mobility profile p(x) are assumed to be homogeneous in the y z plane. If we are to compare experimental results of MOST with the empirical relation (1.26) it is convenient to replace N by I2/qpR, where R is the sample resistance, and replace a: by alatt(p/platt)z:
SV - mlatt(p/platt)'~4 @- alattq p*.l/fR -V2 12f 12f .

(10.7)

In this way the noise mobility is introduced to take into account the reduction in a: in comparison with alatt. For homogeneous samples pl/if= p3/p'1att and pI/fd p d latt. In a MOST the concentration and mobility are inhomogeneous in the x direction. Even if by neglecting this inhomogeneity, the llfnoise in MOST is described by equation (10.7), agreement between experiments and calculations is often found, at least at high channel resistances. However, there are often deviatioiis between the measured l/f noise and values predicted by equation (10.7) at low values of R. Therefore, the homogeneous channel is divided into sublayers with thickness dx. The channel conductance is obtained by adding the conductance of all sublayers. No correlation between the noise in the sublayers is assumed. Relation (1 ,26) is applied to each sublayer and pl/f is calculated as (Vandamme 1980a)

(10.8)
by analogy with (8.15). A mobility profile starting from a low value p(0) at the interface and a concentration profile decreasing with depth lead to a reduction in pl/f. The effective or surface conductance mobility perf was defined by Sze (1969) as (10.9) From experimental results of the conductance against gate voltage we can find peff against VG. In MOST modelling Merckel et a1 (1972) introduced the following expression for peff:
(10.10)

Experimental studies on llf noise

527

where po is the mobility at low VGand e is called the mobility degradation factor. This mobility degradation factor 0 will be used here to distinguish between different categories of MOST. Vandamme and de Werd (1980) demonstrate that the voltage fluctuations in the drain voltage at large drain currents comprise two contributions: (i) mobility fluctuations and (ii) number fluctuations of the charge carriers due to fluctuations of the effective gate voltage induced by the above-mentioned mobility fluctuations. At saturation with VD 2 VGthe equivalent input noise voltage becomes
(IO. 11)

Vandamme (1980a) used simple expressions for n(x) and p(x) in equations (10.8) and (10.9) from which followed analytical expressions for peff and pllf. For the n profile

n(x) =n(O) exp ( - x / t )

for

O c x <4t=6

(10.12)

was used as an approximation of the quantum-mechanical solutions obtained by Gnadinger and Talley (1970), Stern (1972) and Hsing (1977). Here n(0) is the carrier concentration (cm-3) at the interface and t is about the average distance for the carriers from the interface. The value of n(0) increases and t decreases with increasing gate voltage. At present, no theoretical model for the mobility p(x) is known. The following relation for p(x) is proposed:
p(x) = platt [I - exp (- x/v)l

(IO.13)

where v is a characteristic length related to the mean free path of the charge carriers in f then follows that the bulk. For peff and p ~ /it (10.14)
platt

pl/f=(l + v / 2 t ) (1 + v / 3 t ) (1 +Y/4t)'

(IO. 15)

An increasing gate voltage produces an increase in u/t at large n(0) and a decrease in pelf and an even stronger decrease in pl/f. These decreases are due to a decrease in t with increasing gate voltage. Vandamme (1980b) tried to take this channel shrinkage into account with the following expression

t = to( 1 +n(0)/nc)-1/2

(10.16)

where t o is of the order of 10nm and nc is a characteristic concentration of at least 1017 cm-3, Higher impurity concentrations in the substrate or lower temperatures will result in larger nc values and, in consequence, lower degradation factors 8. At 77 K the exponent 3 in equation (10.16) must be reduced to about 4 while to is smaller compared to the 300 K value. The experimental plot of conductance against the effective gate voltage is characterised by po and 0. po is determined from the initial slope in the G against VGplot while 0 is obtained from a curve-fitting technique. The relations between the model parameters Y and nc and the experimentally observed parameters po and e are given by
PO platt (1

Y/~O>-'

(10.17)

