PaperlayerView exact revision on arXiv

Source-linked AI summary

Intelligent metaphotonics empowered by machine learning

Sergey Krasikov, Aaron Tranter, Andrey Bogdanov, Yuri Kivshar

arXiv:2110.11618v1physics.opticsphysics.app-ph

TL;DR

This review surveys machine learning and artificial intelligence applications to metaphotonics, including nanoantennas and metasurfaces. Its examples span nanoparticle-spectrum prediction and metasurface solar-cell optimization, where the optimized cell achieved a 2.5 times larger short circuit current than solar cells without metasurfaces using the same amount of crystalline silicon.

  • Problem

    Designing nanoantennas can require searching thousands of material and layer-thickness combinations, with many calculations needed to match a desired scattering spectrum.

    Source: §IV · p. 5

  • Method

    The paper summarizes machine-learning concepts and reviews their applications to optical nanoantennas and metasurfaces, with examples across metasystems.

    Source: §I · p. 2

  • Results

    Across the examples, learning methods predict nanoparticle scattering spectra and optimize metasurface devices; the optimized solar cell had a 2.5 times larger short circuit current than any solar cell without metasurfaces, with the same amount of crystalline silicon.

    Source: §IV · p. 5, §V · p. 8

  • Takeaways & Limitations

    The examples apply machine learning to tune nanoantenna scattering responses and improve solar-cell optical performance.

    Source: §IV · p. 7, §V · p. 8

    Deep learning can require extensive data collection and training time, making it most justified for repeated parameterized tasks or intractable feature-extraction tasks.

    Source: §II · p. 3

Abstract

from arXiv · show

In the recent years, we observe a dramatic boost of research in photonics empowered by the concepts of machine learning and artificial intelligence. The corresponding photonic systems, to which this new methodology is applied, can range from traditional optical waveguides to nanoantennas and metasurfaces, and these novel approaches underpin the fundamental principles of light-matter interaction developed for a smart design of intelligent photonic devices. Concepts and approaches of artificial intelligence and machine learning penetrate rapidly into the fundamental physics of light, and they provide effective tools for the study of the field of metaphotonics driven by optically-induced electric and magnetic resonances. Here, we introduce this new field with its application to metaphotonics and also present a summary of the basic concepts of machine learning with some specific examples developed and demonstrated for metasystems and metasurfaces.

I. INTRODUCTION

AI and ML offer data-driven tools for photonics, including the design and optimization of metaphotonic systems. This Perspective introduces metaphotonics and surveys how these tools apply to nanoantennas and metasurfaces.

I. INTRODUCTION

  • Maxwell’s equations can generate training data for AI algorithms, which are applied in photonics to inverse design, optimization, and big-data processing.The authors describe these applications as supporting the development of precise photonic technologies.Source: §I · p. 1
  • Metaphotonics uses resonant dielectric nanostructures as meta-atoms for subwavelength metasurfaces and devices, with optically induced magnetic responses.Low-loss dielectric structures may realize concepts limited in optical plasmonics by metal losses.Source: §I · p. 1
  • High-index dielectric resonators can support geometry-tunable electric and magnetic Mie resonances whose interference shapes scattering and can enhance magnetic fields.The passage identifies potential effects in both linear and nonlinear regimes.Source: §I · p. 1
  • The Perspective reviews AI and ML applications to metaphotonics driven by optically induced electric and magnetic resonances, including nanoantennas and metasurfaces.It also summarizes basic ML concepts and examples for metasystems and metasurfaces.Source: §I · p. 2

II. BASIC CONCEPTS OF MACHINE LEARNING

This section introduces ML and DL, outlines training and generalization considerations, and explains how learned mappings support forward and inverse photonic design. It also notes the data and time costs that can constrain DL use.

