IOPT 1 (2025) 100004

Contents lists available at ScienceDirect

iOptics
journal homepage: [Link]/en/journals/ioptics

AI-empowered optical metasurfaces: From unit-cell optimization to

system-level integration
Yunhui Zeng a,b, Xin Jin a,*
a
Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, PR China
b
Peng Cheng Laboratory, Shenzhen 518055, PR China

A R T I C L E I N F O A B S T R A C T

Keywords: Optical metasurfaces, composed of subwavelength structures, represent an emerging class of optical devices
Artificial intelligence capable of precisely manipulating electromagnetic wavefronts, enabling unprecedented miniaturization and
End-to-end optimization planarization in optical systems. However, critical challenges remain across different stages from unit-cell design
Metasurface imaging
to system-level integration: at the unit-cell level, high computational costs, complex nonlinear design spaces, and
Optical metasurfaces
System-level integration
sensitivity to fabrication errors impede rapid and robust design; at the metasurface level, precise modeling of
Unit-cell design inter-unit coupling, balancing multiple conflicting objectives, and achieving efficient real-time dynamic control
remain difficult; at the system level, integration complexity, susceptibility to hardware noise and environmental
variability, and incompatibility between optical design and backend reconstruction algorithms significantly limit
overall performance. Recent advances in artificial intelligence (AI) have provided effective solutions to these
issues. Specifically, AI techniques substantially reduce computational overhead, navigate complex nonlinear
design spaces, and enhance robustness against manufacturing imperfections at the unit-cell stage; accurately
model inter-unit interactions, facilitate efficient multi-objective optimization, and enable real-time dynamic
metasurface reconfiguration; and at the system level, integrate optical design and backend reconstruction
through end-to-end optimization frameworks, effectively mitigating environmental variability, hardware noise,
and design-reconstruction incompatibility. This review systematically summarizes recent progress in AI-driven
optical metasurface research, highlighting a clear trend from isolated component optimization toward inte­
grated, end-to-end system-level optimization.
Data availability: All data are available in the main text or the supplementary materials from the corresponding
author upon reasonable request.

1. Introduction [19,20]. Moreover, recent innovations such as meta-tweezers for optical

manipulation [21], metavehicles—micro-scale platforms capable of
Optical metasurfaces, composed of two-dimensional arrays of transporting objects via controlled optical forces [22], and novel hybrid
subwavelength-scale unit cells (meta-atoms), have emerged as powerful integrations combining metasurfaces with heterostructures on fiber
platforms capable of precisely controlling electromagnetic wavefronts in platforms [23], highlighting the rapid and ongoing expansion of meta­
unprecedented ways beyond the capabilities of traditional optical surface research frontiers. Compared to conventional refractive and
components [1,2]. By engineering the geometry, arrangement, and diffractive optical elements, metasurfaces exhibit distinct advantages
material properties of meta-atoms, metasurfaces can manipulate such as planarization, ultra-thin profiles, ease of integration, multi­
amplitude, phase, polarization, and spectral characteristics of incident functionality, and compatibility with complementary
light at the subwavelength scale. This extraordinary capability has led to metal-oxide-semiconductor (CMOS) fabrication processes, marking a
transformative advances across numerous photonics-related fields, paradigm shift toward compact and versatile optical systems [24,25].
including imaging systems [3–5], sensing [6–8], optical communica­ Despite substantial progress, the development of optical meta­
tions [9,10], holography [11,12], beam steering [13,14], polarization surfaces still faces major challenges that span multiple design and
manipulation [15,16], spectral filtering [17,18], and quantum optics integration scales. The metasurface development pipeline can be

* Corresponding author.
E-mail address: [Link]@[Link] (X. Jin).

[Link]
Received 20 May 2025; Received in revised form 26 June 2025; Accepted 2 July 2025
Available online 8 July 2025
3051-1771/© 2025 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC
BY-NC-ND license ([Link]
Y. Zeng and X. Jin iOptics 1 (2025) 100004

divided into three stages: unit-cell optimization, metasurface-level efficiency, broader operational bandwidths, enhanced dimensionality,
design optimization, and system-level integration. At the unit-cell and more integrated optical metasurface systems. Accompanying the
level, optimization faces significant challenges, including high compu­ rise of AI methodologies, metasurface research has undergone a signif­
tational costs from full-wave simulations, the complexity of exploring icant paradigm shift, from isolated unit-cell design optimization [45], to
vast and nonlinear parameter spaces, and sensitivity to fabrication im­ global metasurface configuration [46], dynamic response modulation
perfections [4,26]. At the metasurface level, challenges include [47], and backend computational reconstruction [48]. More recently,
inter-unit coupling effects [27,28], which degrade performance if not efforts have converged toward unified frameworks that integrate
properly considered, the complexity of achieving multiple conflicting structural design, optical wave manipulation, and computational
optical functionalities, and difficulties in enabling real-time dynamic post-processing within end-to-end AI-driven optimization pipelines
reconfigurability. Finally, at the system level, complexities stem from [49]. This evolution illustrates a transition from component-level design
integrating metasurfaces with electronic hardware, sensitivity to envi­ toward intelligent, collaborative, and system-level optimization,
ronmental disturbances and noise, and incompatibility between meta­ marking a new phase in the development of AI-empowered optical
surface design and back-end computational algorithms [29,30]. These metasurfaces.
multi-scale challenges collectively limit the widespread implementa­ In this review, we aim to summarize the recent progress of AI-
tion and practical deployment of advanced metasurface-based optical empowered optical metasurfaces, clearly delineating developments at
systems. the unit-cell optimization, metasurface-level design optimization, and
Artificial intelligence (AI), especially deep learning, has achieved comprehensive system-level integration, as shown in Fig. 3. We high­
transformative success in modeling complex nonlinear systems, accel­ light how AI methodologies address critical challenges at each stage,
erating inverse problem-solving, and extracting patterns from high- including reducing computational complexity, navigating nonlinear and
dimensional datasets. These capabilities have revolutionized various high-dimensional design spaces, modeling inter-unit interactions,
scientific fields, including computer vision, natural language processing, enabling real-time dynamic control, and resolving incompatibility be­
and materials science, by improving computational efficiency and tween metasurface design and backend reconstruction algorithms
enabling data-driven discovery. Inspired by these achievements, re­ through end-to-end optimization frameworks. Furthermore, we analyze
searchers are increasingly applying AI-driven methods to tackle ongoing the existing bottlenecks and technical challenges within current AI-
challenges in metasurface design optimization [26,31–34]. To quanti­ driven approaches, providing forward-looking insights and clear per­
tatively assess the evolving research landscape, we collected publication spectives for future research directions. This review delineates the
data from the Web of Science database covering three core themes: methodological evolution from isolated component-level design toward
metasurfaces, AI, and their intersection. As illustrated in Fig. 1, the holistic, AI-driven system-level metasurface optimization, offering a
overall number of publications in these areas has grown substantially perspective on how AI can be systematically embedded throughout the
over the past decade. Notably, research at the intersection of AI and entire optical metasurface optimization pipeline.
metasurface design has witnessed an accelerated surge since 2018,
indicating the rapid emergence of this interdisciplinary field and its 2. AI for unit-cell optimization
increasing academic impact.
To systematically characterize this integration and the progressive As the fundamental building block of metasurfaces, unit-cell design
development of metasurface technologies, we have extensively analyzed directly determines the device’s electromagnetic functionality. Howev­
seminal studies published between 2010 and 2024, as illustrated in er, this process is often hindered by high-dimensional parameter spaces,
Fig. 2. Initially proposed in 2011, optical metasurfaces based on sub­ nonlinear and non-unique structure–response mappings, and sensitivity
wavelength structures enabled the generalized manipulation of optical to fabrication imperfections. Recent advances in AI have introduced a
wavefronts [35], marking a fundamental breakthrough in optics. Since powerful toolkit to overcome these limitations: surrogate models
then, the field has experienced a rapid evolution, transitioning from accelerate forward prediction of electromagnetic responses; inverse
lossy plasmonic metallic metasurfaces [36] towards dielectric materials design frameworks leverage generative and differentiable models to
exhibiting significantly lower optical losses. Concurrently, metasurface explore complex solution spaces; and robustness-aware methods incor­
functionalities have evolved from initial single-wavelength wavefront porate real-world constraints such as fabrication tolerances and struc­
[37] control toward broadband achromatic imaging [38], multidimen­ tural uncertainty. This section reviews these three complementary
sional wavefront manipulation involving complex optical fields [39–41] directions and highlights how AI is reshaping the landscape of unit-cell
and hyperspectral imaging [42], and most recently to comprehensive optimization.
Mueller-matrix polarimetric imaging [43,44]. Collectively, these ad­
vancements indicate a clear developmental trajectory toward higher 2.1. Surrogate modeling

