Synthesis of thick optical thin-film filters with a layer-peeling inverse-scattering algorithm

Johannes Skaar, Ligang Wang, and Turan Erdogan

We present an efcient and accurate method for synthesis of optical thin-lm structures. The method is based on a differential inverse-scattering algorithm and considers therefore both phase and amplitude reectance data. We apply the algorithm to the synthesis of lters with arbitrary index layers and two-material lters consisting of only high- and low-index layers. The layered structure is approximated by a stack of discrete reectors with equal distance between all reectors. This mirror stack is in turn determined from the desired, complex reection spectrum by a layer-peeling inverse-scattering algorithm. The complexity of the design algorithm is approximately the same as that of the forward problem of computing the spectrum from a known structure. 2001 Optical Society of America OCIS code: 310.0310.

1. Introduction

Optical thin-lm structures consisting of alternating layers of high- and low-index dielectric materials are vital components for many applications, including architecture, energy management, automobiles, scientic instruments, data storage, and display devices.1 In addition to their use as high-reection mirrors and antireection coatings, thin-lm structures form the basis for optical lters with sophisticated complex responses reection and transmission . In particular, thin-lm lters enable tailorable dispersion for ultra-short-pulse lasers,2 and extremely thick lms are now a key technology for wavelength-divisionmultiplexed WDM ber-optic communication systems. Thin-lm lters with hundreds of layers have become possible to fabricate,3 and hence the ability to synthesize structures with increasingly demanding lter characteristics is a critical tool. The calculation of the optical characteristics of a particular structure, i.e., the forward problem, is straightforward, whereas the calculation of a structure associated with specied optical characteristics,

J. Skaar johannes.skaar@fysel.ntnu.no is with the Department of Physical Electronics, Norwegian University of Science and Technology, N-7491 Trondheim, Norway. L. Wang and T. Erdogan are with The Institute of Optics, University of Rochester, Rochester, New York 14627. Received 8 June 2000; revised manuscript received 21 November 2000. 0003-6935 01 132183-07$15.00 0 2001 Optical Society of America

the inverse problem, is signicantly more difcult. Many approaches to the inverse problem for thin-lm design can be classied into two main categories: numerical renement and thin-lm synthesis.1,4 Both rely heavily on optimization algorithms and hence tend to be based on a reasonable starting guess at the nal structure. In this paper we propose what we believe is a new method for the synthesis of multilayer optical lters based on a specied complex reection spectrum. The method is a direct approach analogous to the Fourier-transform method.5 However, in contrast to the Fourier method, the layer-peeling inverse-scattering method is exact; i.e., the layers are exactly determined from the associated spectrum. The inverse-scattering algorithm is based on direct inversion of the transfer-matrix model after the following principle: Consider the desired impulse response that is determined by an inverse Fourier transform of the desired spectrum. At the leading edge of the impulse response the light sees only the rst layer, because at the very beginning of the impulse response the light does not have time to propagate more deeply into the structure. By this causality argument one can determine the rst reector and hence the rst transfer matrix. Then one can use the transfer matrix to propagate the elds to the next layer. Now one is in the same situation as at the beginning, since the effect of the rst layer is peeled off. The process is continued to the back of the lter. With this method any thin-lm lter structure can be reconstructed from the associated complex reection spectrum. For synthesis problems, however,
1 May 2001 Vol. 40, No. 13 APPLIED OPTICS 2183

the desired reection spectrum does not necessarily have to be realizable for a thin-lm lter, since the required Fresnel reectivities might be complex. To overcome this problem, we rst synthesize a mirror stack consisting of discrete, complex reectors and then approximate the stack by an inhomogeneous layer lter or a two-index thin-lm lter. The remainder of this paper is organized as follows: In Section 2 we give the basic principles behind the layer-peeling inverse-scattering algorithm. In Section 3 we rst consider the easiest problem, namely, to synthesize lters consisting of layers with arbitrary refractive indices. Although this class of lters might be realizable in practice, at least in the future, for practical reasons it is often desirable to design lters consisting of only two different refractive indices, high nh and low nl . Two-material structures could be achieved by means of transforming the arbitrary index layers with the standard Herpin equivalent-index concept.6 In this paper, however, we compute the two-material lters directly from the discrete reectors, and this procedure is described in Section 3. In Section 4 we present numerical examples of lters in both categories.
2. Layer-Peeling Inverse-Scattering Algorithm

Fig. 1. Stack of discrete reectors. The distance between the reectors is d, and the complex elds before the jth section are Aj and Bj .

