Xray diffraction and imaging
Covariant description of Xray diffraction from anisotropically relaxed epitaxial structures
^{a}Belarusian State University, Minsk, Belarus, ^{b}Xray Research Laboratory, Rigaku Corporation, Akishima, Tokyo, Japan, and ^{c}Rigaku Europe SE, Ettlingen, Germany
^{*}Correspondence email: alexei.zhylik@atomicus.by
A general theoretical approach to the description of epitaxial layers with essentially different cell parameters and inplane relaxation anisotropy has been developed. A covariant description of relaxation in such structures has been introduced. An iteration method for evaluation of these parameters on the basis of the diffraction data set has been worked out together with error analysis and reliability checking. The validity of the presented theoretical approaches has been proved with aZnO on rsapphire samples grown in the temperature range from 573 K up to 1073 K. A covariant description of relaxation anisotropy for these samples has been estimated with data measured for different directions of the diffraction plane relative to the sample surface.
Keywords: epitaxial layers; relaxation parameters; highresolution Xray diffraction; theoretical approach.
1. Introduction
Highresolution Xray diffraction (HRXRD) is one of the most effective tools for investigation and nondestructive control of epitaxial crystalline thinfilm layers (Pietsch et al., 2004). An important part of HRXRD analysis is defined by construction of the appropriate sample model and connection of its parameters with diffraction pattern features. It is well known that the interaction between atoms in adjacent thin films leads to deformation of the crystal unit cells that depends on the layer thickness. In the case of layers having a cubic crystallographic system and being isotropic, inplane characterization of this deformation is a well established procedure with the relaxation R as the only parameter connected with inplane strain of the (mismatch). However, nowadays epitaxial layers with hexagonal materials in different orientations are widely used in industrial lightemitting diode production (Paskova, 2008). In this investigation, a complicated epitaxial relation in combination with an inplane anisotropy appear (Laskar et al., 2011). Consequently more parameters should be used for the accurate characterization of unitcell deformations. In the present paper, a set of parameters allowing one to describe layers with anisotropic epitaxial relations is introduced and a general procedure for evaluation of these parameters is developed, based on the Xray diffraction profiles or maps (RSM). We consider this a covariant description because it can be used in the same form for arbitrary symmetry and orientation of the crystal unit cells in each layer. A series of papers (Yang et al., 1994; Caro & Tapfer, 1995, and references therein; Caro et al., 1996; Bottomley et al., 2001; Brandt et al., 2002) have been devoted to theoretical analysis of similar problems. However, in these works the fully pseudomorphic layers were considered and relaxation anisotropy was not taken into account.
We are proposing a generalized theoretical approach to describe the relationships between multiple crystalline layers that may display relative mismatches and inplane anisotropy. The validity of XRD data evaluation was verified on the basis of an iteration scheme for error analysis and reliability checking. Diffraction data were collected in three different crystallographic directions from aZnO on rsapphire samples grown at different temperatures from 573 to 1073 K. The analysis was performed according to the approach described above.
The paper is organized as follows. In Theoretical approach, we show how the conventional parameters used for the description of the epitaxial layer state for cubic inplane isotropic layers can be generalized for inplane anisotropic layers with arbitrary epitaxial relations. In Fit algorithm and error analysis, we describe a way to get values and errors of sample parameters from Bragg peak positions obtained from the set of diffraction profiles/RSM to describe layers with arbitrary epitaxial relations. Application to aZnO films on rsapphire substrates, grown at different temperatures from 573 to 1073 K, is presented in Experimental: aZnO on rsapphire.
2. Theoretical approach
Here we provide a reminder of the principal definitions of the conventional approach for the cubic crystallographic system (Fig. 1a) and isotropic inplane relaxation. The typical crystallographic representation of the interface between two epitaxial layers is displayed in Fig. 1(b). Using the lattice constant values as a basis for further definitions, the lattice constant mismatch F of the relaxed top layer or initial mismatch is defined as follows:
where a^{L,R} is the lattice constant for the relaxed layer. In this paper, we will consider only the strain of the pseudomorphic layer so the structure of the relaxed layer and the parameter F are supposed as known values. It is also assumed that the other inplane lattice constant b^{L,R} is defined utilizing the same equation (isotropic relaxation).
