Home

Search

Collections

Journals

About

Contact us

My IOPscience

Deep-level transient spectroscopy (DLTS) analysis of defect levels in semiconductor alloys

This content has been downloaded from IOPscience. Please scroll down to see the full text. 1988 Semicond. Sci. Technol. 3 1177 (http://iopscience.iop.org/0268-1242/3/12/005) View the table of contents for this issue, or go to the journal homepage for more

Download details: IP Address: 79.123.192.18 This content was downloaded on 02/03/2014 at 19:10

Please note that terms and conditions apply.

Semicond. Sci. Technol. 3 (1988) 1177-1183. Printed in the UK

Amita Dast, Vijay A Singht and David V LangS t Department of Physics, IIT Kanpur, Kanpur 208016, UP, India
$

AT&T Bell Laboratories, Murray Hill, NJ 07974, US

Received 24 May 1988, accepted for publication 5 July 1988

Abstract. A deep trap in the fundamental gap of a semiconductor has a sharp

(delta function) character. In a semiconductor alloy the presence of disorder introduces a distribution of trap activation energies. We have undertaken a detailed analytical and numerical exercise to examine the effect of such broadening on the capacitance transient and on deep-level transient spectroscopy (DLTS) analysis of traps. In general, the standard DLTS analysis introduces negligible error except in cases of severe broadening where it overestimates the activation energy. We illustrate our analysis by considering both simulated and actual experimental situations.

defects. These resultcan in local deviations from the equilibrium concentration. A defect level in the fundamental gap of the semiconductor maybe identified with Deep-level transient spectroscopy (DLTS) [l] has been widelyused over the past ten years as a tool to study its corresponding defect potential via the Koster-Slater deep levels in semiconductors. The context of the equation. For the reasons discussed above, the defect potential in a semiconductor alloyhas fluctuations original proposal [ l ] was that of defect states with well defined energy levels in crystalline host materials. In which cause the delta-function-like gap level to be the past few years, however, there has been considerbroadened. able work on systems where either the host or the There exists a growing body of experimental data on gap levels in semiconductor alloys. Some of this defect is disordered. In such cases the DLTS spectrum may be broadened and the underlying capacitance or indicates the existence of potential fluctuations and current transient may be non-exponential. Neverconsequent broadening of the gap level. In the present work we shalldiscussthis broadened gap level in the theless, in most cases, thestandard DLTS expression context of DLTS [l]. Let us mention briedly experimenderived for exponential transients [ l ] is used to analyse tal work, other than DLTS, which indicates the existence the data. In this work we analyse the effect of gaussian of broadening. Mooney and co-workers [2] have stubroadening of a deep level energy spectrum on the DLTS signal. We show that in all cases except extreme died the kinetics of charge trapping in semiconductor DLTS analysis introduces alloys and hypothesise a distribution of energy levels broadening the standard capture negligible and error. cross sections. Raman spectroscopy, too, A semiconductor alloy, e.g. Gal-,AI,As, may be suggests the importance of alloy potential fluctuations [3]. Samuelson er a f [4] have attributed the splitting of considered to be a substitutionally disordered system. Ga,-,AI,As has a basic zincblende structure where the photoluminescence lines of Cu in GaAs,-,P, to alloy cation site is occupied by Gaor AI atoms. The occupotential fluctuations. Over the past few years there has been a growing pancy is essentially at random if the probability thata body of literature on DLTS studies of disordered given cation site is occupied by a Ga (AI) atom is 1"x ( x ) . However, fluctuations from the mean occusystems. Omling et a1 [5]have studied the broadened pancy are possible. These fluctuations arise duetoa DLTS spectra in GaAs,-,P,. Murawala [ 6 ] , Singh [7], variety of causes. Some of them are: (i) the Pauling and co-workers havediscussed the broadening of the E3-like level in GaAs,-,Sb,. Kanieswka and electronegativity difference between Ga and As is 0.4 Kanieswka [g] havesimilarly studied the main electron while the difference between AI and As is 0.5, thus trap in GaAs,-,P,r. Lang and co-workers [g] have stufavouring AI as a neighbour of As. (ii) Local inhomogeneities such as the presence of voids and anti-site died the gap states in amorphous hydrogenated silicon
1. Introduction
0268-1242/88/121177+07 $02.50