In optical metasurface design, simulating unit-cell electromagnetic

responses typically relies on full-wave solvers such as the finite-
difference time-domain (FDTD) method [50] and rigorous
coupled-wave analysis (RCWA) [51]. While these physics-based
methods offer high accuracy, they become prohibitively
time-consuming when dealing with high-dimensional geometric pa­
rameters, broadband spectral ranges, and multiple polarization states­
—leading to an exponential growth in computational cost [52,53]. To
address this challenge, researchers have proposed AI-based surrogate
models that replace the direct numerical solution of Maxwell’s equa­
tions with data-driven approximations [54–57]. By training on pre­
computed simulation data, these models can rapidly predict
electromagnetic responses in milliseconds during inference, reducing
computational burden while preserving accuracy. Beyond speed, sur­
rogate modeling also enables differentiable approximations that provide
Fig. 1. Growth of publications in AI, metasurfaces, and AI-empowered meta­ explicit gradients with respect to design parameters—facilitating effi­
surfaces in Web of Science (2010–2024). cient gradient-based optimization—and can be extended to directly

2
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 2. Key milestones in the development of metasurfaces, and AI-empowered metasurfaces. This timeline highlights the evolution of metasurfaces from unidi­
mensional to multidimensional, and the shift in AI-empowered metasurfaces from unit-cell optimization to system-level integration.

Fig. 3. System architecture of AI-empowered metasurface design. This diagram represents the integration of AI methods into the metasurface design process, from
unit-cell optimization to system-level integration, addressing key challenges and AI approaches.

estimate application-specific performance metrics, such as imaging fi­ and accurate evaluation of the S-parameters of meta-atoms, given their
delity or energy efficiency. In the following, we review three represen­ dimensions and crystallization state. This approach reduces the optical
tative categories of surrogate modeling techniques: (i) fast-response performance prediction time of meta-surface to milliseconds, as shown
predictors for electromagnetic fields, (ii) differentiable models that in Fig. 4(A). Building on the reduced prediction time, a surrogate model
support gradient-driven optimization, and (iii) performance-oriented based on DNN can accelerate and optimize the inverse design process in
surrogates that link structural parameters to system-level figures of nanophotonics. For instance, Han et al. [58] developed a DNN surrogate
merit (FOMs). model that enables rapid prediction of electromagnetic responses from
The first major advantage of surrogate modeling is its computational metasurface structure parameters. This model was utilized to guide in­
speed. By approximating the solution process of Maxwell’s equations verse design, achieving a significant reduction in computational cost
through data-driven learning, surrogate models drastically reduce the while maintaining high prediction accuracy and enhancing design effi­
runtime required by traditional numerical solvers while maintaining a ciency compared to conventional methods, as shown in Fig. 4(B).
high level of accuracy [27,58,59]. For example, An et al. [27] proposed a On the other hand, the automatic differentiation capabilities
forward-prediction deep neural network (DNN) surrogate model based inherent in modern deep learning frameworks enable surrogate models
on a convolutional neural network (CNN) architecture to achieve rapid to go beyond fast prediction of electromagnetic responses. They can also

3
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 4. Representative surrogate-modeling strategies for metasurface unit-cell design. (A) Convolutional neural network predictor for generating the S-matrix (real
and imaginary parts) of meta-atoms across the 60–100 THz spectrum [27]. (B) Unit-cell inverse design framework using a forward surrogate model based on deep
neural networks [58]. (C) Statistical machine-learning model establishing a differentiable link between meta-atom structures and optical responses for multifunc­
tional metasurface optimization [61]. (D) Multi-scale differentiable framework using a DNN-based surrogate model to map meta-atom geometry to optical per­
formance, enabling gradient-driven optimization of broadband long-wave infrared meta-optics [60]. (E) Backpropagation neural network predictor for mapping
meta-atom geometry to phase and group delay, enabling achromatic metalens design [62]. (F) Inverse design framework combining adjoint optimization and
machine learning to predict and optimize device FOM [63].

provide differentiable gradients with respect to structural parameters differentiable framework for broadband long-wave infrared meta-optics,
[60,61]. Traditional surrogate models are efficient and accurate in employing a DNN-based surrogate model to map meta-atom geometry to
predicting metasurface responses, but they typically lack explicit gra­ optical performance. This allows for gradient-driven optimization of
dients, which hinders their use in optimization. As a result, optimization meta-optic parameters, achieving a six-fold improvement in average
often relies on black-box, derivative-free algorithms such as genetic al­ Strehl ratio over conventional designs, as shown in Fig. 4(D).
gorithms or particle swarm optimization. However, these methods are Beyond conventional electromagnetic response prediction, surrogate
generally less efficient and exhibit slower convergence. In contrast, models can be further extended to directly forecast system-level per­
differentiable surrogate models can simultaneously predict electro­ formance metrics or composite FOMs that are closely tied to practical
magnetic responses and compute precise gradients of performance applications [62,63]. In many real-world scenarios, the ultimate objec­
metrics, enabling more efficient optimization. By leveraging this tive of metasurface design is not merely to achieve a desired electro­
differentiability, the design process becomes more efficient and precise, magnetic response, but to fulfill specific system-level
with accelerated convergence in optimization. For example, Ma et al. requirements—such as imaging resolution, communication
[61] proposed a statistical machine learning framework, establishing a signal-to-noise ratio, or sensing sensitivity. To address this, a more
differentiable link between meta-atom structures and optical responses advanced class of surrogate models has been proposed, which directly
to enable end-to-end optimization of multifunctional metasurfaces. This learns the mapping between structural parameters of metasurfaces and
approach leverages gradient-based optimization, offering higher design application-specific performance metrics. These models enable the
precision and efficiency compared to traditional non-gradient methods, design process to incorporate complex constraints and performance re­
as shown in Fig. 4(C). Huang et al. [60]presented a multiscale quirements of the target application at an early stage, thereby promoting