cell is therefore described by the transfer-matrix product Td T , j, where Td exp ikd 0 0 exp ikd (2)

is the standard propagation matrix, and T

,j

2 j

1 2

1
j

* j 1

(3)

In this section we review the basic principles behind the differential inverse-scattering algorithm. It was rst developed by geophysicists for the recognition of layered-earth media from scattering data. A thorough analysis and comparison with other algorithms can be found in Bruckstein et al.7,8 Whereas Bruckstein used the algorithm only for real reectors, we extend the algorithm to consider complex reectors. This is done to enable synthesis of asymmetric lter characteristics, which require quasi-periodical structures. We state the inverse-scattering problem as follows. We assume a stack of N discrete, complex reectors with distance d between all reectors. From the corresponding complex reection spectrum r k we wish to reconstruct the complex reector amplitudes j , j 1, 2, . . . , N, where k 2 is the vacuum wave number. This inverse scattering problem is inherently discrete in nature and may be solved exactly by the layer-peeling algorithm. The running time of this algorithm is of the same order O N 2 as the conventional approach for computing the forward problem, i.e., computing the spectrum r k from the reector amplitudes j with the transfer-matrix method.9 For a physical picture of how the layer-peeling algorithm works we imagine that the lter is probed by a delta-function pulse. Thus the forward- and the backward-propagating elds before the rst section, A1 and B1, respectively, are A1 k B1 k 1 r k . (1)

describes a pure, discrete reector. The discrete * as the reector is generally complex, with j and j reection coefcients from the left and the right, re2 1 2 spectively, and 1 as the transmission coj efcient for both directions. If we manage to determine the rst reector, we can use Td T ,1 to transfer the elds to the next section. We then nd ourselves in the same situation as with the rst reector, so in effect the rst layer is peeled off. This procedure can be repeated until the entire stack is determined. To nd the complex amplitude of the rst reector, we note that the impulse response of the reector stack for time t 2d c is independent of the reectors j for j 2, because light does not have sufcient time to propagate to and from the second and higher reectors. Thus, when looking at the impulse response of the stack for t 2d c, we obtain the same response as if reector 1 were alone. Therefore we can compute 1 as the inverse Fourier transform of r k B1 k A1 k evaluated at time t 0. More precisely, because r k is periodic with period d, we nd d
1 period

B1 k dk A1 k

or

1 M

M m 1

B1 A1

,
m

(4)

Each unit cell of the reector stack is composed of the discrete localized reector followed by a pure propagation section, as illustrated in Fig. 1. The jth unit
2184 APPLIED OPTICS Vol. 40, No. 13 1 May 2001

where m denotes the sample number of the ratio B1 A1 and M N is the number of spectral points, in one spectral period. We must ensure that M is chosen to be sufciently large to represent the target r k with a desired accuracy. Strictly speaking, we must have M to make the layer-peeling algorithm truly exact. However, in practical situations, it turns out that M does not have to be considerably greater than N. For the examples in this paper we used M N. One way of getting the layer-peeling procedure to be strictly exact is to transform it to the time domain

and represent the intermediate results elds as time-domain responses.7,8 We can summarize the algorithm in the following simple steps: i Start with a physically realizable reection coefcient r k dened in a design bandwidth d see the remark below . ii Compute 1 from Eq. 4 . iii Transform the elds with the transfer matrices or with the equivalent Schur recursion expression B2 k A2 k B1 k A1 k 1 i2kd . * B1 k A1 k 1 1
Fig. 2. Refractive-index prole of the thin-lm lter. The dened Bragg period, or unit cell, is the region between the dotted vertical lines. The discrete reectors j are indicated below the prole.

exp

(5)

Equation 5 results from the transfer-matrix multiplication and is similar to a recursion formula proposed by Schur for testing the boundedness of an analytic function outside the unit circle of the complex plane.8 iv Return to step ii until the entire lter is determined. When the layer-peeling algorithm is used for synthesis of actual lters, the desired reection spectrum r k is not necessarily realizable, even for a stack of complex reectors. To obtain a realizable reection spectrum, we use the windowing procedure that is common in digital far-impulse-response FIR lter design.10 The same procedure was used by Feced et al. for the synthesis of ber Bragg gratings with a layer-peeling algorithm.11 First, we compute the inverse Fourier transform of the desired spectrum to obtain the impulse response. This response is then forced to be causal and realizable when we multiply it by a window function and shift it so that it is nonzero only for times t 0. The window function should be smooth to reduce the Gibbs phenomenon. The realizable r k is nally obtained by Fourier transform of the windowed impulse response. Note that the target impulse response has a nite duration. Strictly speaking, the impulse response of any stack consisting of two or more mirrors must have innite duration, so the target response is not exactly realizable for a stack of N reectors. Thus there is always an unwanted tail in the realized impulse response. The realized tails cause undesirable uctuations in the realized reection spectrum, but as the examples below and the examples in Ref. 11 indicate, the inuence of the tail is small for most practical lters with 1. N 1 and j
3. Design of Thin-Film Structures

are related to the discrete reection coefcients by the relation qj 1 d * j

j

arctanh

(6)