@ 1988 IOP Publishing Ltd

1177

A Das et al

Unless the level is very shallow, we have & S1 and this scale shall not concern us henceforth. If the potential fluctuations are small, then S and b in equation (1.4) will be small. We shall call the case {b < 1, s<O.l} the 0. l} will be weak disorder case. The regime {b2 1,S zdeemedthestrongdisordercase. Wepredict signifiC(t) = CO exp( - t h ) (1.lu) cantly different behaviour for these two cases. the case of weak disorder In 0 2 we treat l/z=ATexp(-EIkT). (l.lb) { b< 1, S < 0.1). We show that the capacitance transient Here E is the activation energy and A is related to the is weaklynon-exponential.Wevalidatethestandard capture cross section of the deep level under considerDLTS analysis based on equation (1.3). In Q 3, where we ation. T is the temperature, k the Boltzmann constant treat the strong disorder case {b3 1, sSO.~}, the tranthe of and z is referred to as the time constant. For case sientcapacitance is highly non-exponential,akinto a broadened defect level the most general expression glassy and spin-glass relaxation. Further, we demonhas been considered by Singh [7] DLTS analysis will systematically strate that the standard overestimate the activation energy in cases of severe B 4 we apply our experience of the broadening. In 2naq 0 C(r)=*I:dEImdA previoustwosections to two realistic cases: a-Si:H previously studied by Lang et a1 [9] and an E3-likelevel ( E - Eo) (A -A,) in GaAso.86Sb0.13 previously studied by Murawala et a1 xexp( - 2a2 - 20: -2). z(A, E ) [6]. Section 5 encapsulates our conclusions and reflections on related phenomena. Intheaboveexpressiontheeffect of potential fluctuations on both the activation energy and the capture cross section has been taken into account. The broadening assumed is gaussian with a mean at E, and A, 2. Analysis for weak disorder and thehalf width at half maximum (HWHM) of approxiWe first consider the weak disorder case {b = a/kT< 1, mately a and al respectively. A detailed examination As a+O the distribution approaches a u/Eo<O.l}. [7] has indicated that the broadening in the quantity Eo). We are then justified in delta function d(Erelated to the capture cross section may be ignored. We expanding the exponential term in the integrand of then obtain equation (1.2) C(t)

(a-Si:H) where the disorder-induced broadening is expected to be more pronounced than in semiconductor alloys. Under idealconditions,thecapacitancetransient C ( t )of deep level is a simple exponential [ l , 101

=A 6 I Y d E exp( - ( 2a L ) . z(E)

(1.2)

exp( - E/kT) = exp( - E,/kT) exp[ - ( E - E,)/kT]

= exp( - E,/kT)

This expression has been widely employed in recent years [5-81. In thiswork we first demonstrate how expression (1.2) implies long-lived a and nonexponential capacitance transient. We demonstrate this by independent analytical and numerical work. Next, and more importantly, we show under what conditions the standard interpretation of DLTS data is valid for a broadened gap level.At this juncture we will define the term standard interpretation. The time constant for the apparent peak of a DLTS signal was shown by Lang to be [l] 1ln(t2/t,) -~ t ,

X[l-(E-EO)/kT+(E-E,)/2(kT)]. Defining

l/zo= A T 2 exp( - Eo/kT)

(2.1)

C ( t )= C, exp( - t / t o )
equation (1.2) becomes

(2.2)

c@) =a t )

IoGa
P

dE

tr - tl

(1.3)

-P[

-2(

1 (E-Eo) t(E-Eo)a2 2 to +. .

.)l.

This expression contains no information on the broadening or smearing factor a. The use of this expression for extracting the defect parameters (activation energy and the pre-exponent A) will be deemed the standard interpretation. We identify three natural scales in the problem:
E

We now apply the stationary point formula to the above integral. Given

Z(a)=

exp( - a p ( E ) )d E

the formula [l11 states that for large a

= ( 1E .4 o~ /k )T

~(a =)fiexp( - ap(~,))/I- aq(E,,)Il

where E, represents the minima of q ( E ) . Application of this prescription yields

(1.4b) S = a/Eo b (1 = .4 a/ ~k )T .
1178

DLTS

analysis of defect levels in alloys

expression (2.5) haslittleinfluence onthedetermination of the activation energy. For weak disorder the standard DLTS analysis based on equation (1.3) would yield accurate values of the activation energy and the pre-exponential factor. Thus we conclude that the standard analysis is robust with respect to small disorder. We have indeedverified this for several values of the parameters b and alE,.
3. Analysis for strong disorder

0.2

0.4

0.6

0.8 1.2 1.0

t is)
Figure 1. Plot of the capacitance transient for small ulkT (-0.2)and r-l=3.4418S-. Note the slower decay of the broadened gap level as opposed to the , exact exponentially dictated form. T = 180 K. (equation (1.2)); + + , approximate (equation (2.3)); , exponential (equation (2.2)).