4
Y. Zeng and X. Jin iOptics 1 (2025) 100004

an application-driven paradigm in metasurface design. Wang et al. [62] First, latent space compression techniques have been proposed to
proposed a backpropagation neural network to establish a mapping from significantly reduce the design dimensionality. Representative ap­
meta-atom geometric parameters to phase and group delay, which are proaches employ variational autoencoders (VAEs) to embed high-
directly linked to system-level performance indicators such as achro­ resolution geometries into a low-dimensional latent vector space,
matic focusing and imaging capabilities, as shown in Fig. 4(E). This where interpolation and search can be performed more efficiently [79,
approach enables a direct link between metaatom parameters and 80]. This encoding–decoding pipeline not only preserves the physical
system-level performance metrics. Yeung et al. [63] presented an in­ validity of structures but also allows fast convergence of optimization
verse design framework that combines adjoint optimization and ma­ algorithms—such as gradient descent and Bayesian opti­
chine learning to create a mapping from photonic device structures to mization—within a compact latent space of merely tens of dimensions.
system-level performance metrics like power transmission efficiency. As a result, both computational complexity and data requirements are
Their method trains a neural network using optimization results as substantially reduced, while maintaining high spectral accuracy. For
training data, where the network takes structural images as inputs and instance, Ma et al. [79] proposed a deep generative model with a
predicts device FOM and target wavelengths as outputs, as shown in semi-supervised learning strategy, implementing latent space compres­
Fig. 4(F). By leveraging explainable AI algorithms, this framework not sion to map high-resolution meta-atom geometries to low-dimensional
only optimizes device performance but also helps escape local minima in vectors, enabling efficient inverse design with reduced computational
the optimization process. complexity, as shown in Fig. 5(A). Similarly, Kudyshev et al. [80]
In summary, surrogate modeling has become an essential enabler for introduced an adversarial autoencoder framework, utilizing latent space
efficient metasurface design, offering rapid forward prediction, differ­ compression to represent complex nanophotonic designs in a
entiable optimization, and direct estimation of application-level per­ low-dimensional space, achieving rapid optimization and significant
formance metrics. Despite these capabilities, current approaches are still reduction in design iteration time, as shown in Fig. 5(B).
fundamentally approximation-based, and several critical challenges In inverse metasurface design, one of the fundamental challenges lies
remain unresolved. First, existing models typically approximate the in the inherently ill-posed nature of the problem, where multiple
electromagnetic response by regressing from structure to response geometrically distinct meta-atom configurations can yield near-identical
without explicitly encoding the governing physics. This limits their electromagnetic responses. This one-to-many mapping complicates the
generalization and interpretability. A promising direction is to develop training of deterministic networks and limits design diversity [81,82].
physics-consistent, differentiable computational solvers that directly To overcome this issue, generative deep learning models have emerged
accelerate Maxwell-based simulations while providing analytical gra­ as powerful tools, capable of learning rich, probabilistic representations
dients, thus serving as a physically faithful and optimization-ready that span diverse geometric solutions while preserving spectral fidelity
backend for inverse design workflows, by leveraging advanced numer­ [44,46,83–85]. For example, Yang et al. [81] proposed a Tandem
ical techniques [64–67]. Second, most surrogate models are trained on Generative Network (TGN) that combines a probabilistic generative
large precomputed datasets, which are costly to generate and prior with a forward-predictive network to enable high-precision and
domain-specific. Future work should focus on hybrid modeling frame­ high-speed generation of free-form meta-atoms under complex electro­
works that combine data-driven learning with physical priors [68,69], magnetic constraints. TGN achieved a speed-up by a factor of 2990
transfer learning from low fidelity simulations [70,71], or active relative to diffusion-based methods while maintaining a mean absolute
learning guided by uncertainty metrics—to reduce data requirements error below 4 %, as shown in Fig. 5(C). In parallel, Zeng et al. [86]
[59,72] while preserving accuracy and model robustness. Third, surro­ proposed the Anchor-controlled Generative Adversarial Network
gate models are increasingly used in inverse design tasks, yet their (AcGAN), which enhances both structural diversity and electromagnetic
intrinsic prediction errors are rarely considered during optimization. fidelity by incorporating two key mechanisms: an anchor-guided phys­
Incorporating uncertainty quantification into the design loop—by ical constraint module that enforces spectral consistency throughout the
modeling predictive confidence intervals or constructing probabilistic generation process, and a clustering-based controller that facilitates
surrogates—can significantly improve the reliability and reproducibility structured exploration across multiple valid geometric configurations,
of optimized designs, especially under manufacturing or environmental as shown in Fig. 5(D). Together, these methods exemplify how AI-based
variability. Finally, with the growing demand for multiplexed meta­ generative frameworks can fundamentally reshape the inverse design
surfaces capable of manipulating light across multiple degrees of landscape by offering scalable, diverse, and high-fidelity design solu­
freedom [73,74]—such as wavelength, polarization, phase, and spatial tions beyond the reach of conventional approaches.
distribution—future surrogate models should be extended to simulta­ While AI-based models have demonstrated strong capabilities in
neously predict high-dimensional, coupled response spaces. This re­ addressing high-dimensional, nonlinear, and one-to-many mapping
quires not only scalable architectures but also new training strategies challenges in unit-cell optimization, purely data-driven approaches
that can handle multimodal outputs and interdependent physical con­ often suffer from heavy dependence on large-scale labeled simulation
straints. Together, these directions will establish surrogate modeling as a datasets and limited physical interpretability. To overcome these limi­
physically grounded, data-efficient, uncertainty-aware, and tations, physics-informed neural networks (PINNs) have emerged as a
high-dimensional foundation for next-generation metasurface design. promising alternative by embedding electromagnetic priors, such as
Maxwell’s equations, directly into the training process [34,46,68,69,
2.2. AI-based unit-cell inverse design 87]. This enables accurate modeling of optical responses even with
scarce data, while preserving physical consistency and differentiability,
In the unit-cell design of optical metasurfaces, inverse design faces which are both crucial for tackling inverse problems and inaccessible
two fundamental challenges: (1) the extremely high dimensionality of gradients. As shown in Fig. 5(E), Jiang and Fan [46] proposed a
geometric and material parameters, resulting in a vast and complex physics-driven conditional generative network for the global optimiza­
design space; and (2) the highly nonlinear and one-to-many mapping tion of dielectric metasurfaces, leveraging adjoint electromagnetic gra­
from a target electromagnetic response to multiple feasible structural dients to guide the learning of topology distributions without relying on
configurations. Traditional optimization approaches—whether heuristic precomputed datasets. Their method achieves global exploration across
[75,76] or gradient-based [77,78]—often suffer from the curse of wavelength and angle parameters with high efficiency and significantly
dimensionality, slow convergence, and entrapment in local optima reduced computational cost. Similarly, Medvedev et al. [87] developed a
within such complex design domains. AI-driven inverse design offers a PINN-based framework for 3D metasurface diffraction modeling, where
systematic solution framework by integrating data-driven and Maxwell’s vector equations, Floquet-Bloch boundary conditions, and
physics-informed strategies. PML absorption are encoded directly into the loss function, as shown in

5
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 5. AI-driven metasurface unit-cell inverse design strategies. (A) Deep generative model with semi-supervised learning for design space compression in meta­
material inverse design [79]. (B) Machine-learning-assisted topology optimization for thermal emitters, emphasizing design space compression through deep learning
[80]. (C) TGN addressing one-to-many mapping for high-precision meta-atom design [81]. (D) AcGAN enhancing electromagnetic fidelity and structural diversity in
metasurface design [86]. (E) Physics-informed neural network for dielectric metasurface optimization, embedding Maxwell’s equations into the training process [46].
(F) Physics-informed deep learning for 3D modeling of light diffraction from optical metasurfaces [87].

