PaperlayerView exact revision on arXiv

Source-linked AI summary

Compounding meta-atoms into meta-molecules with hybrid artificial intelligence techniques

Zhaocheng Liu, Dayu Zhu, Kyu-Tae Lee, Andrew S. Kim, Lakshmi Raju, Wenshan Cai

arXiv:1907.03366v1physics.opticsphysics.comp-ph

TL;DR

Inverse design of metasurfaces with meta-molecules remains unresolved despite their large degrees of freedom. The paper combines compositional pattern-producing networks with cooperative coevolution and reports experimentally verified gradient metasurfaces for polarization and wavefront control.

  • Problem

    Inverse design of metasurfaces with meta-molecules remains unresolved despite their astronomical degrees of freedom.

    Source: §1 · p. 3

  • Method

    The framework combines a CPPN generator with cooperative coevolution, splitting meta-molecule design into independently optimized meta-atom tasks.

    Source: §2 · pp. 3–4

  • Results

    The framework discovered metasurfaces with prescribed objectives, exemplified by experimentally verified gradient metasurfaces for polarization and wavefront control.

    Source: §5 · p. 7

  • Takeaways & Limitations

    The framework is presented as a versatile potential approach to inverse-design tasks with on-demand, user-defined objectives.

    Source: §5 · p. 8

Abstract

from arXiv · show

Molecules composed of atoms exhibit properties not inherent to their constituent atoms. Similarly, meta-molecules consisting of multiple meta-atoms possess emerging features that the meta-atoms themselves do not possess. Metasurfaces composed of meta-molecules with spatially variant building blocks, such as gradient metasurfaces, are drawing substantial attention due to their unconventional controllability of the amplitude, phase, and frequency of light. However, the intricate mechanisms and the large degrees of freedom of the multi-element systems impede an effective strategy for the design and optimization of meta-molecules. Here, we propose a hybrid artificial intelligence-based framework consolidating compositional pattern-producing networks and cooperative coevolution to resolve the inverse design of meta-molecules in metasurfaces. The framework breaks the design of the meta-molecules into separate designs of meta-atoms, and independently solves the smaller design tasks of the meta-atoms through deep learning and evolutionary algorithms. We leverage the proposed framework to design metallic meta-molecules for arbitrary manipulation of the polarization and wavefront of light. Moreover, the efficacy and reliability of the design strategy are confirmed through experimental validations. This framework reveals a promising candidate approach to expedite the design of large-scale metasurfaces in a labor-saving, systematic manner.

1. Introduction

The paper presents a hybrid CPPN–cooperative coevolution framework for designing meta-molecules by dividing their inverse design into smaller meta-atom design tasks. It applies the framework to polarization and wavefront control and experimentally validates selected designs.

1. Introduction

  • Meta-molecules combine varied meta-atoms whose collective effects enable properties such as anomalous deflection, arbitrary phase control, and asymmetric polarization conversion.The paper assumes adjacent meta-atoms are not strongly coupled, allowing overall properties to be predicted from their constituents; it says this holds for most metasurfaces with discrete, spatially variant building blocks.Source: §1 · p. 2
  • Meta-molecule inverse design is hindered by intricate multi-structure mechanisms, numerous possible combinations, and costly 3D full-wave simulations; parameter sweeps and enumeration also restrict candidate shapes.The conventional workflow specifies a parameterized geometry class, sweeps its parameters, and enumerates combinations to meet the objective.Source: §1 · p. 2
  • Adjoint optimization [12-17] and genetic algorithms [18-20] require many simulations and face slow convergence or local minima, while learned models can expedite repeated designs after dataset construction [21-27].The introduction also notes that combining deep learning with optimization can mitigate local-minimum and dimensionality problems, but meta-molecule design remains unresolved because of its enormous combination space.Source: §1 · pp. 2–3
  • The framework encodes candidate meta-atoms as low-dimensional CPPN patterns and uses cooperative coevolution to find combinations that meet a design objective.The CPPN generates two-dimensional patterns pixel by pixel, while cooperative coevolution divides large optimization problems into independently addressable subcomponents.Source: §1 · p. 3
  • The framework designs diatomic metasurfaces for prescribed polarization control and gradient metasurfaces for wavefront manipulation, with selected meta-molecules experimentally characterized.The paper notes that the same framework can be adapted to other objectives, including high-performance dielectric metasurfaces, with minimal alteration.Source: §1 · p. 3