We now consider the case of strong disorder { b = a l k T > l , olEo>O.l}.Definingx=(E-E,)la, and using the definitions (2.1) and (1.4), thecapacitance transient of equation (1.2) may be cast in the form

C([) = C
where

1%
- Eoio

exp( - f ( x ) ) &l&

(3.1)

f ( x ) =2

+ t exp(t o- bx)

(3.2)

C ( t )= C,exp(

--+ zo

O.5(atlkTt0) 1 + (a/kT)2t/to

[l (a/kT)t/t,].

(2.3) Equation (2.3) is out modified formula for the case of weak disorder. Note that for the case d+O we regain the exponential expression (2.2) for the ideal transient. For a DLTS analysis we construct the signal

On account of the nature of the gaussian, the upper limit E, in equation (3.1)has been replaced by infinity. This entails only negligible error. We nextapply the stationarypointformula in amanner similar tothe previous section. Differentiating

f ( x )= x
460 420

- bt

exp( - bx)/to.

S ( T ) = (C(tJ - c(t>)/co

(2.4)

and differentiate it with respect to the temperature to obtain (see the appendix) the equivalent of expression (1.3)
.m _
L

380 340
-

; ;300 m

8 260 -

The subscriptm denotesthemaximum DLTS signal. DLTS Expression (2.5) reduces smoothly to the standard expression (1.3) as alkT,,,+O. The implication is that the standard DLTS analysis may be in slight error and yield an incorrect activation energy. Figure 1 depicts the capacitance transient as calculated using the exact expression (1.2), our approximate equation (2.3) and the pure exponential equation (1.1). The three curves are nearly identical, though our approximate expression (2.3) and the exact (1.2)are in closer agreement. Further, the exponential expression decays faster than the other two indicating the slowly decaying (non-exponential) characterof the broadened gap level. Figure 2 depictsthe DLTS plotfor thegap level discussed in figure 1. The maxima in the signals are at identical temperatures for the three cases discussed in the previous paragraph. This implies that the modified

S 220 vl 180h

140 100 -

60 20

l
140

/
& f

l
T IKI

160 120

l 180

\
200

Figure 2. Simulated DLTS signal for the parameters of figure 1 and rate window t2/tl= 4. The differences between the various expressionsare very small, thus vindicating the standard DLTS analysis based on , exact (equation (1.2)): + +, equation (1.3.). approximate (equation (2.3)); - - -, exponential (equation (2.2)).

1179

A Das et al

0.6

0.5

0.4

L2

b = a/kT is large ( ~ 6 ) The . transient is long-lived and distinctly non-exponential.Theapproximateexpression (3.3) agrees moderately well with the exact one. We have been unable to obtain a closed-form expression for the DLTS time constant extrema analogous to equation (1.3) or equation (2.5). Thedifficulty is due to the complicated character of equation (3.3). Hence, for further analysis of the strong disorder problem, we have to rely on the exact expression (1.2). The DLTS extremacan belocated by differentiating equation (1.2)
=

at

d E exp[ - ( E - Eo)/2d]

x [exp( - t l / t ) / t 2 - exp( - t 2 / t ) / z = ] O (3.5)

0.:

0. i
1

0.2

0.L
f

0.6

0.8

1.0

Is1

Figure 3. Plot of the capacitance transient for Eo = 0.85 eV, A = 4 x IO8 S K-2. The broadening factor U = 0.175 eV and T=340 K, leading to ulkT=6. The transient is long-lived and distinctly non-exponential, The approximate expression (3.3) is in reasonable agreement with the exact one. The parameters selected for this plot are similar to ones encountered in a-Si:H [g]. - -, exact (equation (1.2)); + + +, approximate (equation (3.3));, exponential (equation (2.2)).

