Volume 14, Issue 22

Large language models (LLMs) are rapidly pushing the limits of contemporary computing hardware. For example, training GPT-3 has been estimated to consume around 1,300 MWh of electricity, and projections suggest future models may require city-scale (gigawatt) power budgets. These demands motivate exploration of computing paradigms beyond conventional von Neumann architectures. This review surveys emerging photonic hardware optimized for next-generation generative AI computing. We discuss integrated photonic neural network architectures (e.g. Mach–Zehnder interferometer meshes, lasers, wavelength-multiplexed microring-resonators) that perform ultrafast matrix operations. We also examine promising alternative neuromorphic devices and platforms, including 2D materials and hybrid spintronic–photonic synapses, which combine memory and processing. The integration of two-dimensional materials (graphene, TMDCs) into silicon photonic platforms is reviewed for tunable modulators and on-chip synaptic elements. Transformer-based LLM architectures (self-attention and feed-forward layers) are analyzed in this context, introducing the mathematical operations associated with the transformers and identifying strategies and challenges for mapping dynamic matrix multiplications onto these novel photonic hardware systems. Overall, we broadly introduce state-of-the-art photonic components, AI algorithms, and system integration methods, highlighting key advances and open issues in scaling such photonic systems to mega-sized LLM models. We find that photonic computing systems could potentially surpass electronic processors by orders of magnitude in throughput and energy efficiency, but require breakthroughs in memory especially for long-context windows and long token sequences and in storage of ultra-large datasets, among others. This survey provides a comprehensive roadmap for AI hardware development, emphasizing the role of cutting-edge photonic components and technologies in supporting future LLMs.

Optical metasurfaces, as a booming research field, have provided new methods for modulating the amplitude, phase, and polarization of light through artificial birefringent structures or structural resonances. It has been used to design planar optical components such as ultra-thin lenses, ultra-wideband achromatic lenses, and orbital angular momentum (OAM) generators. However, existing surveys typically examine either metasurface fundamentals or a single display modality, leaving no comprehensive roadmap that connects meta-atom design to full-device performance, hereafter, the term meta-atom are denoted to be an individual sub-wavelength building block of a metasurface. Here we present the first cross-scale review that quantitatively bridgices phase-dispersion engineering at the nanostructure level with system-level figures-of-merit across three mainstream 3D display paradigms, computer-generated holography, light-field projection, and near-eye/retinal displays. By critically benchmarking more than 150 demonstrations published between 2019 and 2025, we extract practical lookup charts that guide practitioners from material choice and meta-atom geometry to field-of-view, depth acuity, efficiency, and form-factor targets. Thanks to metasurfaces’ high integration density and functional diversity, its application in the light field display has attracted great interest. Metasurface can effectively improve the shortcomings of low spatial resolution, low diffraction efficiency, and narrow field of view common in traditional display components. In this paper, we first review the phase modulation method and structure resonance principle of metasurface. Then, we examine their application in the holographic display field and review the approaches for achieving structural-color printing. We summarize the 3D display methods of holographic display, light field display, and near-eye display and discuss how metasufaces enhance each modality. Finally, we distill emerging inflection points: AI assisted inverse design, dynamically tunable multifunctional platforms, and quantum or cascaded architectures into a looking forward commercialization roadmap that addresses the challenges still facing the 3D display industry.

Mid-infrared (MIR) spectroscopy plays a pivotal role in molecular identification and biosensing due to its ability to probe characteristic vibrational fingerprints of biomolecules. Plasmonic nanostructures have been explored for MIR applications but suffer from low efficiencies and broad spectral responses caused by intrinsic ohmic losses. All-dielectric metasurfaces, with low optical losses, offer an attractive alternative; however, their functionality is often restricted to reflection-mode operation. This work introduces a hybrid metal–dielectric metasurface designed to operate in transmission mode, specifically tailored for molecular identification in MIR biosensing applications. The metasurface comprises germanium (Ge) on aluminum (Al) cylinders atop a calcium fluoride (CaF 2 ) substrate, optimized to exhibits a high transmission efficiency of 80 % at a wavelength of λ = 4.2 µm, with a narrow full-width-half-maximum of 0.5 µm. By leveraging the hybridization of modes between the Ge and Al layers, the device enables precise spectral filtering. We demonstrate the potential of this metasurface for molecular identification and biosensing applications through numerical simulations and experimental validation. The straightforward fabrication process further highlights the practicality of this approach, paving the way for miniaturized MIR biosensing platforms.