In this paper, however, our goal is a layered thin-lm lter. For simplicity we neglect all kinds of loss in the structure. First, we consider the simplest problem, namely, to nd a layered structure with arbitrary indices. The indices nj must satisfy nl nj nh, where nl and nh are the lowest and the highest realizable indices, respectively. From the Fresnel reection coefcients it is straightforward to obtain index jumps that realize real j provided that
j

nh

nl

nh

nl .

(7)

In this section we show how to convert the stack of complex, discrete reectors into a physical lter. If we were designing a corrugated lter, e.g., a ber Bragg grating, this procedure would be simple: Comparison with Feced et al.11 shows that the samples of the complex coupling coefcient of the grating

The refractive-index prole must be held between the limits nl and nh. If the required index of the jth layer becomes too small or too large, one can simply insert a section corresponding to a round-trip phase shift of to reverse the index jump. The phase of j must be realized by means of deviating the positions of the index jumps compared with the positions of the discrete reectors. This is clearly an approximation, since the induced phase response will be frequency dependent linear phase , whereas the discrete reectors are independent of frequency. However, this approximation is apparently good for bandwidths that are much less than the center frequency, k k0, which usually is true for the usable bandwidth of thin-lm lters. In practice it is often more desirable to design twomaterial thin-lm lters consisting of only high nh and low nl refractive-index layers. Such lters can be designed from the corresponding inhomogeneous layers above by the standard Herpin equivalentindex concept.6 In this study, however, we present another, more direct, approach of obtaining the layer thicknesses from the desired stack of discrete reectors. By using the transfer matrices, we show that a Bragg period see Fig. 2 can be approximately represented by a discrete reection coefcient. We assume that the grating is perfectly matched to
1 May 2001 Vol. 40, No. 13 APPLIED OPTICS 2185

the surrounding media the refractive indices of the surrounding media are assumed to be nl . Furthermore, we dene dh, j and dl, j as the thicknesses and k nh dh, j and l, j k nl dl, j as the phase shifts in h, j the high- and the low-index layers, respectively, all in the jth Bragg period or unit cell see Fig. 2 . Using a standard transfer-matrix method, we nd the complex reection coefcients of each unit cell to be9 rj 2ir sin exp i
2 h, j h, j

r exp

,
h, j

(8)

where the parameter r nh nl nh nl denotes the Fresnel reectivity at each index step. Equation 8 results from a transfer-matrix multiplication of the two index-step matrices inserted between the appropriate propagation matrices. Now, we want to represent the unit cells with discrete, complex reectors to be able to apply a layer-peeling synthesis algorithm. To force the reection coefcient 8 to be independent of wavelength, we must make an approximation. Therefore we dene the amplitude of the wavelength-independent discrete reector to be
j

rj

2r sin
k kB

h, j 1 2 h, j k kB

2r cos 2

(9)

where kB denotes a design Bragg wave number and we have assumed 0 . The phase of the h reection associated with the jth unit cell is
r, j

pears when 2kB d m 2 for integer m. We recall that the spectrum of the reector stack is periodic with period equals the design bandwidth k d, and, consequently, the choice 2kB d m 2 or kB m k ensures that the design wave number corresponds to k 0. The goal spectrum will therefore be centered around k 0, which makes j j 1 small. In practice we will therefore choose the design bandwidth according to k kB m with a suitable integer m and rather pad the desired spectrum with zeros if necessary. The integer m serves then as a parameter that scales the inverse lter bandwidth and the layer thicknesses for a given design wave number. For the examples in this paper we have set m 1. Now that we have modeled the thin-lm lter as a stack of discrete, complex reectors j with identical distances between all reectors, then once the discrete reectors i are found by use of the layer-peeling method, the layer phase shifts can be computed from Eq. 11 and through inversion of Eqs. 9 and 10 : From the known j and j , Eq. 9 gives h, j for all j. The remaining unknowns l, j are found by substitution of Eq. 10 into Eq. 11 . The solution of Eq. 9 is conveniently computed by numerical packages. However, for most practical situations, the Fresnel reectivity of each interface is small, or r 1, so that Eqs. 9 11 may be approximated by the relations
j