= 0 is given by the For large b the solution of f(x,,) recursive expression

x=- In

1 b

tblt,, tblt,, In ln(tb2/z,,. . . .)

and employing a non-linear numerical procedure such asNewton-Raphson to extractthethreeparameters E,,, A and a. A question of central importance is the error made if thestandard analysis based on equation (1.3) is applied to the DLTS data for the strong disorder case. The analysis of the previous section indicates that the error made is almostneglible forthe weakdisorder case. We consider a typical deep level centred at E,,= 0.5 eV, andA = lo9S - K-, and characterised by anincreasing broadening U = 0.1,0.25and 0.4 eV, respectively. Using a rate window t2/tl= 4, DLTS signals employing the exact broadened expression (1.2) were generated. From the maxima of these signals, T,, we construct Arrhenius plots using the standard expression (1.3). Figures 4-6 depict Arrhenius plots for the three distinct broadenings considered (a=0.1, 0.25 and 0.4 eV respectively). An examination of figures 4-6 leads us to conclude(i)there is no deviationfrom linearity over the four time decades considered in the DLTS simulations (0.1 ms to 1 S ) . Hence an examination

Theaboveexpressionholdsunderthecondition tb/z,,> e. We shall use a truncated form of the above expression, i.e. stop the recursion after the first denominator. A series of algebraic manipulations then yields

50

55
l l k T , lev)

60

Equation (3.3) is a radical departure from the Debye relaxation (pure exponential) form. If the first term in the exponent is ignored, it reduces to (3.4) C ( t ) llt where a = l/b?, a form of non-exponential relaxation currently popular in spin-glass and related areas. Figure 3 depicts a typical capacitance transient encountered in a DLTS analysis of a-Si:H [9]. Here 1180

Figure 4. The Arrhenius plot for a simulated DLTS signal with rate window t2/t,=4. A typical broadened deep level was considered and the parameters employed for the simulation of equation (1.2) were Eo = 0.5 eV, A = IO9 S K-, and u=0.1 eV. The Arrhenius plot is linear over four time decades (0.1 m s to 1 S ) . Note the difference between the extracted parameters (A, A ) and the input (E . , A ) . ;,=0.515 eV, A = 1 . 3 IO9 ~
S-1

K-2

DLTS

analysis of defect levels in alloys

45

llkT, ( e V 1 )

Figure 5. The Arrhenius plot for a typical broadened deep level with broadening U = 0.25 eV. The plot is linear over four time decades (0.1 ms to 1 S). Note the difference between the extracted activation energy E;, = 0.607 eV, and Eo= 0.5 eV. A = 2.5 x IO* S- K. Other parameters employed are the same as figure 4.

0.2

0.4 olE,

0.6

0.8

of the Arrhenius plots provides little clue to the broadening of the given gap level. (ii) The extracted gap level Eh is invariably greater thanEo,the actual peakin broadened density of states. In fact as figure 6 depicts, for U = 0.4 eV, E ; = 0.72 eV, while Eo= 0.5 eV, a 50% deviation! (iii) The extracted value of the pre-exponent A is also larger than A . However no systematic trends were found for A . Figure 7 displays the extracted value E ; as a function of ulE,,. It reinforces and encapsulates the conclusion based on figures 4-6. For ulEo>O.l there is a systematic overestimation of the activation energy. For ulE,<O. 1, this error is negligible.
4. Simulation on real systems

Figure 7. Plot of the extracted activation energy h against the dimensionless broadeninga/&. The parameters used are the sameas figures 4-6. There is a systematic overestimation of the activation energy for u/E0<0.05. The inset displays Et, in the weak disorder limit (u/Eo>O.l).

We shall apply our analysis to two experimental situations: (i) a-Si:H; (ii) an E3-like level in GaAso.s6Sbo.14. For the a-Si:H case [9] the disorder is large and a priori we expect U to belarge.Indeed,as figure 8 shows, the fit totheexperimental DLTS signal [9] is accomplished by using U = 0.175 eV, Eo= 0.82 eV and A =4 X 1 0 sK - 2 in equation (1.2). This implies that ulkT,,,=6 and alE0=0.2. This fit was possibleonly

after some trial and error. The corresponding Arrhenius plot is depicted in figure 9. Note that it is linear over five time decades. The value of the activation energy extracted from the slope of the data presented in figure 9 is EA= 0.87 eV. Recall that the fit in figure 8 used the input value of Eo= 0.82 eV. Thus the overestimates the activation standard DLTS analysis energy in this case by 0.05 eV. Lang er a1 [9] have done numerical analysis for the case of a-Si:H, taking into account both large trap densities and broadened energy spectra, and find similar results. Thus, our main thesis encapsulated in figure 7 that the standardDLTS analysis overestimates the activation energy in the limit of severe broadening, is borne out. The nextcase we consider is the E3-like level in GaAso,,,Sb0,,,. This has been subjected to a DLTS study by Murawalaandco-workers[6].A close fit tothe experimental DLTS plot is obtained by simulating equation (1.2) with Eo= 0.23 eV, A = 2.31 X lo4sK-*, and U = 0.055 eV. This corresponds to ulkT, = 4 and ulEo= 0.24. We have not reproduced the DLTS signal fit here since it is analogous to figure 8. However, we do show the Arrhenius plot in figure 10. The plot is again linear