The coupling between the electromagnetic field and the motional degrees of freedom of a nanoscale object offers an interesting control-knob for light–matter interaction. However, this interaction is rather weak so, in general, a precise control over the mechanical vibrations of the object becomes desirable for enhancing and fine-tuning this coupling. Phonon insulation is a base-line to achieve nanoscale vibration control by inducing constructive and destructive phonon interference, which opens phononic band gaps at various frequencies. Phonon insulators are essential for managing and directing mechanical vibrations in advanced mechanical systems, such as resonators and acousto-optic modulators. Conventional fabrication methods, such as electron-beam lithography, are highly robust and efficient but also very demanding and limited in accessibility due to their cost and technical requirements. Self-assembly offers an alternative route for fabricating mechanical structures from diverse materials, allowing for innovative designs with improved functionalities. By using natural self-organizing processes, production is simplified and accelerated, with components assembling automatically in a low-cost, scalable manner. Here, we present a simple self-assembled device that functions as a phonon insulator in the GHz frequency range and can be easily integrated onto silicon. This integration opens the door to a wide range of applications, including optical communication, quantum computing, and ultra-sensitive sensors, where efficient control over mechanical vibrations is vital.

The long-metallic-strip (LMS) array proposed by J. B. Pendry has been widely studied in electromagnetic windows. However, it could only work for the single-band window, not the dual-band one, due to the plasma-like oscillation attenuating with increasing frequency. To solve the problem, we have analyzed the condition of EM windows and then proposed to introduce “parasitic” rings (PRs) into the LMS array. The design is called the LPR metasurface, which could open the dual-broadband window at large incident angles (60°–85°). In C band, the LPR metasurface could conserve the LMSs’ plasma-like oscillation, thus opening a broadband window for transverse-electric polarization. In K band, where the plasma-like oscillation has almost disappeared, the PR generates a capacitive resonance. It could open an additional window at large angles and provide a great out-of-band suppression concurrently. Unlike related studies, the PRs here are structurally easy and integrated into the LMSs. It has conserved the LMSs’ simplicity and continuity, thus could better meet processing and protective materials. Additionally, with the great performance at large angles, the LPR metasurface may find wide applications in hypersonic aircraft radars, 5G/6G base stations, and others.

Precise control of the polarization state of light at the nanoscale is a critical and transformative technology for advancing next-generation nano-optical components. Graphene, despite its highly symmetric lattice and weak in-plane anisotropy, exhibits limited optical response to polarized light in the visible spectrum. This constraint significantly hinders its application in visible-light polarization detection and broader optoelectronic integration. In contrast, two-dimensional (2D) layered rhenium disulfide (ReS 2 ), characterized by its unique distorted 1T-phase crystal structure, exhibits strong in-plane anisotropy in the visible range. This intrinsic property not only enriches its optical characteristics but also substantially enhances the photocurrent and photoresponsivity of graphene when integrated, thereby offering new opportunities for advanced optoelectronic applications. In this study, we fabricated a vertically stacked graphene/ReS 2 van der Waals heterostructure with high precision. Polarized Raman spectroscopy was employed to systematically analyze the polarization dependence of the heterojunction, enabling an in-depth understanding of its structural and optical anisotropy. Our experimental results reveal that when isotropic single-layer graphene is coupled with anisotropic ReS 2 , the resulting heterostructure exhibits anomalous polarization-dependent Raman scattering. This finding highlights the potential of ultrathin 2D heterostructures as nanoscale polarization-sensitive optical elements. This work not only provides a novel approach for designing and implementing next-generation polarization-resolved optical devices based on 2D anisotropic materials but also lays the groundwork for their integration into highly miniaturized and lightweight optoelectronic systems. These findings hold significant promise for advancing the field of optoelectronics and enabling the development of more sophisticated and efficient optical technologies.

This work designs a model named POST based on the vision transformer (ViT) approach. Across single, double, and even triple lattices, as well as various non-circular complex hole structures, POST enables prediction of multiple optical properties of photonic crystal layers in photonic crystal surface emitting lasers (PCSELs) with high speed and accuracy, without requiring manual intervention, which serves as a comprehensive surrogate for the optical field simulation. In the predictions of quality factor ( Q ) and surface-emitting efficiency (SE) for PCSEL, the R-squared values reach 0.909 and 0.779, respectively. Additionally, it achieves nearly 5,000 predictions per second, significantly lowering simulation costs. The precision and speed of POST predictions lay a solid foundation for future ultra-complex model parameter tuning involving dozens of parameters. It can also swiftly meet designers’ ad-hoc requirements for evaluating photonic crystal properties. The database used for training the POST model is derived from predictions of different photonic crystal structures using the coupled-wave theory (CWT) model. This dataset will be made publicly available to foster interdisciplinary research advancements in materials science and computer science.

