Proceedings of Spie: Deep-Learning Enabled Modeling Tool Meta-Magus For Metadevice Optimization and Design
Proceedings of Spie: Deep-Learning Enabled Modeling Tool Meta-Magus For Metadevice Optimization and Design
Proceedings of Spie: Deep-Learning Enabled Modeling Tool Meta-Magus For Metadevice Optimization and Design
SPIEDigitalLibrary.org/conference-proceedings-of-spie
ABSTRACT
Over the past few years, Deep-learning (DL) based modelling solutions have been presented as an alternate to the time-
tedious and computationally draining conventional design and optimization procedure of metasurfaces. While designing a
phase-based transmission meta-device, such as meta-lenses and meta-holograms etc., the most crucial part is to optimize
its unit-cell to ensure maximum electromagnetic (EM) transmission amplitude and full phase coverage (0-2π). Most of the
DL-based solutions have resulted in accurate optimization to provide the desired transmission amplitude. But the abrupt
discontinuities in the phase response, makes it more challenging to map and predict the optimized structure for full phase
coverage. Here, we present a novel DL-based tool named as “Meta-Magus” to design transmission based metaholograms.
Meta-Magus consists of two parts: (i) unit-cell optimization, and (ii) phase mask generation. Here, the first component
takes target transmission amplitude, phase, material properties, and the wavelength aimed as input, process it via regression
based tandem neural network, and provide optimized unit-cell structural parameters as output. Target image whose
hologram is to be generated is fed to the second component which comprises of deep convolutional neural network to
generate the corresponding phase mask as output. A full-wave commercial simulator then maps the optimized unit-cell
onto the generated phase mask and generates the intended meta-hologram. Simulation results of the generated designs
exhibit perfect holography, and validates that the model yields excellent predictions of a complete metasurface design from
scratch within a matter of seconds.
Keywords: Metasurface, Deep-learning (DL), metaholograms, Meta-Magus, U-Net, tandem neural network.
1. INTRODUCTION
Metasurfaces, the most successful and flourishing type of artificially engineered 2D structures, are
comprised of periodic assemblies of nano-resonators exhibiting exceptional powers to control and
maneuver the incident Electromagnetic (EM) waves [1-13]. The geometric dimensions and the spatial
orientation of these nano-resonators generates strong resonances causing sudden phase changes,
providing full control over the phase, amplitude, polarization of the EM waves [14-21]. Due to these
unique features and compactness, Metasurfaces have emerged as the best alternative to the hulking
conventional optical components and a perfect on-chip integrable solution for Photonic Integrated IC’s
[22-26]. Thus they have completely transformed the optics and photonics industry by providing
immensely efficient, smart and compact, on-chip portable devices [27-34]. Owing to these capabilities
they provide many interesting applications such as absorbers [35-38], holograms [39-42], meta-
mirrors [43-45], OAM multiplexing [46], reflectors [47-49], 3D integrated circuits [50-54], on-chip
devices [55, 56], amplifiers [57], self-healing and non-diffracting beams generation [58-60].
Widespread use of these metasurface’s in practical applications is limited by their overpriced
fabrication along with their sluggish and tiresome design procedure. The fundamental dimensional
parameters and the orientation of a meta-atom, that serves as the building block of a metasurface,
2. METHODOLOGY
A block diagram of the proposed metasurface optimization and design development tool “Meta-
Magus” is shown in the figure 1. A prototype of the cylindrical nano-rod meta atoms under
consideration is also illustrated in this figure. Optical response of each cylindrical meta-atom depends
upon its geometrical dimensions such as its radius (R), height (H), Period (P) and its material
characteristics such as the refractive index (n) and extinction coefficient (k). The multipole resonances
created in the dielectric nano-rod will change if any of these dimensions or the material attributes
changes, which will significantly alter the output EM amplitude and phase. Therefore, the Meta-Magus
takes the target transmission amplitude and phase (for eight distinct radii within the range defined by
Rmin and Rmax), the material properties refractive index (n), extinction coefficient (k) and operating
wavelength (λ) as input, pass it to its first component i.e., the Tandem neural network (TNN) which is
further divided into two parts which are the forward mapping neural network and the inverse design
neural network. Tandemly training these two networks yields the optimized set of dimensions for the
corresponding meta-atom to achieve the target response. The second component of the Meta-Magus
is a convolutional neural network based U-Net that takes the image whose hologram is to be generated
as input and provides the corresponding phase mask for the complete metasurface with any numerical
calculations and iterative Fourier and inverse Fourier transforms. Mapping this phase mask to the
optimized set of dimensions of the nano-rod will yield the final hologram. Here we discuss the two
components or Meta-Magus, their architecture and their functionalities in detail.