Fig. 5(F). This enables high-fidelity prediction of polarization and formulations. Third, existing models tend to generalize poorly across
far-field responses within milliseconds, eliminating the need for structural domains, especially when encountering previously unseen
repeated numerical simulations and demonstrating exceptional scal­ topologies or geometries. Robust inverse design frameworks should
ability across input configurations. Collectively, these works show that incorporate structural priors, adaptive normalization schemes, or
physics-informed AI can balance data efficiency, physical rigor, and localized data augmentation to improve stability under distributional
optimization performance in metasurface design. shifts. Advancing along these directions will further enhance the
In summary, recent advances in latent space compression, generative controllability, reliability, and generalization capacity of AI-driven unit-
modeling, and physics-informed design have collectively shaped a cell inverse design, facilitating its broader adoption in practical meta­
promising AI-based inverse design paradigm for metasurface unit cells. surface engineering.
While these methods have demonstrated clear advantages in acceler­
ating design and enhancing structural diversity, several key challenges 2.3. AI-based robust design
remain. First, current latent representations often lack explicit geo­
metric interpretability, limiting the ability to control or constrain Despite the success of AI-based frameworks in rapidly generating
generated structures. Future work should explore the development of high-performance metasurface structures, these designs often suffer
semantically meaningful and geometry-aware latent spaces that support from significant performance degradation under real-world perturba­
fine-grained structural manipulation aligned with optical objectives. tions such as fabrication tolerances, material variability, or incident-
Second, generative models typically produce multiple candidate solu­ angle deviations. Standard neural networks tend to optimize for nomi­
tions, yet lack effective mechanisms for solution quality control or nal performance, neglecting the stability of designs under realistic, non-
preference alignment. Addressing this issue requires the integration of ideal conditions. A common workaround is to introduce robustness-
structure-performance co-optimization strategies, such as feedback- enhancing procedures in the post-design stage, though these strategies
guided sampling, discriminative filtering, or diversity-aware loss typically incur trade-offs. As shown in Fig. 6(A), Yeung et al. [88]

6
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 6. AI-driven unit-cell robust design strategies for metasurfaces. (A) Global inverse design across multiple photonic structure classes using generative deep
learning [88]. (B) Dual-band optical collimator based on deep-learning designed, fabrication-friendly metasurfaces [89]. (C) Establishing exhaustive metasurface
robustness against fabrication uncertainties through deep learning [90]. (D) Uncertainty qualification for metasurface design with amendatory Bayesian network
[91]. (E) Velocity-adaptive electrohydrodynamic printing for microscale conformal circuits on freeform curved surfaces [92]. (F) Data-driven concurrent nano­
structure optimization based on conditional generative adversarial networks [93].

applied a Gaussian filter to the generated structures after training a offering generalized, fabrication-aware solutions that drive the shift
generative adversarial network across structural and material domains. from idealized to manufacturable metasurface designs. For instance, as
While this operation increases fabrication compatibility, it also com­ shown in Fig. 6(C), Jenkins et al. [90] proposed a deep learning
promises optical performance by smoothing out fine design features. framework that not only quantifies the impact of fabrication errors but
Alternatively, as illustrated in Fig. 6(B), Ueno et al. [89] proposed a also dynamically reshapes the design objective to favor Pareto-optimal
selector-based approach, where a neural network generates a large set of structures with built-in resilience, such as smooth transitions and
candidate meta-atoms, and only those satisfying fabrication constraints non-resonant regions, across large structural-uncertainty spaces. Simi­
are retained. This brute-force filtering ensures manufacturability but at larly, Fig. 6(D) illustrates how Zhang et al. [91] introduced an amen­
the cost of efficiency, requiring substantial computational overhead to datory Bayesian network that quantifies both model and data
produce a viable design subset. These studies underscore the lack of uncertainty by embedding topologically distorted patterns into training
intrinsic robustness modeling in existing AI pipelines, and highlight the and estimating predictive confidence alongside electromagnetic
need for integrated approaches that embed fabrication and perturbation response. This provides a statistical reliability map for device selection.
awareness directly into the learning process. Further extending this idea from robustness modeling to process-driven
Beyond reactive post-processing and brute-force filtering, recent design, Yue et al. [92] in Fig. 6(E) employed AI-guided velocity-adaptive
advances in AI-based robust design have demonstrated intrinsic capa­ electrohydrodynamic printing for freeform metasurface fabrication,
bilities that are difficult to achieve with conventional optimization where machine learning algorithms control print speed, nozzle position,
methods. Deep learning models are naturally well-suited to handling and pattern fidelity under real curvature constraints. Finally, Baucour
high-dimensional, nonlinear design spaces, modeling uncertainty, and et al. [93] in Fig. 6(F) integrated conditional generative adversarial
capturing implicit physical constraints. Leveraging generative archi­ networks into an iterative design loop to align structural generation with
tectures, multi-task training, and probabilistic modeling, these ap­ real fabrication process parameters, mapping directly from process
proaches can learn data-driven mappings from process variations to space (e.g., deposition rate, annealing time) to physically realizable
device performance distributions, enabling joint optimization of effi­ nanostructures. Together, these efforts highlight a paradigm shift in
ciency and robustness. Instead of relying on predefined tolerance robust design, shifting from discrete tolerance handling to continuous,
models, these methods learn statistical behaviors under perturbations, data-driven, and fabrication-informed optimization embedded within