Nonreciprocal devices are key technologies for modern photonic applications, and nonlinearly induced nonreciprocity based on metasurface platforms opens up the potential for compactness and miniaturization of such devices in free-space optical paths. In this work, a design of quasi-bound states in the continuum (QBICs) via scaled lattice constants is employed, featuring a plate-hole tetramer unit cell with single-step etching. Under normal incidence, QBICs exhibit polarization-insensitive excitation, with the scattering dominated by the combination term of magnetic dipoles, magnetic toroidal dipoles, electric quadrupoles as well as electric toroidal quadrupoles – rendering the vertical field distribution sensitive to parametric variations. Embedding Kerr nonlinearity yields nonreciprocal responses under upper and lower port excitation. The steep edges between peaks and valleys in the Fano-line-shaped transmission spectra facilitate balancing the nonreciprocal intensity range (NRIR) and isolation. The NRIR originates from electromagnetic asymmetry introduced by varying the substrate refractive index and can be flexibly tuned via structural parameters – validated by simulations of nonreciprocal responses with varying plate thickness alone. Additionally, Rabi splitting from varying hole depth induces abrupt electromagnetic asymmetry changes in the strong coupling region, offering a new NRIR tuning freedom. This design strategy provides fresh insights for nonreciprocal device research, with the structure holding promise for sensing and nonlinear applications.

Traditional metasurface design is limited by the computational cost of full-wave simulations, preventing thorough exploration of complex configurations. Data-driven approaches have emerged as a solution to this bottleneck, replacing costly simulations with rapid neural network evaluations and enabling near-instant design for meta-atoms. Despite advances, implementing a new optical function still requires building and training a task-specific network, along with exhaustive searches for suitable architectures and hyperparameters. Pre-trained large language models (LLMs), by contrast, sidestep this laborious process with a simple fine-tuning technique. However, applying LLMs to the design of nanophotonic devices, particularly for arbitrarily shaped metasurfaces, is still in its early stages; as such tasks often require graphical networks. Here, we show that an LLM, fed with descriptive inputs of arbitrarily shaped metasurface geometries, can learn the physical relationships needed for spectral prediction and inverse design. We further benchmarked a range of open-weight LLMs and identified relationships between accuracy and model size at the billion-parameter level. We demonstrated that 1-D token-wise LLMs provide a practical tool for designing 2-D arbitrarily shaped metasurfaces. Linking natural-language interaction to electromagnetic modelling, this “chat-to-chip” workflow represents a step toward more user-friendly data-driven nanophotonics.

Subwavelength grating-based color-shifting devices exhibit dynamically tunable spectral responses under specific resonant conditions, offering advanced applications in optical anticounterfeiting and surface decoration. However, conventional devices remain limited by structural complexity, narrow tuning ranges, low energy efficiency, and high manufacturing costs. Thereby, we propose a design methodology for multicolor shifting devices capable of dynamic color shifts through azimuthal angle rotation (i.e., in-plane rotation of the device). The proposed evaluation function can be adjusted to yield the desired spectral response, enabling diverse color shifts from a single template. Optimization of the geometric parameters and azimuth angles of rectangular gratings using rigorous coupled-wave analysis (RCWA) and an immune algorithm, followed by conversion into a sinusoidal structure to simplify large-area fabrication while maintaining performance. Using this approach, quad- and penta-color shifting devices based on subwavelength sinusoidal gratings were designed, achieving peak reflectance exceeding 45 % for all target colors, with maxima reaching 89 %. Additionally, the quad-color shifting device was fabricated using low-cost exposed holographic interferometry, validating the method. The results simplify the structure compared to conventional subwavelength grating filters, while enhancing optical performance and fabrication feasibility, offering a new approach for cost-effective, high-performance dynamic optical devices.