The difference between the relaxed and the actual lattice constants is characterized by the relaxation degree R,
This is closely connected with the actual mismatch of the top layer, which is usually defined with the diffraction data,
and its strain
The outofplane lattice constant c can be found on the basis of Hook's law, which is expressed as the Poisson ratio for the considered isotropic case:
In order to generalize equations (1)–(5) for anisotropic relaxation (twodimensional) we should introduce a `sample' orthogonal coordinate system S: S_{1}, S_{2}, S_{3} with S_{3} parallel to the sample normal and S_{1}, S_{2} describing the interface plane. There is some degree of freedom for selection of S_{1,2} that will be used below.
Let us assume that the pseudomorphic state interface between layer and substrate has an inplane translation symmetry, forming a coincidencesite lattice (CSL) (Zur & McGill, 1984), which is in general different from those of layer and substrate. The actual inplane translation symmetry of the CSL can be parameterized by defining three nodes in the layer and in the substrate which coincide. We will call them anchors. The anchors are given in in crystallographic reference systems for the layer C_{L} and for the substrate C_{S}; hence they enable the description of the mutual arrangement of C_{L}, C_{S} and S. One of the anchors is by definition [000] in both C_{L} and C_{S}. The other two anchors are in general given by the sum of CSL translation vectors with integer factors.
For example, the trivial case of relaxed Ge on Si in terms of anchors can be described in the following way:
or by
both leading to the same CSL. Compared with the epitaxial relations, anchors define not only the directions of CSL translation vectors but their magnitude as well.
In order to proceed with the generalization of equations (1)–(5) let us introduce the following quantities:
for the relaxed top layer anchor projections on the sample axis,
for the actual top layer anchor projections on the sample axis, and
for the bottom layer (substrate) anchor projections on the sample axis; the anchor projections on S_{3} are zero by definition.
Then the mismatch between the relaxed layer and the substrate is described by the known 2 × 2 matrix F_{ij} [compare with equation (1)]:
The matrix F_{ij} describes the strain only; the net rotation is supposed to be excluded from the transformation (10), therefore F_{ij} is a symmetrical matrix. The vectors S_{1}, S_{2} can be chosen as the principal axes Φ_{1}, Φ_{2} of F_{ij}. In this inplane coordinate system, F_{ij} has a diagonal shape and can be described by two eigenvalues F_{1}, F_{2}:
The actual mismatch matrix connects the anchor projections in the following way [compare with equation (3)]:
The relaxation matrix R_{ik} connects the actual mismatch and the initial mismatch F_{kj} as follows:
In this coordinate system , the relaxation tensor will follow the strain tensor and have a diagonal form. It is described by two values, R_{1}, R_{2}, that can be considered as the covariant relaxation parameters. Consequently the actual mismatch tensor can be expressed in the following way:
The inplane strain tensor of the top layer connects the projections of the actual top layer and the relaxed top layer:
The strain can also be expressed through the relaxation and the initial mismatch [compare with equation (4)]:
In order to reconstruct a complete deformation status of the crystallographic
all the components of the threedimensional strain tensor have to be calculated. Hook's law, linking the strains and the stresses by the stiffness tensor, is used:Taking into account that the inplane strain tensor components are known, the strain and the stress tensors are symmetric, and the vertical components of the stress tensor equal zero because only inplane forces σ_{13} = σ_{23} = σ_{33} = 0 exist, the values can be found from equation (17). As the result, the complete strain tensor is obtained:
where the components are defined by relaxation parameters (16) and are calculated from equation (2).
Knowing the threedimensional strain, it is straightforward to find the parameters of the strained crystallographic cell. The connection of the crystallographic cell basis e_{1}, e_{2}, e_{3} with the cell parameters is
The decomposition of the relaxed crystallographic cell basis vectors in the principal axes basis of the initial mismatch is
Hence, the decomposition of the strained crystallographic cell basis vectors in the principal axes basis of the initial mismatch is given by
Finally, the parameters of the strained crystallographic cell are
where in equation (22) summing over the index i is not performed.
The proposed method is valid for every crystallographic system. It can be shown that the expressions for strain and unitcell deformation, as obtained from Romanov et al. (2006) and Laskar et al. (2011) for specific cell configurations, can be derived from the equations above.