7
Y. Zeng and X. Jin iOptics 1 (2025) 100004

the AI design pipeline. 3. AI for metasurface design

In summary, AI has emerged as a versatile enabler for robust meta­
surface design, supporting manufacturability-aware filtering, suppres­ Beyond unit-cell optimization, the design of full metasurfaces pre­
sion of performance sensitivity to fabrication perturbations, uncertainty sents a new layer of complexity. As devices scale in aperture, function,
quantification, and real-time process-informed optimization. These ca­ and adaptivity, key challenges emerge across three dimensions: (1)
pabilities collectively address the long-standing trade-off between modeling near-field and long-range interactions between densely
structural complexity and fabrication feasibility. Moving forward, future packed meta-atoms and multilayer structures; (2) optimizing for mul­
research should prioritize the systematic integration of process con­ tiple, often conflicting, performance objectives spanning electromag­
straints into design models—by formally encoding lithographic limits, netic, mechanical, and thermal domains; and (3) enabling real-time
material deposition tolerances, or patterning resolution as structured, dynamic control in response to environmental changes or task demands.
learnable inputs—to enable fabrication-aware optimization across Traditional approaches, which rely on periodic approximations, rule-
diverse manufacturing platforms. Furthermore, it is essential to develop based parameter sweeps, or heuristic controllers, struggle to address
algorithmic frameworks capable of explicitly differentiating and adap­ these demands efficiently. In contrast, AI offers a flexible and scalable
tively balancing between hard constraints (strict fabrication limits) and toolkit for capturing nonlocal interactions, learning multi-physics trade-
soft constraints (performance trade-offs), and dynamically adjusting offs, and generating intelligent control policies. This section reviews
their influence during optimization, thereby facilitating more nuanced recent advances in AI-driven full-metasurface design across these three
and practically aligned constraint management. Lastly, advances in AI- directions, outlining the growing role of AI in bridging structural,
based design may increasingly enable the realization of complex opti­ functional, and temporal complexity in metasurface systems.
cal functionalities—such as broadband, polarization-insensitive, or
multi-wavelength control—within simplified fabrication schemes, for 3.1. Modeling inter-unit interactions
instance using single-material, single-layer structures. Such progress
would mark a shift from merely accommodating fabrication constraints As metasurfaces evolve toward increasingly compact, broadband,
to actively leveraging AI to co-optimize both device performance and and multi-functional platforms, the long-standing assumption of inde­
process simplicity, ultimately contributing to the scalable realization of pendently operating unit cells under local phase approximation is
high-performance, robust metasurfaces in practical applications. becoming inadequate. In high-NA systems [94–96], wide field-of-view

Fig. 7. AI-driven metasurface design through modeling inter-unit interactions. (A) Deep convolutional neural networks to predict mutual coupling effects in met­
asurfaces [27]. (B) Multiplexed super-cell metasurface design and optimization with tandem residual networks [100]. (C) Physics-data-driven intelligent optimi­
zation of large-aperture metalenses [107]. (D) Deep-learning-enabled electromagnetic near-field prediction and inverse design of metasurfaces [108]. (E) General
characterization of intelligent metasurfaces with graph coupling network [109]. (F) Building multifunctional metasystems via algorithmic construction [30].

8
Y. Zeng and X. Jin iOptics 1 (2025) 100004

imaging [97–99], and tightly packed supercell architectures [100–102], effectiveness in capturing either local or global coupling, there remains a
near-field coupling between adjacent meta-atoms can substantially alter lack of unified architectures that can simultaneously account for mul­
the local electromagnetic response, rendering periodic-boundary-based tiscale interactions. Future efforts should focus on developing integrated
simulation models invalid. Traditional methods address such inter-unit modeling frameworks that combine local receptive-field learning (e.g.,
interactions through full-structure simulations or exhaustive para­ CNNs) with global structural context (e.g., GNNs or Transformers),
metric sweeps, which are computationally prohibitive and difficult to enabling comprehensive mapping from structural layouts to electro­
generalize. In contrast, AI-based models, particularly deep learning magnetic responses. Second, current models are primarily trained on
networks, offer a scalable and flexible alternative by learning the com­ small-scale, periodic structures, and often fail to generalize to large-area
plex mappings between local structure, surrounding context, and elec­ [107,110] or functionally multiplexed metasurface arrays [111,112].
tromagnetic behavior directly from data [103–106]. For instance, as Expanding the modeling boundary to encompass long-range coupling
shown in Fig. 7(A), An et al. [27] proposed a CNN that predicts the true across massive metasurfaces—such as those used in large-aperture
phase and amplitude of a meta-atom based on its own geometry and that metalenses or wide-angle holography—will require scalable,
of its immediate neighbors, enabling efficient correction for mutual context-aware learning paradigms that can adapt to irregular layouts
coupling effects. This context-aware predictor significantly improves the and broad structural diversity. Third, most existing works treat
performance of large-area devices such as beam deflectors and metal­ inter-unit coupling as an undesirable perturbation to be mitigated. A
enses. A complementary strategy, illustrated in Fig. 7(B), was introduced promising alternative is to consider coupling as an additional design
by Yeung et al. [100], who employed tandem residual networks to degree of freedom. To this end, future models should incorporate
generate complex multiplexed supercell designs. Their approach cap­ mechanisms for learning and exploiting controllable coupling behaviors,
tures coupling-aware structure–response relationships across a design allowing designers to actively leverage structural proximity and inter­
space of trillions of configurations, enabling accurate and diverse action effects to enhance wavefront shaping, broadband spectral con­
spectral targeting. Further extending this concept, Fig. 7(C) presents the trol, and multifunctional responses. Such advances will redefine the role
work of Ha et al. [107], who proposed a physics–data-driven intelligent of coupling from a constraint to a resource in next-generation meta­
optimizer for large-aperture metalens design. By embedding local surface design.
context into super meta-atom inputs, the model adaptively modifies
meta-atom dimensions in response to neighboring geometries, achieving 3.2. Multi-objective and multi-physics optimization
93.4 % focusing efficiency and five orders of magnitude acceleration
over traditional inverse design methods. Collectively, these studies As metasurface applications move toward multifunctionality and
demonstrate that AI is not only capable of modeling local coupling ef­ real-world deployment, design objectives have expanded from single-
fects, but also of leveraging them as functional degrees of freedom in parameter targets (e.g., phase control) to complex multi-objective and
metasurface optimization. multi-physics optimization tasks [113–115]. Conventional techniques
While the previous studies focus on capturing near-field interactions such as genetic algorithms or topology optimization often struggle with
among adjacent meta-atoms within local neighborhoods, metasurface the exponential complexity of such tasks, facing major limitations in
design at the system level necessitates modeling of long-range and hi­ exploring Pareto frontiers or integrating cross-domain objectives like
erarchical couplings across the entire structure. As devices scale in electromagnetic response and mechanical robustness [116,117]. Recent
aperture and functionality, particularly in wide-area, multilayer, and studies demonstrate that AI-driven approaches significantly alleviate
high-NA systems, traditional local-phase approximations become this burden. For instance, generative neural networks efficiently pro­
insufficient for capturing non-local electromagnetic interactions and duce optimized metasurface topologies with reduced computational cost
cascading layer effects. To address these challenges, recent works have compared to adjoint-based optimization [46]. Additionally, shape
turned to global AI models, such as graph neural networks (GNNs) and optimization methods leveraging AI-driven Fourier decomposition
coordinate-aware Transformers, which can learn spatially distributed effectively balance high efficiency and structural simplicity, enabling
interaction patterns across unit cells and between stacked layers. These faster convergence and easier manufacturability [118]. Such AI-based
models treat each meta-atom as a node within a graph, with structural methods enable rapid exploration of high-dimensional design spaces
features as node attributes and spatial proximity defining the adjacency and nonlinear trade-offs through unified latent representations and
relations. This formulation allows the model to learn complex wave­ composite loss strategies.
–structure interactions across the entire metasurface under various In multifunctional metasurface design, engineering objectives often
boundary conditions and incidence scenarios. For example, as shown in involve simultaneously satisfying multiple electromagnetic performance
Fig. 7(D), Kanmaz et al. [108] proposed an encoder–decoder network to metrics—such as phase control, transmission efficiency, and diffraction
directly map full metasurface configurations to near-field electromag­ behavior. Traditional approaches typically rely on sequential optimi­
netic responses. By testing triangular and randomized pillar layouts, zation or large-scale parameter sweeps, both of which suffer from
their model captures both local and global coupling effects, achieving exponential growth in search space and frequent objective conflicts,
millisecond-level predictions of full-field behaviors. Extending this idea leading to slow or stalled convergence. Early deep learning methods
to even larger metasurfaces, Wu et al. [109] introduced a GNN frame­ addressed part of this challenge by enabling one-shot prediction of
work for intelligent metasurface characterization, as illustrated in Fig. 7 single-target designs, but often failed in multi-objective scenarios due to
(E). Their model integrates diffraction theory and near-to-far-field mismatched dimensionalities and difficulty in balancing gradients
transformations to accurately predict scattered field distributions from across tasks. To overcome these limitations, An et al. [119] proposed a
large-scale structures, and was experimentally validated in microwave target-driven surrogate model that directly regresses a composite output
domains. Going beyond single-layer modeling, Zhu et al. [30] proposed vector containing multiple optical parameters, enabling simultaneous
a hybrid neural architecture for the design of multifunctional multilayer optimization of phase and efficiency within a unified framework, as
metasystems, as shown in Fig. 7(F). Their model simultaneously con­ shown in Fig. 8(A). Similarily, Zhu et al. [120] introduced a residual
siders inter-layer phase accumulation, coupling, and function multi­ neural network architecture to decouple amplitude and phase into
plexing, enabling the design of metasurfaces that perform second-order parallel sub-tasks, as shown in Fig. 8(B). Their approach leverages re­
differentiation, beam shaping, and full space-polarization-wavelength sidual branches and joint loss functions to generate high-resolution
multiplexing. holographic patterns without parameter sweeps, significantly
Looking ahead, several key directions remain to be explored in improving both image fidelity and encoding capacity.
advancing AI-enabled modeling of inter-unit interactions in meta­ As the design space expands into other physical domains, Zhou et al.
surfaces. First, while existing approaches have demonstrated [121] extended surrogate modeling to elastic metasurfaces, enabling