J
1

40

45

llkT, (eV1

Figure 6. The Arrhenius plot for a typical broadened deep level with broadening a=0.4 eV. The plot is linear over four time decades (0.1 ms to 1 S). Note the difference between the extracted activation energy EA= 0.72 eV, and Eo= 0.5 eV. A = 2.4 x IO9 S K-2. Other parameters employed are the sameas figure 4.

100

150

200

250
T

300 450 400 350

IK)

Figure 8. Comparison of the experimental DLTS signal [g] with the simulated one for a-Si:H. The parameters , are identical to figure 3. The rate window f2/tl = 4. experiment; - - -, simulated.
1181

A Das et a/

-6

"l
.
c I

tt

i
4

"r

-10

-11

1
28

suspect systems. Second, an examination of the capacitance transient is recommended. If it is nonexponentialthenone must presumebroadening.In fact, it was this trait which led earlier workers to posit we have demonstrated in broadening [5-71. What thiswork is an analyticproof of a long-lived nonexponential character of the capacitance transient both forweakdisorder (equation(2.3)and figure 1) and strong disorder (equations (3.3) and (3.4) and figure 3 ) . Third,the DLTS signal is itself broadenedoverand above its normal value. Murawala er a1 [6] have proposed phenomenological a relationship connecting ATIT,, the full width at half maximum to a AT/Tm=O.l+a/E,. This relation may help one to extract a rough value of a.Lastly, the Arrenhius plot provides little clue to the existence of broadening. As figures 4-6, 9 and 10 indicate, the Arrhenius plot is linear over four to five time decades, even in the presence of significant broadening. Regarding the second question, the following conclusions may be gleaned from our work. (i) The conventional DLTS analysis overestimates the activationenergyforseverely broadened levels. This overestimation is often quite small. It may, however, lead to small but systematic discrepancies between DLTS and other experimental probes such as optical studies, Hall effect etc. In the case of a-Si:H studied in Q 4, the input value was E,,= 0.82 eV while the extracted value was EA=0.87 eV. The precise factor responsible for this discrepancy can be traced to the expression for the time constant

30

32 IlkT, IeV"1

34

36

Figure 9. The Arrhenius plot for a-Si:H. The parameters are the same as in figures 3 and 8.The plot is linear over five time decades (0.01 ms to 1 S). There is an overestimation of the extracted activation energy: h= 0.87eV while Eo= 0.82 eV. A' = 4.4x IO8 S - ' K-2.

over four time decades. The extracted activation energy is 0.25 eV, in slight excess of the peak in the broadened gap level (E,, = 0.23 eV).
5. Conclusion

In this section we first address ourselves to two issues: (i) What are the experimental signatures in capacitance transientspectroscopy of abroadenedgap level as opposed to a sharp delta-function-like gap level? (ii) To what extent would a standardDLTS analysis of a broadened defect level be in error? Regarding the first question, there are at least three indicators of broadened gap level. First, if the system under study is patently inhomogeneous, then the qualitative arguments presentedin 9 1 indicate that the level should be broadened. Semiconductor alloys are typical suspect systems. An extreme example is GaAsSb since it is a semiconductor alloy with a large lattice constant mismatch between GaAs and GaSb. Amorphoussemiconductorssuchasa-Si:Hwouldbeanotherset of

F(alkTm,alE,,) < 1.

(5.2)

Figure I O . The Arrhenius plot for an E3-like level in GaAs,-Sb, (x=0.14). The plot is linear over four time decades (0.1ms to 1 S). There is a negligible difference between the extracted activation energyE(,= 0.25 eV and Eo= 0.23 eV. A' = 3.4 x IO4 S - ' K-'.