Our main objective is to harness and generalize the complicated inverse relationship between
the transmitted spectrum of a metasurface and its physical and geometrical characteristics using a deep
neural network. However, due to the inverse design's non-uniqueness problem, training a deep neural
network is extremely difficult. More specifically, numerous distinct metasurface configurations can
provide the same EM field response resulting in some contradictory training samples. This will create
major convergence issues for the inverse neural networks because the training data set will ineluctably
include some instances with different output labels for the same input. To overcome this issue, we
demonstrate a tandem neural network (TNN) which is composed of two sub-networks i.e., the forward
mapping neural network (FMNN) and the inverse optimization neural network (IONN).
First the FMNN is trained using the dataset collected from commercial EM software simulations. For
dataset procurement, a glass substrate and a high indexed dielectric nano-rod; for instance,
hydrogenated amorphous silicon (a-Si:H), titanium dioxide (TiO2) and Gallium nitride (GaN) are
selected for the meta-atom due to their transparency window in the visible regime and substantially
high efficiencies. Since the intrinsic properties of a material are also a function of the operating
wavelength (λ), so λ must also be incorporated in the mapping network. For instance, we select three
visible wavelengths i.e., 404nm, 532nm and 633 nm. An ideal cylindrical meta-atom must provide
maximum possible transmission amplitude and full phase coverage (0 to 360o) when the radius is
increased from a minimum (Rmin) to maximum (Rmax) value. Therefore, we simulate dielectric nano-
cylinders (a-Si:h, TiO2 and GaN) with different combination of Rmin, Rmax, H and P at the selected
visible wavelengths. These simulations are performed in the commercial EM software with periodic
boundary conditions and circular incident light. For each simulation, all the input structural dimension
values are saved as the input of the forward mapping neural network (FMNN) and output of the inverse
optimization network (IONN). The operating wavelength and the material’s n and k are also saved as
the secondary input of both the networks. The output transmission amplitude and phase at 8 discrete
levels of the radius between Rmin to Rmax are saved as the output of the FMNN and input of the
IONN.
Fig. 2. Depicts the overall layout of the proposed Tandem neural network. First the FMNN
explained above is trained over the collected dataset of 2000 samples. The Pre-trained FMNN is then
used to generate a larger dataset of 5000 samples for the inverse optimization neural network since it
requires more training samples to converge. This pre-trained FMNN is tandemly connected with the
proposed inverse optimization (IONN) network. The FMNN is kept frozen and rest of the tandem
network is trained to predict the best suitable set of nano-rod’s structural parameters to achieve a
specific target response at a given operating wavelength. The architecture and hyper parameters if the
IONN are exactly the same as the FMNN, but the inputs and outputs are interchanged. The input of
the IONN is a [16 × 1] defined as XIONN = [AR1, AR2, …, AR8, ΦR1, ΦR2 ,…, ΦR8] that corresponds to
the target transmission amplitude and phase at 8 discrete levels of the radius between Rmin to Rmax
Here, LossG is calculated at the intermediate layer, G and G’ represents the set of ground-truth
geometrical parameter’s and predicted geometrical parameters respectively. Rest of all the terms are
the same as that for MSEFMNN.
IONN
A tandem arrangement of inverse optimization neural network and forward mapping neural network
is trained to minimize the MSETNN given by equation 2 and 3 and accurately predict the nano-rod
geometrical parameter. Adam optimizer with a learning rate of 10-3 and 1000 epochs is used. The
average test MSEIONN achieved is 4.3 × 10-3. Fig. 5 shows the comparison of the predicted and the
ground-truth geometrical parameters. Three test example are chosen to be displayed from each
belonging to a different material and different wavelength to show the model’s accuracy for all the
materials and wavelengths. It is evident from fig. 5 that the predicted amplitude and phase values
perfectly matches the simulated values for a set of radii between Rmin and Rmax. The material
characteristics and the operating wavelength for each example is written in the middle in the fig. 5. in
red color.