II. BASIC CONCEPTS OF MACHINE LEARNING

  • ML algorithms learn mappings between input and output data, but they can find correlations without understanding the data’s physical meaning.Models may also incorporate physical intuition or constraints such as Maxwell’s equations.Source: §II · p. 2
  • DL is a subclass of ML based on layered artificial neural networks that learn data features to map inputs to outputs.Networks are trained by minimizing a loss function through weight adjustment; convolutional layers are used for image processing.Source: §II · p. 2
  • Reliable generalization requires representative training data, while high-dimensional networks can overfit and ML models are generally interpolative rather than extrapolative.Validation-set loss can be used to monitor overfitting.Source: §II · pp. 2–3
  • DL can require substantial training data and time, so its use is usually justified for repeated tasks with slightly varying parameters or when feature extraction is intractable.The passage says data needs typically depend on problem complexity and that more data can help mitigate overfitting.Source: §II · p. 3
  • For photonics, forward design predicts a structure’s physical response, while inverse design seeks structure parameters that produce a desired response.DL can serve as a surrogate model: a data-driven approximation rather than a simulation.Source: §II · p. 3, Figure 2 · p. 4
  • For a spherical particle, an ANN learns a radius-to-scattering-cross-section mapping and can be used in reverse to estimate parameters for a target response.The example’s transformations are not constrained to follow a physical equation, theory, or model.Source: §II · p. 3
  • With enough data, prediction accuracy may approach unity, and estimation speed may suit many real-time tasks, though real-world accuracy can require specialized architectures and algorithms.The passage cautions that the inverse-design procedure may not be straightforward in real-world tasks.Source: §II · p. 4

III. METASYSTEMS AND METASURFACES

Metasystems and metasurfaces use structured subwavelength elements to manipulate light, with dielectric resonances offering control over multiple optical properties. Their thin form and potential functionalities motivate their development as optical components.

III. METASYSTEMS AND METASURFACES

  • Metadevice functionality can be realized by structuring functional matter on the subwavelength scale, extending metamaterial concepts toward integrated photonic building blocks.The passage identifies nanoantennas and metasurfaces as materials building blocks to study and optimize.Source: §III · p. 4
  • Resonant plasmonic and dielectric nanoantennas can manipulate light, and combining electric and magnetic modes can modify radiation patterns or localize electromagnetic energy.The passage connects localization to destructive interference of leaky modes in bound states in the continuum.Source: §III · p. 5
  • Metasurfaces use arrays of subwavelength resonators to control phase, amplitude, and polarization with subwavelength resolution, unlike conventional components relying on phase accumulation during propagation.Their unit-cell geometry determines optical response, and the arrays are arranged for the intended functionality.Source: §III · p. 5
  • Metasurfaces have been explored for flat-optics devices that may offer ultra-thin, lightweight, ultracompact alternatives to conventional optical components.The passage describes their novel functionalities and potential to overcome several limitations of traditional counterparts.Source: §III · p. 5
  • All-dielectric metastructures can combine electric and magnetic Mie resonances to control optical properties, while plasmonic metastructures suffer significant losses and low efficiency.Optimization depends on the application; relevant resonances include Mie-type scattering, bound states in the continuum, Fano, and anapole resonances.Source: §III · p. 5
  • Advances in nanofabrication are bringing low-cost, large-area, mass-production approaches for metasystems and metasurfaces closer to maturity.The passage anticipates dielectric-metasurface flat-optics components appearing in daily life.Source: §III · p. 5

IV. ADVANCED NANOANTENNAS

This section surveys machine-learning approaches to designing nanoantennas, from fast prediction and inverse design of layered spheres to designs targeting complex scattering responses. It also describes transfer learning and applications extending from individual nanoantennas to arrays.

IV. ADVANCED NANOANTENNAS

  • Nanoantenna design can require iterating through thousands of material and layer-thickness combinations to match a desired scattering spectrum, motivating advanced optimization.The section frames DL as useful for this search over core-shell structures.Source: §IV · p. 5
  • A trained network predicts layered-sphere scattering spectra in under a second with mean relative error below 1.5%, and can be retrained for inverse design and spectral optimization.The model takes layer thicknesses of 30–70 nm as inputs and predicts spectrum samples from 400–800 nm.Source: §IV · p. 5
  • A hybrid two-network architecture jointly learns forward and inverse design while selecting three-layer-sphere materials from seven feasible refractive indices and tuning electric and magnetic dipole resonances.The approach supports resonance co-location, the first Kerker condition, and designs of media with negative refractive index.Source: §IV · p. 6
  • Beyond neural networks and spheres, Bayesian optimization designs cylindrical nanoantennas for Kerker or anti-Kerker scattering, while DL studies multipole responses and anapole states [35].The review also notes inverse designs for plasmonic nanoantenna emission and near- and far-field properties [49], while related chemistry and biology applications lie beyond its scope.Source: §IV · p. 6
  • DL-based forward designs of core-shell nanoantennas are 100–1000 times faster than FDTD simulations, with prediction accuracy about 95%.The review suggests these methods may become conventional numerical tools and notes the prospect of real-time web-based design tools.Source: §IV · p. 6
  • A five-layer Ag–SiO2 particle was designed for scattering efficiency below 10^-2 from 400–700 nm and below 10^-4 from 510–550 nm.Another design uses phase-change materials for switching between near-zero scattering and superscattering; an all-dielectric shell was also designed for optical cloaking.Source: §IV · p. 6
  • Transferring trained-network layers from nanoparticle scattering to multilayer-film transmission reduced film learning error by almost 20%, from 7.1% to 5.7%.The review presents transfer learning as useful when data for the new task are insufficient.Source: §IV · p. 6
  • ANN optimization of core-shell metallic nanoparticles in halide-perovskite solar cells targets optical absorption, illustrating a move from isolated nanoantennas toward arrays [53].The section then points to multilayer-sphere designs in which a DL algorithm maps target spectra to materials and layer thicknesses.Source: §IV · p. 7, Figure 3 · p. 7