3. Fit algorithm and error analysis
Once the parameters from equation (20) are known, the direct diffraction problem can be solved, i.e. the positions of Bragg peaks in diffraction profiles and/or RSM can be calculated for an arbitrary sample. However, in most cases the solution of the inverse problem, that is the evaluation of the sample parameters (like concentration, relaxations R_{1}, R_{2} etc.) from a series of Bragg peak positions, is required. In the onedimensional case this problem can be solved analytically. In the considered twodimensional case, a quite large system of linear equations has to be solved. Following from equations (7)–(23) this should be carried out for each set of available diffraction data, e.g. symmetric and asymmetric profiles, or RSM, etc. Direct solution of these equations is not effective because the experimental data could be limited or, in the opposite case, the system of equations could be overcomplete. Therefore we have evaluated these parameters by fitting all experimental Bragg peak positions that could be extracted from the diffraction profiles and RSM in a concrete experiment. The proximity of the measured Bragg peak position to the theoretical one has been used as a partial cost function which should be equal to zero for exact values of the sample parameters.
The actual form of the cost function depends on the experimental setup. If an open detector is used, the Bragg peak position in a diffraction profile corresponds to the direction of the incoming wavevector k_{in} which satisfies the Bragg condition H^{(S)} = k_{out} − k_{in}, while the direction of the outgoing wavevector k_{out} cannot be obtained owing to the large detector aperture. In this case, we have used the following cost function:
Here H^{(S)} is the vector calculated on the basis of sample parameters of interest from equations (7)–(23).
In the case of RSM, both k_{in} and k_{out} can be found from the scattering experiment and the cost function is defined as
where Q = k_{out} − k_{in} is the transferred wavevector.
The cost function from a single peak provides only some link between parameters, e.g. the peak position in a symmetric scan can give only the connection between relaxation and concentration. To determine each of them an additional asymmetric scan is needed. For an arbitrary set of diffraction data we have built up the complete cost function . Then the parameters were evaluated by means of minimization of this cost function.
In order to find out whether the parameters of interest can be obtained from a given set of diffraction data, a procedure for evaluation of their errors has been developed. Generally, some information for the estimation of the errors can be gained from the analysis of the optimization algorithm convergence. In the case of fitting overcomplete sets of measured data this kind of error estimation can be sufficiently accurate, because of the random character of the experimental errors of different points in the considered data set (Giacovazzo et al., 2002). However, when the number of the characteristic parameters of the measured data has the same order as the number of optimized sample parameters, the error estimation on the basis of the cost function can be incorrect. For example, when the numbers of the data and sample parameters coincide, the optimization can be made perfectly with the achieved value of the cost function being exactly equal to zero. However, the obtained results may still contain errors; they are just undetected by this method. This situation, commonly encountered in the problem of cell parameter determination from Bragg peak positions in set of diffraction profiles, is analyzed below.
Consider the set of normalized sample parameters {x_{j}}, (like concentration, relaxation etc.), which are determined on the basis of fitting the set of parameters of the measured data (like Bragg peak positions) that are defined with some experimental errors assumed to be independent. The correct values {x_{j}} of the sample parameters correspond to the accurate values y_{β}, containing no experimental errors, and are determined by minimizing the cost function :
The values determined from the real (inaccurate) data are connected in the same way to the measured values :
Expanding equation (27) to the first order of the errors and taking into account equation (26), we obtain the following equation for the inaccuracies :
where
is the cost function Hessian for the sample parameters and
is the matrix of the mixed derivatives.
If the Hessian H_{ij} is a nonsingular matrix, equation (28) provides all the necessary information for estimation of the errors :
The deviations of the measured data from the exact values are unknown (otherwise the exact value could be found) and random in nature. From the experimental conditions and the methods of data processing, only the variances of the quantities can be estimated. Taking into account the independence of the random variables, the following expression for the mean squared errors of the estimated sample parameters is found:
A more general consideration, accounting for both singular and nonsingular Hessians, is provided in Appendix A.