(IO. 18)
34

528

F N Hooge, T G M Kleinpenning and L K J Vandamme

Knowing po, 0,p l a t t and the gate oxide capitance COper unit area and taking to= 10 nm we can calculate U, nc and the llfnoise from the above equations. Three classes of MOST can be distinguished according to 0,po and the experimentally observed noise. (i) Category I contains the MOST with low 6' and higher values of pO/p*.latt. As a conseVG for quence, experimentally as well as by calculations we find S v / V z R and Svgscc 6 '< 0.03 V-l and p o / p l a t t 2 0.4. (ii) Category IIa contains MOST with higher 6 'values (6'> 0.03 V-l) and lower p,o/vlatt. This leads to SvlV2ccRy with y - 2 and S v g s c c V ~ with ? ~ ~ 0 . 7 . (iii) Category IIb MOST have the same characteristics at 300 K as category IIa as far 'is concerned. However, at 77 K category IIb shows a strong increase in as p o / p l a t t and 6 6 'or a maximum in the plot of G against VG. In the following table the three categories are summarised.
Category

8 at 300 K
8vG< 1

po/patt

at 300 K

877/8300

Sv/V2ccRr Svgscc Vd
y=l
y22
y"3

>0.4

41 <1 >1

7=1 7<0.7

8q0.03V-I

iL

} {8VGNl b0.03
V-1

<0.4

- 0 . 2 < 7<0.7

The fourth column gives the ratio between 6 'at 77 and 300 K. This criterion discriminates between category IIa and IIb. The category to which a MOST belongs can be seen from the G against VGplot at 300 and 77 K that gives 6 ' and po values. MOST belonging to category I and IIa show agreement between calculated and experimentally observed noise in spite of a measured surface state density ranging from 3 x 1010 cm-2 eV-1 to 2 x 1012 cm-2 eV-l for diflerent samples. For category I and IIb MOST,6'77 is smaller than 8300 which is in agreement with a less pronounced channel shrinkage at 77 K than at 300 K. The category IIb MOST show 100 times less noise than predicted by the model and the ' with decreasing temperature also can not be understood with this simple increasing 6 model. Experimental results and lines calculated from G against VG plots are presented in figures 13 and 14 and show agreement. Experimentally, category I MOST show larger VT and large net impurity concentrations. This is also in agreement with the model in which large impurity concentrations go hand in hand with large values of the model parameter nc and therefore with low 6 'values. Modern MOST have lower thresholds and thin gate oxide layers leading to a higher 6 'value. In general they belong to category IIa or IIb.
10.3. Comparison of the two models

In order to examine the validity of the oxide trap model a survey was made of the experimentally obtained noise and the density of the surface states. That survey did not then the survey is an argument show a proportionality between noise and Nss. If N s s ~ N T against the oxide trap model. Vandamme (1980b) compared two relations, one based on the oxide trap model and the other on the a-noise model. We can calculate Svgs according to the &-noise model with equation (10.11) and according to the oxide trap model

Experimental studies on I l f noise

529

IO"'

c
io3 R(R)
104

Figure 13. llfnoise of category I MOST. The full lines were calculated from parameters obtained from G against VG plots. Dots were obtained experimentally at 300 K. Circles were obtained experimentally at 77 K. The noise was measured in the ohmic region.

(Klaassen 1971) with equation (10.5). The ratio RI of the two Svgs expressions is

which is of the order of 1. This means that the experimental results for Klaassen's MOST and category I MOST in general are in agreement with both models.

10"

1 V,CVl

10

Figure 14. Experimental and calculated values of Svgs of the category IIa MOST at 300 K. The n-channel MOST was biased in the saturation region with VD=1.5 VG.f= 1 kHz.

530

F N Hooge, T G A 4 Kleinpenning and L K J Vandamme

The noise in category I1 MOST can be described with equation (10.6). The ratio of the results of the 8p model and the Sn model becomes (10.20) which is also of the order of 1. Again, experimental results can be interpreted both by Sp and Sn models. Noise measurements on MOST are no more a proof of the correctness of the oxide trap model which originally had had reasonable success in MOST research. 11. Summary and conclusions The majority of results of l/f noise measurements can be described by an empirical relation S G / G= ~a:/NJ (1 .26) Noise obeying this relation is called a: noise. It has been proved experimentally that 01 noise is a fluctuation in the conductance (92) caused by fluctuations in the mobility ($7). Measurements on samples where lattice scattering is mixed with some other scattering mechanism show that only lattice scattering causes a: noise (@). In consequence, the constant a in (1 .26) becomes
( 9 ( 2 x 1 0 3 ) .
platt

(8.10)