9
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 8. AI-driven multi-objective and multi-physics metasurface optimization. (A) Deep-learning-based approach for objective-driven all-dielectric metasurface
design [119]. (B) Deep-learning-empowered holographic metasurfaces with simultaneous amplitude and phase customization [120]. (C) An inverse design paradigm
of multi-functional elastic metasurface via data-driven machine learning [121]. (D) Spin-multiplexed metasurface inverse-design based on a bi-directional DNN for
terahertz wavefront control [73]. (E) Efficient gradient-based metasurface optimization toward the limits of wavelength-polarization multiplexing [122]. (F)
Frequency-domain transfer and inverse design of metasurfaces under multi-physics coupling via Eulerian latent dynamics and data-driven regularization [123].

rapid inverse design of structures optimized for both mechanical and encoding diverse physical constraints, learning coupled cross-domain
dynamic wave properties, as shown Fig. 8(C). Their model reduces response behaviors, and supporting inverse design across electromag­
design time from days to minutes and provides a transferable paradigm netic, mechanical, and thermal domains, AI-based frameworks unify
for acoustic and elastic system design. With growing demand for previously fragmented design spaces under a single optimization pipe­
high-capacity multiplexing, And Wei et al. [73] (Fig. 8(D)) developed a line. Looking ahead, future research should move beyond fixed-weight
bi-directional network for inverse design of spin-multiplexed terahertz optimization strategies toward adaptive task-balancing schemes that
metasurfaces, supporting independent control of linear and circular dynamically manage objective trade-offs and physical conflict. More­
polarization channels. In parallel, Bao et al. [122] (Fig. 8(E)) combined over, developing unified latent representations and multi-scale encoding
gradient-based inverse design with DNNs to realize metasurfaces sup­ mechanisms across heterogeneous physical domains will be essential for
porting nine independent channels (three wavelengths × three polari­ achieving truly collaborative, cross-physics metasurface design. Finally,
zations), maintaining over 90 % task accuracy while reducing design incorporating physics-informed priors—such as symbolic constraints,
iterations by an order of magnitude. Finally, Zhu et al. [123] addressed data-driven physical consistency, and model-guided augmenta­
cross-frequency and multi-physics coupling by integrating DNN and tion—may enhance not only prediction accuracy but also the inter­
Eulerian latent dynamics. Their framework enables model-free electro­ pretability and generalizability of AI models. Collectively, these
magnetic–thermal design transfer and achieves sub-5 % prediction error directions will advance metasurface optimization from empirical,
even in data-scarce regimes, demonstrating the scalability and gener­ target-driven routines to mechanism-aware, system-level intelligent
alization of AI in complex, multi-objective metasurface design, as shown design paradigms.
in Fig. 8(F).
Together, these advances demonstrate that AI not only accelerates 3.3. Real-time dynamic control
the metasurface design process but also enables a fundamental paradigm
shift toward multi-objective and multi-physics optimization. By jointly Dynamic metasurfaces refer to reconfigurable platforms whose

10
Y. Zeng and X. Jin iOptics 1 (2025) 100004

electromagnetic responses can be modulated in real time under external control, researchers have explored on-site learning and integrated
stimuli, enabling functions such as adaptive wavefront shaping [129], sensing-communication capabilities. Liu et al. [126] in Fig. 9(D) inte­
on-demand spectral tuning [130], and programmable polarization grated a programmable diffractive DNN with digital coding meta­
control [131]. Unlike static designs, these systems typically integrate surfaces, introducing on-site learning via RL algorithms to handle
tunable materials (e.g., phase-change media, liquid crystals, or 2D diverse sensing and communication tasks. Wu et al. [127] in Fig. 9(E)
semiconductors) or embedded control circuits to enable spatiotemporal further combined DNN-enhanced genetic algorithms with metasurface
manipulation of optical behavior [13,128,132–134]. beam design, achieving precise multi-target detection and adaptive
However, achieving high-speed, low-latency, and robust control re­ response for integrated electromagnetic sensing. Lastly, Li et al. [128] in
mains challenging. Traditional methods—based on predefined look-up Fig. 9(F) developed a flexible intelligent surface platform (FISP) as a
tables, heuristic rules, or slow-response physical switches—are often solution to achieve flexible RM with highly stable performance under
inadequate for dynamic environments, especially when inputs vary dynamic deformation, demonstrated its applications in electromagnetic
rapidly or involve complex feedback interactions. These limitations illusion, carpet cloaking, and data transmission.
underscore the need for learning-based strategies that can autonomously With advances in programmable materials and embedded control
model system dynamics, predict optimal control actions, and adapt to architectures, dynamic metasurfaces are rapidly evolving from static
unseen conditions in real time. AI-based strategies, particularly rein­ wavefront modulators into intelligent physical systems capable of
forcement learning (RL), recurrent neural networks (RNNs), and model perception, decision-making, and adaptive response. Recent de­
predictive controllers, offer an alternative paradigm. These methods velopments have leveraged AI techniques to realize real-time geometric
enable end-to-end learning of control policies directly from temporal adaptation, object tracking, wireless channel management, and multi­
signals, leveraging nonlinear state-to-action mappings without requiring modal sensing across a variety of electromagnetic scenarios. These
explicit modeling of metasurface responses. As shown in Fig. 9(A), Li studies collectively demonstrate the unique advantages of AI in
et al. [124] demonstrated an intelligent metasurface imager and modeling complex system dynamics, learning nonlinear control policies,
recognizer, employing three hierarchical ANNs to transform microwave and adapting to feedback in real time. Nevertheless, several critical
measurements into whole-body images, classify designated spots (hand challenges remain. First, current control policies are typically task-
and chest), and instantly recognize human hand gestures at the Wi-Fi specific and lack generalization across applications. Future research
frequency of 2.4 GHz. Similarly, Fig. 9(B) presents a self-adaptive mi­ should explore transferable RL frameworks or memory-augmented
crowave cloak developed by Qian et al. [47], where deep learning en­ sequential decision models to support rapid policy deployment across
ables millisecond-scale response to dynamic incident fields without diverse environments. Second, there remains a disconnect between
human intervention. At the system level, Fig. 9(C) illustrates a sensing, decision-making, and hardware-level actuation, limiting
neuro-metasurface architecture proposed by Fan et al. [125], combining system-wide responsiveness and integration. To address this, co-
perception, decision, and actuation into a single AI-driven loop for optimization of AI control strategies with circuit-level implementa­
real-time wireless channel management in dynamic environments. To tions and tunable material responses will be essential to achieve low-
further advance metasurface intelligence beyond direct adaptive latency, high-frequency reconfiguration. Third, most existing systems