4. Experimental: aZnO on rsapphire
A set of aZnO films were grown on rAl_{2}O_{2} substrates. ZnO films were grown in the temperature range of 573–1073 K with an interval of 100 K. The ZnO film thickness was about 280 nm. Structural characterization for the aZnO/rAl_{2}O_{3} nonpolar heterostructures was performed with a Rigaku fivecircle SmartLab system, equipped with a highpower Xray source (45 kV, 200 mA) and a Bartelstype Ge(220) monochromator. For each sample, HRXRD measurements were performed at 110, and 004 reflections (see Fig. 3a). The mutual orientation of the aZnO crystallographic cell on rsapphire is presented in Fig. 2(a). According to epitaxial relations referred to in the paper by Han et al. (2012), pairs of anchors were specified as {ZnO[003]; Al_{2}O_{3}} and {ZnO; Al_{2}O_{3}[110]} (see Fig. 2b). Relaxation parameters were found using the fitting algorithm with errors that are presented in Table 1. Strain and mismatch parameters were calculated from the relaxation. Relaxation, strain and mismatch results are shown in Fig. 3(b). Crystallographic lattice constants, including angle γ distortion of the ZnO layer, are calculated according to the proposed theoretical approach and the results are shown in Fig. 3(c).

5. Conclusion
A general formalism for describing anisotropic relaxation in layers with arbitrary epitaxial relations has been proposed. An iteration scheme of layer parameter determination from peaks of arbitrary sets of one and twodimensional diffraction data was considered. It is based on construction of a cost function and its minimization over layer parameters. Error evaluation and reliability checking of the obtained results has been described. Application of the proposed approaches has been demonstrated on Xray diffraction data of aZnO on rsapphire samples grown in the temperature range from 573 K up to 1073 K.
APPENDIX A
Error analysis. General case
A1. Equations
The number of measured parameters can be less than the number of sample parameters that a researcher would like to determine. Although the desired complete determination of the sample parameters is impossible in this case, some useful information can be retrieved from the experimental data. Below we show that, in general, the sample parameters can be divided into the following groups: (i) determined, (ii) partially determined and (iii) undetermined. The first type of parameters can be found from the provided experimental data with reasonable inaccuracies. The correct value of each parameter from the second group, considered separately, cannot be determined sufficiently accurately. However, there are some constraints on the set of `partially determined' parameters imposed by the measurement results. This means that, fixing the values of some of these parameters on the basis of some other information (e.g. information concerning preparation of the sample), one can find the correct values of the other partially determined parameters. The `undetermined' parameters are those for which the measured data provide no useful information.
If all the parameters are determined, equation (32) provides the estimated values of the parameters' errors. However, if there are some `partially determined' or `undetermined' sample parameters, the Hessian is singular and the matrix H^{1} is not correctly defined. To resolve this problem one can consider an eigenbasis of the Hessian matrix and exclude eigenvectors corresponding to zero (or too small) eigenvalues, from determination of the errors. More precisely, the errors corresponding to such eigenvectors must be estimated from some conditions different from those of equation (28). Below it is shown that these conditions are provided by normalization of the parameters x_{j}.
We slightly modify equation (27) by introducing the cost function error , independent of errors of the measured parameters y_{β}:
This error accounts for calculation inaccuracies and is introduced to make the error evaluation procedure more stable.
Let the orthogonal matrix describe the linear transformation of the sample parameters, diagonalizing the Hessian:
The solution provided by equation (31) is then simplified:
where
characterizes the influence of the finite calculation precision on the derivatives of the cost function.
Equation (35) is valid for nonzero eigenvalues only. Normalization of the parameters {x_{j}} implies that . Then one can assume that the following condition is satisfied:
Finally, from equations (35) and (37) one can find the following estimation for the mean squared error of the transformed sample parameters :
where
Then, the errors of the initial sample parameters can be calculated as
To complete the description of the error estimation protocol, one needs to assign a certain value to . This quantity was introduced to resolve the ambiguity arising when the cost function does not depend on some sample parameter (e.g. x_{n}). Such a parameter corresponds to the zero eigenvalue of the Hessian: λ_{n′} = 0, x_{n′} = x_{n}. On the other hand, one also has Z_{n} = 0 and, finally, , which means that introduction of nonzero quantity enables us to detect undetermined parameters by large estimated errors (otherwise, the estimated error is indeterminate and can happen to have a small value). For correctly determined parameters, the introduction of should not change the error estimation significantly: , where N is the number of the found sample parameters. Then, for all j′ the quantities can be assigned the following value:
where in calculations one can take (the value must be much less than the anticipated relative errors of the parameters but larger than the calculation errors caused by finite machine precision).