Fig. 5. Comparison of ground-truth and TNN predicted nano-rod geometrical parameters (nm)
Fig. 6. Comparison of the GS algorithm generated ground-truth phase masks and the predicted phase mask (obtained by
performing inverse PCA transform on the CNN’s output)
After getting the optimized unit cell dimension’s and the predicted phase mask, a meta-hologram
device containing 100 × 100 nano-rods is designed in the FDTD Lummerical software, and full-wave
analysis is performed to get the desired hologram. Fig. 7 shows the comparison of the holograms
generated by GS algorithm phase mask and the CNN predicted phase mask. It can be clearly seen that
both the holograms are exactly alike and they reproduce the input image accurately in the metasurfaces
far field. Thus we can replace the conventional time consuming unit cell optimization as well as phase
mask generation’s parts with DL enabled models and automate the whole procedure via trained
networks.
4. CONCLUSION
A novel, DL-enabled tool named as “Meta-Magus” is demonstrated to speed up the conventional
design and optimization procedure of metasurfaces. Meta-Magus exhibit two constituent components
i.e., (i) unit-cell optimization and the (ii) phase mask generation components. The first component
consists of a tandem neural network (TNN) formed by the tandem arrangement of a pre-trained
forward mapping neural network (FMNN) and an inverse optimization neural network (IONN). The
FMNN is trained earlier to take all the structural and material characteristics of nano-rods as input and
predict the transmission amplitude as well as phase for a set of radii between R min and Rmax. The
average test MSE achieved for FMNN is 3.5 × 10-3. Tandemly training the FMNN and IONN results
in excellent predictions of the optimum set of geometrical parameters to achieve the target
electromagnetic response. Test MSE achieved for TNN assembly is 5.2 × 10-3. The second component,
which consists of a deep convolutional neural network, receives the target image whose hologram is
to be produced as input and produces the associated phase mask as an output. The dimensionality of
the ground truth phase masks is reduced using principle component analysis and then provided at the
output of the proposed CNN. Once the CNN is trained, it predicts the output with an MSE as low as
3.4 × 10-3, this output is passed through inverse PCA transform to achieve the predicted phase mask.
The optimized unit-cell is then mapped onto the created phase mask using a full-wave commercial
simulator, which creates the desired meta-hologram. A complete metasurface design may be created
from scratch in a matter of seconds, according to simulation findings of the developed designs, which
show perfect holography.
REFERENCES
[1] M. Q. Mehmood, J. Seong, M. A. Naveed et al., “Single‐Cell‐Driven Tri‐Channel Encryption Meta‐Displays,” Advanced
Science, 2203962 (2022).
[2] I. Javed, J. Kim, M. A. Naveed et al., “Broad-Band Polarization-Insensitive Metasurface Holography with a Single-Phase
Map,” ACS Applied Materials & Interfaces, (2022).
[3] M. A. Naveed, J. Kim, I. Javed et al., “Novel Spin‐Decoupling Strategy in Liquid Crystal‐Integrated Metasurfaces for
Interactive Metadisplays,” Advanced Optical Materials, 2200196 (2022).
[4] M. D. Shafqat, N. Mahmood, M. Zubair et al., “Highly Efficient Perfect Vortex Beams Generation Based on All-Dielectric
Metasurface for Ultraviolet Light,” Nanomaterials, 12(19), 3285 (2022).
[5] D. R. Smith, W. J. Padilla, D. Vier et al., “Composite medium with simultaneously negative permeability and permittivity,”
Physical review letters, 84(18), 4184 (2000).
[6] S. Ijaz, A. S. Rana, Z. Ahmad et al., “The Dawn of Metadevices: From Contemporary Designs to Exotic Applications,”
Advanced Devices & Instrumentation, 2022, (2022).