2r sin
j 1

h, j

, m 2
h, j h, j 1

(12) . (13)

arg r j

arctan tan

h, j

1 1

r2 , r2

2 (10)

l, j

where we have ignored the phase shift that is due to the factor i in Eq. 8 , since it is the same for all unit cells. We account for the variation in the phase shifts r, j resulting from variation in the layer thicknesses by assigning phase factors to the discrete reection coefcients j , such that j j exp i j . That is, we let the distance between the discrete reectors be constant and account for chirp that is due to variation in the layer thicknesses, using j . At the design wave number k kB the phase difference between reections from two neighboring reectors must equal the phase difference between reections from two neighboring unit cells so that
j j 1

In fact, this approximation can be justied even if the Fresnel reectivity is relatively large, since the approximation that the reection coefcient of each unit cell is independent of wave number is usually more signicant. Therefore Eqs. 12 and 13 are sufcient to compute the layer phases from the synthesized stack of discrete reectors. Finally, we obtain the layer thicknesses from the denitions h, j kBnhdh, j and l, j kB nl dl, j, where kB denotes the design Bragg wave number.
4. Numerical Examples

2kd

r, j

r, j 1

h, j 1

l, j 1

(11) for k kB. Here we have moved the reection reference planes of the unit cells j and j 1 to the same position, namely, to the beginning of unit cell j 1. This transformation, which is done to enable comparison of the phases, yields the nal two terms on the right-hand side in Eq. 11 see Fig. 2 and the roundtrip phase term 2kd on the left-hand side. Equation 11 should ideally be valid for all wave numbers, but since the difference j j 1 is independent of k, whereas the other terms are not, this is clearly not possible. The best approximation is obtained when j j 1 is as small as possible, which usually ap2186 APPLIED OPTICS Vol. 40, No. 13 1 May 2001

As a numerical example we rst apply the synthesis algorithm to design dispersionless square lters with high reectivities. Such lters are potentially important for the wavelength-division-multiplexed communications technology of the future. The target lter is a square lter with maximum reectivity of 99.9%, no dispersion, and a bandwidth of 0.13 m. To make the target lter causal and realizable, the impulse response is apodized with a Kaiser10 window10 and shifted so that it starts at time t 0. We take the number of reectors to be 150. We have synthesized both an inhomogeneous layer lter Fig. 3 and a thin-lm lter consisting of high- and lowindex layers two-material lter, Fig. 4 . In the latter case the indices of the materials are set to 1.5 and 2.0, which are close to the indices of the standard dielectric materials SiO2 and Ta2O5, respectively. The total thicknesses of the resulting lters are 39

Fig. 3. Refractive-index prole of the inhomogeneous layer nondispersive bandpass lter.

and 52 m, respectively. The lter performances were computed by an exact transfer-matrix method and are shown in Figs. 5 and 6. In Fig. 5 the reectivity spectra for the two lters are compared. We observe a small ripple of 4 10 4 in the passband and a sidelobe level of 24 dB for both lters. As can be observed in Fig. 6, the ripple or deviation of the group-delay dispersion is 20 fs2 inside the passband for both the inhomogeneous layer lter and the twomaterial lter. Next, we show an example of synthesis of a dispersion-compensating bandpass lter. This type of lter has become important in ultrafast laser physics.2 The target dispersion for this example is 275 fs2; that is, the group delay is linear as a function of wave number. The dispersion is specied after the windowing process. The number of layers is set to 200. The maximum reectivity and the indices of the materials are the same as in the previous example. The result of the synthesis is shown in Figs. 710. As expected, the performance of the lters is about the same as that of their nondispersive counterparts, but the group-delay dispersion is nearly

Fig. 5. Reectivity of the nondispersive bandpass lter. Dashed curve, spectrum of the inhomogeneous layer lter; solid curve, spectrum of the two-material lter. The spectra are shown in both linear and logarithmic scales.

constant as a function of wave number in the reection band. Although we choose a linear group delay for the sample dispersion response, the power of the layer-peeling method is its ability to synthesize an arbitrary group-delay spectrum based on the known dispersive properties of the system that requires compensation. One of the main advantages of the layer-peeling algorithm compared with the conventional optimization methods is the efciency: Because the running time is of the same order as that of the forward-transfer-matrix calculation of the spectrum, we can design extremely thick lters quickly. The two calculations in this paper had associated run times of less than 0.2 s using the program MATLAB on a 300-MHz Pentium computer. In addition, it is evident from the examples that the layer-peeling method is particularly suitable for

Fig. 4. Layer thicknesses of the two-material nondispersive bandpass lter. Dots, thicknesses of the low-index layers dl, j ; dotted curve, ve times the thicknesses of the high-index layers dh, j .