A2. Parameters classification
The parameters for which equation (40) provides reasonably small errors (Δx_{j} < η ≃ 0.5) are classified as determined from the measurement results. The parameters that do not satisfy this condition can be either partially determined or undetermined. To distinguish between these two groups of sample parameters, we evaluate the constrained errors.
The abovedescribed (unconstrained) error estimation was based on the assumption that all the sample parameters are initially unknown and can take any values (consistent with normalization). One can also make a constrained error estimation for each of the parameters, assuming that the considered parameter is the only unknown quantity, while the values of all other parameters are fixed. For a parameter with a large unconstrained error Δx_{j} > η, the constrained error can turn out to be small: Δ_{c}x_{j} < η. In this case the parameter is found to be partially determined. If both the unconstrained and the constrained errors are large, the parameter is undetermined.
Similarly to the abovedescribed calculations, the constrained error can be determined in the following way:
Summarizing, the procedure of error evaluation and parameter classification can be divided into the following steps:
(a) the unknown sample parameters are normalized, , and the cost function is constructed;
(b) the optimal values {x_{j}}, minimizing the cost function, are found;
(c) the Hessian H_{ij} and the mixed derivatives matrix are calculated using equations (29) and (30);
(d) the Hessian is diagonalized; transformation matrix is constructed from normalized eigenvectors of the Hessian;
(e) unconstrained errors are calculated on the basis of equations (38)–(41);
(f) constrained errors are calculated on the basis of equation (42);
(g) the type is determined in the following way for each parameter:
(i) if , the parameter x_{j} is determined;
(ii) if , , the parameter x_{j} is partially determined;
(iii) if , , the parameter x_{j} is undetermined.
References
Bottomley, D., Omi, H. & Ogino, T. (2001). J. Cryst. Growth, 225, 16–22. Web of Science CrossRef CAS
Brandt, O., Waltereit, P. & Ploog, K. H. (2002). J. Phys. D Appl. Phys. 35, 577585. Web of Science CrossRef
Caro, L. D., Giannini, C. & Tapfer, L. (1996). J. Appl. Phys. 79, 4101–4110.
Caro, L. D. & Tapfer, L. (1995). Phys. Rev. B, 51, 4381–4387.
Giacovazzo, C., Monaco, H. L., Artioli, G., Viterbo, D., Ferraris, G., Gilli, G., Zanotti, G. & Catti, M. (2002). Fundamentals of Crystallography, 2nd ed., ch. 2, pp. 67–147. Oxford University Press.
Han, S.K., Oh, D.C., Song, J.H., Inaba, K., Yao, T. & Hong, S.K. (2012). Appl. Phys. Express, 5, 081101. Web of Science CrossRef
Laskar, M. R., Ganguli, T., Rahman, A. A., Mukherjee, A., Hatui, N., Gokhale, M. R. & Bhattacharya, A. (2011). J. Appl. Phys. 109, 013107. Web of Science CrossRef
Paskova, T. (2008). Editor. Nitrides with Nonpolar Surfaces: Growth, Properties, and Devices. Weinheim: WileyVCH.
Pietsch, U., Holy, V. & Baumbach, T. (2004). HighResolution Xray Scattering: From Thin Films to Lateral Nanostructures, Advanced Texts in Physics, 2nd ed. Berlin: Springer.
Romanov, A. E., Baker, T. J., Nakamura, S., Speck, J. S. & Group, E. U. (2006). J. Appl. Phys. 100, 023522. CrossRef
Yang, K., Anan, T. & Schowalterc, L. J. (1994). Appl. Phys. Lett. 63, 2789–2791. CrossRef Web of Science
Zur, A. & McGill, T. C. (1984). J. Appl. Phys. 55, 378–386. CrossRef CAS Web of Science
This is an openaccess article distributed under the terms of the Creative Commons Attribution (CCBY) Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original authors and source are cited.