Fig. 9. AI-driven Real-time metasurface dynamic control. (A) Intelligent metasurface imager and recognizer [124] (B) Deep-learning-enabled self-adaptive mi­
crowave cloak without human intervention [47]. (C) Homeostatic neuro-metasurfaces for dynamic wireless channel management [125]. (D) A programmable dif­
fractive deep neural network based on a digital-coding metasurface array [126]. (E) Integrated electromagnetic sensing system based on a
deep-neural-network-intervened genetic algorithm [127]. (F) Flexible intelligent microwave metasurface with shape-guided adaptive programming [128].

11
Y. Zeng and X. Jin iOptics 1 (2025) 100004

optimize for single-task objectives, such as beam steering or target reconstruction from degraded or entangled metasurface outputs,
recognition. A promising direction is to investigate multi-modal infor­ enhancing interpretability without modifying optical hardware. On the
mation fusion and multi-task adaptive scheduling, enabling meta­ other, end-to-end differentiable optimization connects nanostructure
surfaces to dynamically couple sensing goals with control strategies and design directly to application objectives—such as imaging fidelity or
paving the way toward next-generation intelligent platforms with recognition accuracy—through joint training pipelines. This section
autonomous, context-aware functionalities. reviews recent advances in both directions, highlighting how AI trans­
forms metasurface systems into intelligent, co-optimized photonic
4. AI for system-level optimization pipelines.

While most metasurface research has traditionally focused on opti­

mizing optical structures in isolation, practical deployment requires 4.1. Backend-processing reconstruction
tight integration with backend perception, reconstruction, and task-
specific decision modules. At the system level, imperfections from In practical metasurface systems, the optical signals collected at the
front-end optics—such as limited resolution, chromatic aberration, and output often suffer from severe degradation due to design constraints,
compressed encoding—should be jointly addressed alongside hardware fabrication imperfections, and operational bandwidth limitations. These
constraints like sensor bandwidth and processing latency. Conventional raw measurements—such as speckle patterns, compressed encodings, or
methods typically isolate optical design from downstream data pro­ defocused projections—are typically noisy, distorted, and structurally
cessing, resulting in fragmented workflows and suboptimal end-task ambiguous, posing significant challenges for traditional decoding algo­
performance. AI provides a unified framework to bridge this gap. On rithms [135–137]. AI-based backend processing, which enhances signal
one hand, learning-based backend processing enables robust signal interpretability and task performance without modifying frontend
hardware, has emerged as a key enabler for real world metasurface

Fig. 10. AI-driven backend processing reconstruction for metasurface imaging. (A) Computational complex optical field imaging using a designed metasurface
diffuser [138]. (B) 3D imaging using extreme dispersion in optical metasurfaces [136]. (C) Trilobite-inspired neural nanophotonic light-field camera with extreme
depth of field [48]. (D) A meta-device for intelligent depth perception [141]. (E) Neural-network-enhanced metalens camera for high-definition, dynamic imaging in
the long-wave infrared spectrum [142]. (F) High-spatial and spectral-resolution snapshot spectral imaging based on near-field-coupled metasurfaces [143].