2. The hybrid framework

The framework combines CPPN-generated meta-atom patterns with cooperative coevolution to design meta-molecules, and uses a neural-network simulator to accelerate metasurface design.

The hybrid framework

  • A CPPN trained on a user-defined geometric dataset generates patterns pixel by pixel from coordinates and a latent vector, enabling user-controlled shapes.For pixel i, the network predicts s_i = f(x_i, y_i, r_i, v), where v controls the generated pattern’s shape.Source: §2 · pp. 3–4
  • The CPPN can generate complex corners, curves, and slits, features described as crucial for strong resonance and coupling in novel photonic structures.Source: §2 · p. 4
  • Cooperative coevolution splits whole-molecule design into independent meta-atom tasks and round-robin optimizes each atom’s latent-vector population while treating other species as optimized.Each update decodes patterns through the CPPN, evaluates fitness, and applies selection, reproduction, and mutation until the desired fitness criteria are reached.Source: §2 · p. 4
  • A neural-network simulator trained on 8,000 full-wave meta-atom simulations approximates complex transmission and reflection coefficients with accuracy above 97% from 170 to 600 THz.The design example uses gold meta-atoms on glass, each 40 nm thick in a 320 nm square cell.Source: §2 · p. 4

3. Diatomic meta-molecules for polarization control

The framework designs diatomic metasurfaces for polarization conversion and experimentally validates selected designs against their target polarization states.

3. Diatomic meta-molecules for polarization control

  • At 800 nm, designs rotated x-polarized light to 15°, 30°, 45°, or 60°, or converted it to circular polarization; each was accomplished within 20 seconds.Network-simulator and FEM full-wave simulations showed polarization states matching the target rotation angles and ellipticities.Source: §3 · p. 5
  • A diatomic meta-molecule comprises two distinct meta-atoms, whose sufficiently weak coupling lets their scattered far-field waves be approximated as a superposition.This approximation allows the framework to break the meta-molecule design into separate meta-atom designs.Source: §3 · p. 5
  • For diatomic metasurfaces, simulated horizontal-input outputs had orientation angles Ψ = 31.3° and 43.7°, respectively, with minimal ellipticities, closely matching the 30° and 45° design objectives.Separately, measured polarization states perfectly matched the design objectives despite marginal fabrication-process disagreements.Source: §3 · p. 5
  • As incident polarization varied from 0° to 180°, the output states mapped to a great circle slightly tilted from the Poincare-sphere equator.This mapping was simulated for the designed metasurfaces.Source: §3 · p. 5

4. Gradient metasurfaces

The framework designs eight-meta-atom gradient metasurfaces for polarization conversion and wavefront control. Experimental measurements and simulations confirm the intended polarization and phase-gradient behavior, while revealing asymmetric responses under opposite circular polarizations.