V. TRANSFORMATIVE METASURFACES

The section surveys machine-learning designs for metasurfaces, where engineering and optimization support practical optical functions. Examples range from absorbers and metalenses to solar cells, beam steering, and proposed spacecraft light sails.

V. TRANSFORMATIVE METASURFACES

  • Metasurface optimization is presented as a route from established optical physics toward practical applications such as bending light, metalenses, and metaholograms.The review notes that many studies emphasize design methods over particular applications.Source: §V · pp. 7–8
  • Inverse design can represent meta-atoms by fixed-shape geometric parameters or pixelated images, with free-form designs expanding the range of possible cell geometries.Fixed-geometry gold-cross supercells were used to design narrow-band, broadband, and multiresonant absorbers.Source: §V · p. 8, Figure 4 · p. 9
  • DL-assisted metasurface designs span perfect, multiband, ultrathin, and moth-eye absorbers, switchable reflectors, chiral surfaces, beam steering, and achromatic or bifocal metalenses.The moth-eye design reached average absorption of 90% from 400–1600 nm.Source: §V · p. 8
  • DL-based metasurface design also targets structural coloration, where selective absorption or scattering can support displays, sensing, and information storage.The cited design work includes plasmonic and dielectric metastructures.Source: §V · p. 8
  • A DL-optimized silicon solar cell with an active nanopixel metasurface achieved 2.5 times the short-circuit current of cells without metasurfaces using the same crystalline silicon.Its short-circuit current was above 12 mA/cm2 for the other reported polarizations and incidence angles.Source: §V · p. 8
  • For LIDAR beam steering, ANN-generated metasurface ensembles achieved deflection angles of 45°, 55°, and 65° with absolute efficiency 0.7 in each case.The device combines a metasurface with voltage-controlled liquid crystals.Source: §V · pp. 8, 10
  • For near-eye display waveguide gratings, deflection efficiency was about 91% with integrated RGB projectors and 95% with separated projectors, with overall efficiency about 80%.The designed out-coupling grating also exceeded 85% deflection efficiency for designated diffraction efficiencies of 33%, 50%, and 100%.Source: §V · p. 10
  • Free-form silicon metastructures for a light sail converged to one-dimensional gratings, with acceleration distance D = 1.9 × 10^9 m and mean reflectivity around 0.81.The design sought optical and weight constraints for reaching 0.2 of light speed.Source: §V · p. 10

VI. CHEMICAL AND BIOLOGICAL SENSING

Metasurfaces support both sensor design and machine-learning analysis of sensor outputs, including chemical and biological classification. Examples span optimized sensing structures, biomolecule monitoring, and detection of metamaterials.

Sensing approaches

  • Metasurfaces can optimize sensor structures or support classification of measured outputs, and classical ML methods may perform as well as DL for some classification tasks.Examples include chemical recognition using plasmonic mid-infrared spectral filters or metasurfaces integrated with microfluidic channels.Source: §VI · pp. 10, 12

Colorimetric sensing

  • A dual-resonant double-bar metasurface designed with DL distinguished refractive-index changes of 0.0086, with mean square error within 0.005.The design optimized color change sensitivity to small spectral shifts.Source: §VI · p. 10

DNA sensing

  • Combined DL and classical ML improved double-negative plasmonic metasurface sensitivity for DNA oligomer detection by up to 13 times over conventional plasmonic sensors.A multilayer perceptron predicted angular reflectance, and spectral characteristics were clustered before the algorithms were applied.Source: §VI · p. 10

Biomolecule monitoring

  • DL analysis of real-time reflectance-derived absorbance spectra distinguished biomolecule analytes over time, enabling label-free monitoring of interactions including liposome perforation and cargo release [136].The platform tracked sucrose and nucleotide cargo release and subsequent formation of a supported lipid bilayer.Source: §VI · p. 12

