JTu2A.

119 CLEO 2023 © Optica Publishing Group 2023

Inverse Design of Photonic Structures Using

Yi Huang,1 Hong Tang,1 Bowen Zheng,1 Yunxi Dong,1 Mohammad Haerinia,1

Abstract: We present an automatic differentiable R-DIT for the fast inverse design of pho-
tonic structures. We demonstrated that the proposed method could achieve a 30% speedup
per optimization on GPUs compared with the differentiable RCWA. ©2023TheAut hor(s
)

Automatic differentiation (AD) has become a powerful tool for different inverse design applications where AD
provides an alternative to the adjoint methods, avoiding the need to derive derivative formulas. Rigorous coupled-
wave analysis (RCWA) [1] is a widely-used technique capable of predicting the electromagnetic response of
spatially-periodic composites. For meta-optics applications, Colburn and Majumdar [2] have successfully imple-
mented rigorous coupled-wave analysis (RCWA) on AD frameworks by generalizing complex eigenproblems with
the Lorentzian broadening technique. However, the performance of this AD-RCWA, as well as the performance
of all RCWA algorithms, are significantly constrained by the requirement of the eigen decomposition, the time-
consuming task that decelerates the computing on both central processing unit (CPU) and graphics processing unit
(GPU) platforms, when solving for large, complicated models with many spatial modes.
Rigorous diffraction interface theory (R-DIT) [3] formalism has been proposed to speed up the optics anal-
ysis of thin periodic composites. Here we present an extension of the R-DIT framework based on automatic
differentiation. As an example, we applied the developed automatic differentiable R-DIT to optimize a guided
mode resonance filter (GMRF), demonstrating significant speedup over AD-RCWA. Furthermore, the proposed
approach could accelerate the numerical calculation and optimization of other complex metasurfaces.
One of the key features of R-DIT is solving Maxwell’s equations in the Fourier domain without solving eigen-
value problems, which is briefly explained in the following. Maxwell’s equation in the Fourier domain can be
expressed below (the negative sign convention e− jkz is used in this text):
d E⃗ 0 P̂ E⃗
    
= (1)
dz̃ H⃗ Q̂ 0 H⃗

where E⃗ and H⃗ are the tangential Fourier coefficients of the electrical field ⃗E and the normalized magnetic field
⃗ = − jpµ /ε H,
H̃ 0 0 ⃗ z̃ = k0 z is the normalized coordinate, and P̂ and Q̂ are block matrices containing the information
of the permittivity/permeability distributions of the specific layer. Instead of solving the eigenvalue problem for Eq.
(1), R-DIT relies on Taylor expansion of electromagnetic fields and reduces the optics of diffractive metasurface
layer to the generalized boundary condition relating the fields at the interfaces surrounding the metasurface:
" 2
# 
h3
I + h8 P̂Q̂ + · · · h
2 P̂ + 48 P̂Q̂P̂ + · · · E⃗
h h3 h2 H⃗ z̃=−k0 h
2 Q̂ + 48 Q̂P̂Q̂ + · · · I + 8 Q̂P̂ 2
" #
h2 h3
− 2 P̂ − 48 P̂Q̂P̂ + · · · E
h ⃗
 
I + 8 P̂Q̂ + · · ·
= h3 2 (2)
− h2 Q̂ − 48 Q̂P̂Q̂ + · · · I + h8 Q̂P̂ H⃗ z̃=+k0 h
2

where h is the thickness of the metasurface layer, z = 0 is its center,and I is the identity matrix. It is also worth
mentioning that since only matrix multiplications are included in the derivation, there is no need to rewrite the
backward computations on AD frameworks explicitly.
To demonstrate the efficacy of the proposed method, a guided mode resonance filter (GMRF) [4] operating at the
free-space wavelength of around λ0 = 1540 nm high-frequency selectivity was designed [Fig. 1 (a,b)]. We started
 JTu2A.119 CLEO 2023 © Optica Publishing Group 2023

Fig. 1. (a) Geometry of the hexagonal unit cell of the proposed GMRF: a = 575 nm, b = 995.9 nm,
r is the radius of the holes to be optimized; (b) Cross-section of the GMRF under investigation:
top SiO grating (εSiO = 2.101, h1 = 230 nm), SiN waveguide (εSiN = 3.748, h2 = 345 nm) and
fused silica substrate (semi-infinite layer with εfs = 2.280), the TE-mode plane wave is normally
incident from the air; (c) Spectrum with the initial radius r = 400 nm and it with the optimized
radius r = 450 nm; (d) Radius and gradient of the transmission efficiency with respect to the radius
over the iterations; (e, f) Forward and backward simulation times and speedup of RCWA and the
10th order of proposed R-DIT.

from an initial value r = 400 nm, with a reflection peak at λ = 1541 nm [Fig. 1 (c)]. The goal of the inverse design
is to move the resonance to 1537 nm by changing the radius r. Based on the gradient descent method, the loss
function was reduced by either maximizing the reflection or minimizing the transmission at 1537 nm according to
the gradient information from the 10th order of R-DIT (10th order of Taylor expansion). The optimized response
is shown in Fig. 1 (c). The derivative of the transmission efficiency with respect to the radius converges to 0 over
iterations, and the optimal radius converges at r = 450 nm [Fig. 1 (d)].
To better understand how the proposed R-DIT performs during inverse design, we used a batch of GMRF
models (batch size 100 and 10 frequencies) to go through the same optimization process as presented above with
different spatial modes on the GPU platform (NVIDIA Tesla T4 offered by Google Colab). As a reference, we
also developed a differentiable RCWA based on [2]. Fig. 1 (e) and (f) show the performance comparison between
these two approaches. It is observed that, compared with RCWA, the developed R-DIT (10th order) has sped up
the forward simulation time for about 20% to 70% (increased with the spatial modes since larger matrices perform
much better with pure matrix multiplications than the complicated eigendecomposition). On the other hand, the
backward simulation time is very similar because the back-propagation computation within AD frameworks relies
on the chain rule, in which only generated computing graphs are used. Overall, the developed R-DIT achieved a
10% to 30% speedup for an optimization iteration compared to RCWA.
In summary, we presented an automatic differentiable R-DIT platform, which can perform fast gradient-based
inverse design. To validate the proposed approach, a GRMF was designed and optimized. It is demonstrated that
the proposed R-DIT can accelerate the computational process compared to existing methods (i.e., RCWA). It is
expected that the proposed platform can be combined with machine learning techniques for fast inverse design of
non-intuitive photonic structures.
This research was supported by the National Science Foundation (NSF) (awards No. 2132929 and 2118787).

References
1. M. G. Moharam and T. K. Gaylord, “Rigorous coupled-wave analysis of planar-grating diffraction,” Journal of the
Optical Society of America, vol. 71, p. 811, 7 1981.
2. S. Colburn and A. Majumdar, “Inverse design and flexible parameterization of meta-optics using algorithmic differen-
tiation,” Communications Physics, vol. 4, p. 65, 12 2021. RCWA.
3. C. M. Roberts and V. A. Podolskiy, “Rigorous diffraction interface theory,” Applied Physics Letters, vol. 110, 4 2017.
4. A. A. Mehta, R. C. Rumpf, Z. A. Roth, and E. G. Johnson, “Guided mode resonance filter as a spectrally selective
feedback element in a double-cladding optical fiber laser,” IEEE Photonics Technology Letters, vol. 19, pp. 2030–2032,
12 2007.