Of the theories proposed in the last few decades two still survive. (i) McWhorters theory, based on generation-recombination in surface states. Originally proposed for semiconductors its importance is now with the noise in discontinuous metal films ($5). / (ii) Clarke and Vosss theory based on local temperature fluctuations leads to a 1 like spectrum which is possibly found in continuous metal films in a limited frequency range around the reciprocal of the thermal relaxation time ($6). In the large field of llfnoise in semiconductors we have no accepted theory. There is, however, strong experimental evidence for a llfnoise source in the lattice scattering ($8). The llfnoise in electronic devices can well be described by the empirical relation for a noise ($9). This, in turn, gives additional support to the validity of that relation. With MOS transistors the situation is ambiguous in the sense that many experimental data agree both with the McWhorter model and with the relation for cy noise. Therefore MOST data do not provide decisive arguments for or against any l/fnoise model ($10).

References
Beck H G E 1979 Solid St. Electron. 22 475-7 Beck H G E and Spruit W P 1978J. Appl. Phys. 49 3384-5 Bell D A 1960 Electrical Noise (London: Van Nostrand) pp240-3 Bergero F, Mazzetti P and Montalenti G 1968 Nuovo Cim. B Ser. X 56 62-70 Berz F 1970 Solid St. Electron. 13 621-47 Boudry M R 1973 Appl. Phys. Lett. 22 530-1 Broux G, Van Overstraeten R and Declerck G 1975 Electron. Lett. 11 97-8 Butterweck HJ 1975 Philips Res, Rep. 30 316*-21* (special issue)

Experimental studies on llf noise

531

Celasco M, Masoero A, Mazzetti P and Stepanescu A 1978 Phys. Rev. B 17 2553-63,2564-74 Christensson S, Lundstrom I and Svensson C 1968 Solid St. Electron. 11 797-812 Clarke J and Hsiang T Y 1976 Phys. Rev. B 13 4790-800 Clarke J and Voss R F 1974 Phys. Rev. Lett. 33 24-7 Declerck G, Van Overstraeten R and Broux G 1973 Solid St. Electron. 16 1451-60 Eberhard J W and Horn P M 1978 Phys. Rev. B 18 6681-93 Fu H S and Sah C T 1972 IEEE Trans. Electron Devices ED-19 273-85 Gnadinger A P and Talley H E 1970 Solid St. Electron. 13 1301-9 Hanafi HI and van der Ziel A 1978 Physica B, C 94 351-6 Hilsum C 1974 Electron. Lett. 10 259-60 Hooge IFN 1969 Phys. Lett. 29A 139-40 -1972 Physica 60 130-44 Hooge F N and Hoppenbrouwers A M H 1969a Phys. Lett. 29A 642-3 __ 1969b Physica 45 386-92 Hooge FN, Kedzia J and Vandamme L K J 1979 J. Appl. Phys. 50 8087-9 Hooge F N and Vandamme L K J 1978 Phys. Lett. 66A 315-6 Hoppenbrouwers A M H and Hooge F N 1970 Philips Res. Rep. 25 69-80 Hsing C T 1977 PhD Thesis University of Florida Hsu S T 1970 Solid St. Electron. 13 1451-9 -1971 IEEE Trans. Electron Devices ED-18 882-7 Hsu S T, Fitzgerald D J and Grove A S 1968 Appl. Phys. Lett. 12 287-9 Jacoboni C and Reggiani L 1979 Ado. Phys. 28 493-553 Jones B K and Francis J D 1975a J. Phys. D: Appl. Phys. 8 1172-6 -1975b J. Phys. D: Appl. Phys. 8 1937-40 Kedzia J and Vandamme L K J 1978 Phys. Lett. 66A 313-4 Retchen M B and Clarke J 1978 Phys. Rev. B 17 114-21 Klaassen F M 1970 Philips Res. Rep. 25 171-4 -1971 IEEE Trans. Electron Devices ED-18 887-91 Kleinpenning T G M 1974 Physica 77 78-98 __ 1976a J. Phys. Chem. Solids 37 925-8 __ 1976b Physica B, C 84 353-61 __ 1978 Physica B, C 94 141-51 -1979 Solid St. Electron. 22 121-8 -1980a Physica B, C 98 289-99 __ 1980b J. Appl. Phys. 51 3438 -1981a Physicu B, C 103 340-4 __ 1981b Physica B, C 103 345-7 Leuenberger F 1967 Electron. Lett. 3 400 J,eventhal E A 1968 Solid St. Electron. 11 621-7 Liu S H 1977 Phys. Rev. B 16 4218-23 Lorteije J H J and Hoppenbrouwers A M H 1971 Philips Res. Rep. 26 29-39 MacDonald D K C 1962 Noise and Fluctuations (New York: Wiley) Mandelbrot BB 1977 Frcrctuls: Form, Chance and Dimension (San Francisco: W H Freeman) May E J P and Aniagyei H M 1974 Electron. Lett. 10 250-1 McWhorter A L 1955 PhD Thesis Massachussetts Institute of Technology, Cambridge, Mass. -1957 Semiconductor Surface Physics ed R H Kingston (Philadelphia: University of Pennsylvania Press) pp207-28 Merckel G , Bore1 J and Cupcea N Z 1972 ZEEE Truns. Electron Devices ED-19 681-90 Montagnon N B 1954 Wireless Engr. 31 255-63, 301-5 Nicolet MA, Bilger H R and Zijlstra R J J 1975 Phys. Stat. Solidi b 70 9-45, 415-38 Ortmans L H F and Vandamme L K J 1976 Appl. Phys. 9 147-151 Pai S Y 1978 PhD Thesis University of Minnesota Peyriguer J M, Martinez A and Esteve D 1973 C.R. Acad. Sci., Paris B 276 511-4 Sah C T and Hielsher F H 1966 Phys. Rev. Lett. 17 956-8 Simmons J G 1964 J. Appl. Phys. 35 2655-8 Stern F 1972 Plzys. Rev. B 5 4891-9 Storm L 1978Proc. 5th Int. Conf. on Noise in Physical Systems, Bad Nauheim, FRG ed D Wolf Springer Series in Electrophysics vol 2 (Berlin: Springer-Verlag) p240 Stroeken J T M and Kleinpenning T G M 1976 J. Appl. Phys. 47 4691-2