4. Gradient metasurfaces

  • The eight-atom designs require adjacent atoms to scatter with equal amplitude and constant phase difference Δ𝜙, addressed using locally defined fitness functions for neighboring atoms.Although cooperative coevolution is greedy and does not guarantee optimal solutions, it empirically converged to the desired amplitude and phase distribution in the experiments.Source: §4 · p. 6
  • At 800 nm, an eight-atom meta-molecule was designed to convert x-polarized incident light to orthogonal y-polarized light and deflect it to a specific angle; polarization analysis found the conversion orthogonal.The framework identified a set of meta-atoms within 10 minutes on a GPU machine; FEM simulation illustrated the phase gradient and amplitude distribution induced by the eight meta-atoms.Source: §4 · p. 6
  • At 800 nm, a fabricated gradient metasurface flipped left circularly polarized light to right circularly polarized light and imparted a phase gradient for beam steering.Quarter-waveplate measurements established the polarization orthogonality; simulations confirmed the converted field’s polarization, phase gradient, and amplitude distribution.Source: §4 · pp. 6–7
  • The circular-polarization design combines geometric phase and material-induced phase delay, which constructively form the desired gradient for LCP incidence but destructively alter the phase profile for RCP incidence.Under RCP incidence, the phase shift deviates from a constant gradient, producing multiple diffractions and asymmetric behavior.Source: §4 · p. 7
  • Because inverse design does not impose a particular physical mechanism, predefined constraints, or human intervention, the framework can uncover devices and photonic phenomena involving complex light-matter interactions.The section also notes additional gradient-metasurface examples at different wavelengths and polarization states in the Supporting Information.Source: §4 · p. 7

5. Discussion and Conclusion

The framework combines CPPN-generated meta-atoms, cooperative coevolution, and neural-network simulation to design metasurfaces for polarization and wavefront control. Its gradient-free objectives and component-based approach support broader photonic and cross-disciplinary inverse-design tasks, while final physical-simulation checks can address occasional simulator errors.

Discussion and Conclusion

  • The framework combines CPPN pattern generation, cooperative coevolution, and neural-network simulation to design gradient metasurfaces for polarization and wavefront control.CPPNs create complex nanostructure patterns, while cooperative coevolution divides the multi-element design into individual subtasks; the work includes experimental verification.Source: §5 · p. 7
  • Stochastic evolutionary searches can approach a global minimum across fast runs, while inexpensive neural-network simulation reduces failed-design costs but warrants physical-simulation checks in final iterations.The authors recommend replacing the simulator with physical simulation in the last few optimization iterations to eliminate occasional flawed solutions caused by approximation errors.Source: §5 · p. 7
  • Gradient-free evolutionary optimization permits user-defined objectives spanning scattering, dispersion, chirality, absorption, nonlinear responses, and nearfield enhancement.The authors also describe applications to metasurfaces and photonic devices, including multilayered metasurfaces and achromatic metalenses, and potential exploration in materials science, condensed matter physics, and chemical syntheses.Source: §5 · pp. 7–8

Figures and captions

The captions show the hybrid framework’s CPPN and cooperative-coevolution modules, define the meta-molecule design setup, and illustrate polarization-manipulation and gradient-metasurface designs.

Fig. 1

  • Fig. 1 illustrates a CPPN that maps pixel coordinates and a bias vector to meta-atom patterns, alongside cooperative coevolution that iteratively optimizes species in round-robin fashion.The algorithm evaluates individuals and applies selection, reproduction, and mutation until all species meet preset fitness criteria.Source: Figure 1 · p. 9

Fig. 2

  • Fig. 2 defines each metasurface meta-molecule as several distinct meta-atoms in a lattice, with an example diatomic design using gold particles on glass.The stated common parameters are a 40 nm meta-atom thickness and a 320 nm square lattice side length.Source: Figure 2 · p. 10

Fig. 3

  • Fig. 3 shows designed diatomic meta-molecules for prescribed polarization rotation or conversion, with simulated states and fabricated rotators for 30° and 45° rotation.The caption includes neural-network and FEM simulations, SEM images, and measured output polarization states.Source: Figure 3 · p. 11

Fig. 4

  • Fig. 4 depicts gradient metasurfaces that convert polarization and steer beams for linear and LCP incidence, with opposite-handed circular outputs steered in different directions.For the LCP-designed metasurface, the caption also shows relative phase delays and simulated diffraction under LCP and RCP incidence.Source: Figure 4 · p. 12
Loading 1907.03366v1…