The functional form of F for small disorder is given by equation(2.5).For large disorder, the complicated form of equation (3.3) prohibits asimple derivation. F, nevertheless, is always less than unity. This character of F is reponsible for the overestimation. (ii) The error involved in a standard DLTS analysis based on equation (1.3) is minor if ( a ) alkT,< 1 and alE,<0.1, ( b ) olkT,>l but alE,<O.l, (c) thepreexponent A < lo6s-' K-' . Alternatively stated, our experience shows that the error involved is significant (>0.05eV) only if alE0>0.1 and A>10"s"K-'. Thus, standard DLTS analysis is fairly robust with respect to disorder. We would like to end with a few speculative remarks. The power law decay form in equation (3.4) was derived under the assumption of additive contributions embodied in equation (1.2). Thisimplies that the carriersareemittedindependentlyintotherelevant band. Suchaprocess is deemed parallelrelaxation. Mutual correlations and subsequent modification of the defect density of states during the emission process are ignored. Preliminary investigations in the relaxation of

1182

DLTS

analysis of defect levels in alloys

glassy phenomena indicate that non-exponential relaxation arises as a natural consequence of serial as opposed to parallelrelaxation [12]. Similarconsiderations in the study of capacitance transients may yield interesting results. The form of the defect density of states has been assumed to be gaussian (equation (1.3)). This form is premised onthe central limit of theorem. A microscopic theory for the defect density states in semiconductor alloys should be pursued [13]. Finally, detailed analytical and numerical exercise similar to the current one must be carried out for other experimentally studied quantities of interest (e.g. Hall effect, thermally stimulated capacitance etc) to ensure thattheconventionalinterpretation is valid forthe disorder-broadened case.
Appendix

We obtain next theDLTS extrema by differentiating (2.4) and using equations ( A . l ) and (A.4) forthe signal. This expression is trivially equation (2.5).

References

We will indicate the derivation of the modified expression (2.5). Differentiating equation (2.3)
-

DLTS

1aqt) - exp[ - t / t , 0.5(at/kTto)] T aT

{[tlto- (~t/kTt,)/to] &o/aT- (ot/kTto)/T}. ( A . l )

If

Iat,/atl S tom

(A4

then the last term on the RHS of ( A . l ) may be neglected. This condition translates to

2 + EoIkTS 1 .

(A.3)

Such a condition is easily met. Next we exponentiate the termin the curly brackets of (A.l) and expand the subsequent logarithm:
t h o- (ar/kTto)=t exp[ln(l - ( c ~ / k T ) ~ t / t ~ ] / t ~
2

t exp[ - (a/kT)2t/to]/zo.

(A.4)

[l] Lang D V 1974 J. Appl. Phys. 45 3014-23 [ 2 ] Mooney P M, Caswell N S, Solomon P M and Wright S L 1986 Materials Research Society Meeting, San Francisco 15-18 April 1985 Theis T N, Kuech T F, Palmateer L T and Mooney P M 1985 Gallium Arsenide and Related Compounds 1984 (Inst. Phys. Conf. Ser. 74) Ch. 4, p241 [3] Parayanthal P and Pollak F H 1984 Phys. Rev. Lett. 52 1822 [4] Samuelson L, Nilsson S and Grimmeiss H-G 1984 Phys. Rev. Lett. 53 1501 [ 5 ] Omling P, Samuelson L and Grimmeiss H-G 1983 J . Appf. Phys. 54 5117 [6] Murawala P A , Singh Vijay A , Subramanian S, Chandvankar S S and Arora B M 1984 Phys. Rev. B 29 4807 Proc. 13th Int. Conf. on Defects in Semiconductors 1985 ed. L C Kimerling and J M Parsey Jr (New York: AIME) p 1165 [7] Singh V A unpublished [8] Kanieswka M and Kanieswka J 1985 Solid State Commun. 53 485 [9] Lang D V, Cohen J D and Harbison J P 1982 Phys. Rev. B 25 109 Cohen J D, Lang D V, Harbison J P and Seregent A M 1983 Sofar Cells 19 119 [lo] Bleicher M and Lange E 1973 Solid State Electron. 16 375 [l11 Bender C M and Orszag S A (ed.) 1978 Advanced Mathematical Methods for Scientists and Engineers (New York: McGraw-Hill) p 267 [l21 Palmer R J , Stein D L, Abrahams E and Anderson P W 1984 Phys. Rev. Lett. 53 958 [l31 Bhat G R , Singh V A and Rangawala A A 1988 Phys. Rev. B (submitted) Bhat G R 1986 PhD Thesis Bombay University

1183