Biological classification

  • The SARS-CoV-2 saliva sensor reached 95.2% sensitivity and specificity in clinical trials, with one false positive and one false negative among 69 samples.It also distinguished Alpha, Beta, and wild-type variants, with cumulative variance 99.7%.Source: §VI · p. 12

Metamaterial detection

  • ML classification of THz response spectra identified metamaterial in lactose mixtures with 100% accuracy, versus 87.9% for support vector machines and below 57% for human recognition.This use detects metamaterial presence rather than using a metasurface as the sensor.Source: §VI · p. 12

VII. SELF-ADAPTING METASYSTEMS

Machine learning can act as a feedback mechanism for reprogrammable metasurfaces, adjusting their responses to inputs and environmental changes. Demonstrations include adaptive cloaking, imaging, and re-trainable optical neural networks.

Feedback framework

  • For reprogrammable metasurfaces, ML can redefine programming sequences through feedback as environmental conditions change.The approach is discussed in connection with wireless communications, including 6G and internet-of-things technologies.Source: §VII · p. 12

Adaptive imaging

  • A 2-bit coding metasurface imager uses four digital responses, corresponding to phases 0, π/2, π, and 3π/2, to produce different scattering patterns for object recognition.An FPGA controls the metasurface after ML algorithms are trained on desired radiation patterns.Source: §VII · p. 12

Microwave imaging

  • In a metasurface microwave imager, an ANN reconstructs human-body images and a second ANN identifies regions such as hands or heads [169].A Gerchberg-Saxton algorithm defines coding sequences that focus radiation on target regions, and the system was demonstrated experimentally.Source: §VII · pp. 12, 14

Adaptive cloaking

  • A DL algorithm processes real-time sensed waves and calculates varactor voltages so a metasurface cloak can adjust its scattering across changing incidence angles, frequencies, and backgrounds.The engineered scattered field is made similar to that of the surrounding environment without the intruding object.Source: §VII · p. 14

Optical neural networks

  • Feedback signals and FPGA-controlled redefined coding sequences enabled a metasurface optical neural network to perform image and feature recognition, signal coding and decoding, and dynamic multi-beam focusing.The demonstrated system learned from environmental interaction without prior knowledge.Source: §VII · p. 14

Laser control

  • The review also describes ML feedback systems that automatically adjust laser-element positions and orientations, including waveplates and polarizers.Such systems may affect experimental characterization and manufacturing of metasystems.Source: §VII · p. 14

VIII. PERSPECTIVE AND OUTLOOK

The outlook surveys emerging intersections of AI with metaphotonics, including topological structures, high-Q resonances, physical neural networks, and optimization. It also notes unresolved practical challenges in fabrication and modulation.

Outlook

  • ML-assisted design is expected to help discover unconventional optical structures and advance imaging, sensing, and other metaphotonic-device functions.The review anticipates applications in smartphones, robotics, and self-driving cars using flat optics.Source: §VIII · p. 14

Topological photonics

  • AI methods have supported topological photonics through inverse design of 1D structures and prediction of topological transitions, while non-DL design has extended to 2D topological insulators and waveguides [178] [179] [187] [188].The review notes the shift toward smaller topological metadevices [189].Source: §VIII · p. 14

Bound states in the continuum

  • DL inverse design has identified photonic-crystal parameters for symmetry-protected BICs, with experiments confirming two Γ-point BICs at wavelengths 700 and 750 nm [180].The designed structure was a suspended Si3N4 photonic-crystal slab with circular holes.Source: §VIII · p. 16

Physical neural networks

  • Photonic physical neural networks offer distributed processing that maps to distributed information tasks, with an energy-efficient approach and inherently parallel computation [194] [195].Demonstrated neuron implementations include Mach–Zehnder interferometers, phase-change materials, and diffractive elements.Source: §VIII · p. 16

Biology-inspired systems

  • A retina-like array of photonic synapses forms a single-layer neural network that recognizes images with up to 90% accuracy.The review suggests intelligent metasurfaces could add reprogrammable functions or optical data processing to such systems.Source: §VIII · p. 16

High-dimensional optimization

  • Online ML optimization uses an agent that predicts parameters and updates its belief from feedback, and has optimized cold-atom laser cooling with up to 63 controllable parameters [211].The review presents metasurface design and real-time programmable systems as potential applications of this approach.Source: §VIII · p. 16

Practical challenges

  • Accuracy-critical fabrication and metasurface modulation remain challenges for practical, end-specific metaphotonics applications.The authors anticipate that emerging materials, new physics, and advanced nanofabrication may help bypass these challenges.Source: §VIII · p. 16
Loading 2110.11618v1…