532

F N Hooge, T G M Kleinpenning and L K J Vandamme

SutcliRe H 1972 Colloq. Int. CNRS. Le bruit de fond des composants actifs semi-conducteurs, Toulouse, 1971. Conf. Rep. CNRS No 204 pp39-42 Sze S M 1969 Physics of Semiconductor Devices (New York: Wiley) p519 Vaes H M J and Kleinpenning T G M 1977 J. Appl. Phys. 48 5131-4 Van Bokhoven W M G 1978 Arch. Electron. Uebertragungstech.32 349-52 Vandamme L K J 1974a J. Appl. Phys. 45 4563-5 -1974b Phys. Lett. 49A 2 3 3 4 -1976a Appl. Phys. 11 89-96 __ 1976b PhD Thesis University of Technology, Eindhoven -1980a Solid-St. Electron. 23 317-23 -1980b Proc. 2nd Int. Symp. on llf Noise, Gainesville,USA (Florida: University of Florida) pp228-43 Vandamme L K J and Kedzia J 1979 Thin Solid Films 65 283-92 Vandamme L K J and de Kuijper A H 1979 Solid St. Electron. 22 981-6 Vandamme E K J and van Bokhoven W M G 1977 Appl. Phys. 14 205-15 Vandamme L K J and de Werd H M M 1980 Solid St. Electron. 23 325-9 Van der Ziel A 1959 Fluctuation Phenomena in Semiconductors (London: Butterworths) -1970 Noise Sources Characterization Measurement (Englewood Cliffs, NJ: Prentice-Hall) __ 1979 Adv. Electron. Electron Phys. 49 225-97 Van Helvoort G J M and Beck H G E 1977 Electron. Lett. 13 542-4 Van Vliet K M and Mehta H 1981 Phys. Stat. Solidi b in press Voss R F 1978 Phys. Rev. Lett. 40 913-6 Voss R F and Clarke J 1976a Phys. Rev. Lett. 36 4 2 4 -1976b Phys. Rev. B 13 556-73 Wall E L 1976 Solid St. Electron. 19 389-96 Wax N (ed) 1954 Selected Papers on Noise and Stochastic Processes (New York: Dover) Weissman M B 1980 Physicu B, C 100 157-62 Zettler R A and Cowley A M 1969 IEEE Trans. Electron Devices ED-16 58-63

Potrebbero piacerti anche