Fig. 6. Group-delay dispersion for the nondispersive bandpass lters. Dashed curve, dispersion spectrum of the inhomogeneouslayer lter; solid curve, dispersion of the two-material lter. 1 May 2001 Vol. 40, No. 13 APPLIED OPTICS 2187

Fig. 7. Refractive-index prole of the inhomogeneous layer dispersive bandpass lter.

Fig. 10. Group-delay dispersion of the dispersive bandpass lters. Dashed curve, dispersion spectrum of the inhomogeneous layer lter; solid curve, dispersion of the two-material lter.

designing lters when the entire spectrum is relevant to the designer, not only a small bandwidth. For example, one observes that the spectra of the designed lters match well to the goal spectrum in both the reection pass band and also the stop band. The disadvantage is clearly that one cannot do trade-offs in the same way as with optimization methods. However, if a certain parameter is critical, for example, the group-delay dispersion, one can use numerical renement techniques to optimize for that parameter with the layer-peeling design as a starting point.
5. Conclusion
Fig. 8. Layer thicknesses of the two-material dispersive bandpass lter. Dots, thicknesses of the low-index layers dl, j ; dotted curve, ve times the thicknesses of the high-index layers dh, j .

We have proposed a method for synthesis of thin-lm lters, using a layer-peeling inverse-scattering algorithm. This method is particularly useful for designing lters with many lm layers and has the advantage that no knowledge of a reasonable starting guess at the nal structure is necessary. The method has been used for designing lters consisting of arbitrary index layers and for designing twomaterial thin-lm lters. The running time of the synthesis algorithm is approximately the same as the running time of the conventional transfer-matrix method for computing the spectrum from a known structure.
References and Note
1. J. A. Dobrowolski, Numerical methods for optical thin lms, Opt. Photon. News June 1997 , pp. 24 33. 2. N. Matuschek, F. X. Kartner, and U. Keller, Analytical design of double-chirped mirrors with custom-tailored dispersion characteristics, IEEE J. Quantum Electron. 35, 129 137 1999 . 3. S. R. A. Dods, Z. Zhang, and M. Ogura, Highly dispersive mirror in Ta2O5 SiO2 for femtosecond lasers designed by inverse spectral theory, Appl. Opt. 38, 4711 4719 1999 . 4. See, for example, commercial thin-lm design software on the Internet: http: www.qis.net bnichols opticsnotes thincad.htm. 5. J. A. Dobrowolski and D. Lowe, Optical thin lm synthesis

Fig. 9. Reectivity of the dispersive bandpass lter. Dashed curve, spectrum of the inhomogeneous layer lter; solid curve, spectrum of the two-material lter. The spectra are given in both linear and logarithmic scales. 2188 APPLIED OPTICS Vol. 40, No. 13 1 May 2001

program based on the use of Fourier transforms, Appl. Opt. 17, 3039 3050 1978 . 6. J. A. Dobrowolski and S. H. C. Piotrowski, Refractive index as a variable in the numerical design of optical thin-lm systems, Appl. Opt. 21, 15021511 1997 . 7. A. M. Bruckstein and T. Kailath, Inverse scattering for discrete transmission-line models, SIAM Soc. Ind. Appl. Math. Rev. 29, 359 389 1987 . 8. A. M. Bruckstein, B. C. Levy, and T. Kailath, Differential

methods in inverse scattering, SIAM Soc. Ind. Appl. Math. J. Appl. Math. 45, 312335 1995 . 9. H. A. Macleod, Thin-Film Optical Filters Adam Hilger, Bristol, UK, 1985 . 10. J. G. Proakis, Digital Signal Processing Prentice-Hall, Englewood Cliffs, N.J., 1996 . 11. R. Feced, M. N. Zervas, and M. A. Muriel, An efcient inverse scattering algorithm for the design of nonuniform bre Bragg gratings, J. Quantum Electron. 35, 11051115 1999 .

1 May 2001

Vol. 40, No. 13

APPLIED OPTICS

2189