Transparent metafilms with spectrally selective properties have shown great potential in energy-efficient window systems. Most previous studies focused on optimizing materials and thicknesses to enhance visible transmittance and near-infrared (NIR) reflectance. However, few have considered how the position of the metafilms on the glass affects overall optical and thermal performance, especially in the mid-infrared (MIR) range critical for radiative cooling. In this work, we propose and analyze a five-layer TiO 2 /Ag/TiO 2 /Ag/TiO 2 structure and systematically evaluate its performance under two typical installation scenarios. Numerical simulations based on the transfer matrix method show that both configurations maintain a high visible transmittance (∼0.88) and an effective NIR reflectance (∼0.98). Notably, a substantial difference is observed within the atmospheric transparency window 8–14 μm, where the interior-coated configuration possesses a high emissivity of 0.8. This value significantly exceeds the average emissivity of 0.01 found for the exterior-coated configuration, thereby resulting in superior passive radiative cooling capability. Moreover, we also compared the net radiative cooling power under the two configurations. These findings reveal that the position of the transparent metafilms critically influences MIR radiation. Coating placement on the interior surface not only maintains favorable solar modulation but also markedly enhances the thermal dissipation. This study offers theoretical guidance and practical insight into the design and implementation of metafilms in energy saving window systems aimed at reducing energy consumption, especially in regions with hot climates.

Electromagnetic devices with multiple polarization modes are urgently needed in remote sensing detection and radar imaging due to their ability to obtain scattering information from targets through manipulation of full-parametric Jones matrix components ( J xx , J xy , J yx , J yy ). Although metasurfaces exhibit exceptional capability for polarization control, they typically facilitate conversion between specific linearly polarized (LP) and circularly polarized waves on Poincaré sphere. Here, we find that identical phases in two co-polarized components is exactly half of the sum of phases in two cross-polarized components by deriving Jones matrix J AAL with full-parametric Aharonov–Anandan (AA) phase Jones matrix. On this basis, a novel spin-decoupled paradigm is proposed by merging of AA phase and Pancharatnam–Berry (PB) phase mechanisms. Such a paradigm in diatomic metasurface is promised to achieve elegant amplitude-phase controlling and generate arbitrary polarized waves. For verification, two types of meta-devices were designed, fabricated, and experimentally characterized. Compared to the combination of propagation and PB phase, the proposed method enables simultaneous broadband arbitrary LP-to-LP conversion and wavefront control with a relative bandwidth of 43.5 %. Our strategy establishes theoretical foundation for spin-decoupled phase manipulation and amplitude-phase control of AA phase, providing a solid platform and guidance for the design of devices with arbitrary polarization and wavefront control.

We demonstrate a label-free, far-field super-resolution imaging approach based on photonic nanojets generated by tapered dielectric fibers. By systematically analyzing the dependence of nanojet confinement and focal distance on cylinder diameter (8–16 μm), we establish a geometric design framework for tunable light localization below the diffraction limit. Using this insight, we fabricate a 12-μm waist-tapered optical fiber that produces a laterally extended nanojet for non-contact imaging. This configuration resolves grating lines with 92 nm width and spacing – dimensions beyond the classical resolution limit. Ray tracing simulations confirm the experimental magnification trend and show that fiber tilt enables tunable control over magnification and field of view. Our fiber platform provides scalable alignment, mechanical tunability, and extended working distances. These findings establish tapered fibers as compact and flexible photonic elements for delivering sub-wavelength light confinement, with applications in optical metrology, field enhancement, and scanning nanophotonic systems.

Symmetry-protected bound states in the continuum (SP-BIC) enable flexible tuning of the wavelengths and linewidths of high-Q resonances, showing great potential in high-performance photonic devices. However, the implementation of SP-BIC at shorter wavelengths, such as in the visible range, requires precise control of smaller feature sizes, which imposes stringent fabrication requirements. This trade-off between geometric accuracy and manufacturing complexity limits the practical application of SP-BIC metasurfaces. This work presents a defect-insensitive design strategy for SP-BIC metasurfaces by constructing position-detuned arrays composed of periodically arranged antisymmetric trapezoidal unit cells. Theoretical studies show that under x-polarized incidence, the dominant resonance mode is a magnetic dipole, with electric fields strongly enhanced in the inter-pillar gap rather than inside the structures body. This field distribution provides high robustness against geometric deformation. The geometric dependence of this mode under x-polarized light is approximately 1/10 to 1/20 that of the electric dipole-dominated mode under y-polarized light, and the design is demonstrating strong tolerance to typical fabrication defects encountered in nanofabrication. Experimental validation is being conducted using metasurfaces fabricated with intentional defects. Overall, this work is offering a practical, fabrication-tolerant approach to realizing high-performance dielectric BIC metasurfaces.