12
Y. Zeng and X. Jin iOptics 1 (2025) 100004

deployment. One representative approach focuses on reconstructing deep learning-based decoding. Their encoder compresses spatial–spec­
explicit information from single-frame, non-structured observations, tral mappings into footprints as small as 5 μm (comparable to typical
such as speckle fields generated in scattering media or compressed CMOS pixel sizes), while deep reconstruction networks recover
metasurface outputs under low-light conditions. As shown in Fig. 10(A), high-resolution spectral images across the visible band. Together, these
Kwon et al. [138] introduced a computational imaging system that uses works illustrate that AI not only supports signal recovery but also
a random dielectric metasurface diffuser with known scattering char­ actively extends the physical performance envelope of metasurface im­
acteristics to encode optical fields into complex speckle patterns. A aging systems, representing a key paradigm shift from algorithmic
deep-learning-enhanced decoding process based on speckle-correlation correction to learned resolution enhancement.
principles enables the recovery of complex amplitude without AI-powered backend reconstruction has shown significant advan­
requiring multi-frame redundancy or frequency-domain priors. Recent tages in metasurface systems by enabling the recovery of target infor­
AI-assisted methods have further advanced phase retrieval capabilities, mation from severely degraded measurements, such as speckle fields,
for instance, Zhou et al. [139] demonstrated precise phase reconstruc­ compressed encodings, or low-SNR projections, and by effectively dis­
tion from two defocused metasurface-generated images, and subsequent entangling high-dimensional physical quantities such as spectral, depth,
breakthroughs showed that deep-learning algorithms can even reliably and polarization information. Furthermore, AI models have demon­
extract phase information from a single captured image, dramatically strated the ability to surpass the physical limits of optical systems by
enhancing the space-bandwidth product and practical efficiency in enhancing resolution, inferring occluded structures, and restoring high-
metasurface imaging systems [140]. Compared to traditional scattering frequency content beyond classical bandwidth constraints. Despite these
media, the metasurface diffuser offers superior reproducibility, robust­ advances, several key challenges remain. First, most existing models are
ness, and ease of simulation, allowing AI models to generalize effectively tailored to specific imaging modalities and lack generalization across
across different environmental conditions. diverse signal types and system configurations. Future work should
Beyond denoising or deblurring single-frame inputs, a critical chal­ explore general-purpose reconstruction frameworks with multi-
lenge in metasurface-based systems lies in the recovery of multi- structure awareness and dynamic adaptability, capable of operating
dimensional information—such as spectrum, depth, and polar­ robustly under varying optical conditions, signal encodings, and task
ization—that is often encoded in highly entangled or compressed optical requirements. Second, the reconstruction pipeline is often developed
signals. Traditional decoding techniques struggle to disentangle these independently of the metasurface design process, limiting the system’s
coupled dimensions, especially under constraints of limited measure­ overall performance ceiling. A promising direction is to establish end-to-
ments, bandwidth, or computational resources. AI-based can learn latent end co-optimization frameworks in which metasurface structural pa­
mappings that decouple physical dimensions from complex rameters and neural decoding architectures are jointly trained under
metasurface-encoded raw data. As shown in Fig. 10(B), Tan et al. [136] unified task-driven objectives, thereby achieving global system-level
exploited the extreme dispersion of metasurfaces to build a passive 3D optimization. Lastly, the strong dependence on large-scale labeled
imaging system, using a single-shot chromatic defocus image to jointly datasets restricts real-world deployment in resource-limited or
recover RGB texture and depth information via a dual-network pipeline. application-specific scenarios. Future research may incorporate physics-
In Fig. 10(C), Fan et al. [48] proposed a trilobite-inspired bifocal informed self-supervised learning, generative augmentation strategies,
light-field camera integrating spin-multiplexed metalenses and a and domain-adaptive transfer learning to enhance model generaliz­
multi-scale CNN, enabling simultaneous macro and telephoto capture ability and robustness without compromising performance.
across kilometer-scale depth ranges with image-level resolution recov­
ery. Extending this idea further, Chen et al. [141] in Fig. 10(D) devel­ 4.2. End-to-end optimization
oped a multifunctional meta-device based on a dense achromatic
metalens array, which supports both passive light-field imaging and Traditional metasurface design often separates the optimization of
active structured light projection. By leveraging deep convolutional optical structure from backend perception, leading to suboptimal
networks, the system decodes depth cues in both modalities and adapts system-level performance due to fragmented modeling and stage-wise
across lighting conditions. These works highlight how AI-driven tuning. End-to-end optimization frameworks, enabled by AI, integrate
dimensional decomposition not only enhances metasurface signal nanostructure parameters, physical propagation models, and task-
interpretability but also expands the scope of compact metasystems to specific losses within a fully differentiable architecture. This design
tackle complex visual tasks like full-scene 3D reconstruction, intelligent paradigm directly links metasurface layout to final application objecti­
depth sensing, and multi-modal holography. ves—such as image fidelity, classification accuracy, or optical effi­
Even when metasurface imaging systems operate within structurally ciency—thus unlocking “task-driven structure design” through gradient
resolvable and input-discernible regimes, conventional reconstruction backpropagation [144–155]. As shown in Fig. 11(A), Tseng et al. [153]
algorithms remain fundamentally constrained by physical limitations proposed a neural nano-optics framework that jointly optimizes a
such as optical resolution, sensor sampling, and device bandwidth. In full-color metasurface and a deconvolutional neural network to achieve
contrast, AI-based post-processing provides a means to transcend these aberration-free imaging with a single metalens, matching commercial
limitations through learned priors and nonlinear mappings that enable six-element lens performance being approximately 550,000 times
structural inference, detail enhancement, and perceptual recovery thinner. Similarly, Chakravarthula et al. [154] introduced a learned
beyond classical thresholds. These models can interpolate spatial de­ nanophotonic camera array operating under natural conditions, where
tails, restore occluded features, and amplify high-frequency components metasurface phase design and generative diffusion-based reconstruction
by aligning physical outputs with learned target distributions. As shown are co-optimized for broadband outdoor scenes, as shown in Fig. 11 (B).
in Fig. 10(E), Wei et al. [142] integrated a high-frequency-enhancing Beyond lens imaging, Seo et al. [49] in Fig. 11(C) leveraged a deep re­
cycle-GAN into a long-wave infrared metalens imaging system, signifi­ sidual model for metalens image recovery under severe chromatic and
cantly improving dynamic imaging fidelity at a frame rate of 125 fps. angular aberrations, pushing the achievable peak signal-to-noise ratio
The model introduces a wavelet-based feedback loop in the adversarial (PSNR) and perceptual scores well beyond classical optics. Furthermore,
loss function, enabling the neural network to selectively recover Yin et al. [148] demonstrated in Fig. 11(D) a direct
high-frequency components suppressed by the metalens’ intrinsic metasurface-to-hologram inverse design pipeline that circumvents Jones
low-pass filtering. In a complementary approach, Zhang et al. [143], as matrix engineering and supports 12-channel multiplexing using only
shown in Fig. 10(F), demonstrated a near-field coupled metasurface for two degrees-of-freedom per meta-atom, substantially reducing fabrica­
snapshot spectral imaging, where the trade-off between spatial and tion complexity. And Chi et al. [147]expanded this paradigm to full light
spectral resolution was resolved via localized resonance engineering and field control across wavelength, polarization, and depth using a

13
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Fig. 11. AI-driven end-to-end design for metasurface imaging. (A) Neural nano-optics for high-quality thin lens imaging [153]. (B) Thin on-sensor nanophotonic
array cameras [154]. (C) Deep-learning-driven end-to-end metalens imaging [49]. (D) Multi-dimensional multiplexed metasurface holography by inverse design
[148]. (E) Neural network-assisted end-to-end design for full light-field control of meta-optics [147]. (F) Beating spectral bandwidth limits for large aperture
broadband nano-optics [144].

neural-assisted global optimization method, supporting tri-polarized research should explore meta-learning strategies or conditional gener­
and non-orthogonal multiplexed metasystems, as shown in Fig. 11(E). ative models that enable real-time structural adaptation and cross-task
Finally, Fig. 11(F) illustrates how Fröch et al. [144] overcame the generalization. Second, while recent studies have achieved co-
inherent spectral-bandwidth tradeoff in large-aperture fiat optics by optimization across physical and task layers, the absence of standard­
co-training meta-optics and a learned post-processing network with a ized differentiable optical propagation models and unified structural
dual-aperture dataset, enabling high-resolution broadband imaging encoding schemes limits cross-platform generality and reproducibility.
with a 1 cm, f/2 metalens. Collectively, these studies exemplify how Developing modular, plug-and-play differentiable optical simulators
AI-driven end-to-end co-design reshapes the metasurface work­ and scalable structural encoding schemes will be essential for advancing
flow—bridging structure, physics, and tasks into a single learning the engineering portability of such systems. Third, current end-to-end
loop—and offers a scalable, fabrication-aware, and task-adaptive pipelines largely focus on maximizing task performance, often neglect­
pathway toward the next generation of intelligent meta-optical systems. ing fabrication constraints, structural robustness, and material
AI-driven end-to-end optimization frameworks are fundamentally compatibility. Future directions should aim to incorporate
transforming the metasurface design pipeline by integrating nano­ manufacturing tolerances, material limitations, and functional zoning
structure configuration, physical wave propagation, and task-level ob­ constraints directly into the joint optimization objective, thereby tran­
jectives into a unified, fully differentiable architecture. This paradigm sitioning end-to-end metasurface design from performance-centric par­
overcomes the inherent fragmentation of conventional design work­ adigms toward engineering-ready workflows that balance optical
flows—where optical layout and signal processing are decoupled—and performance, manufacturability, and reliability.
has demonstrated strong potential across a range of applications, from In summary, this review has systematically explored the evolution
single-layer imaging metalenses and broadband nanophotonic cameras and future directions of AI-empowered optical metasurfaces across
to multi-channel holography and full light-field control. Despite these multiple hierarchical scales, as comprehensively outlined in Table 1. At
advances, several challenges remain. First, most existing frameworks are the unit-cell optimization level, advanced surrogate modeling, inverse
optimized for static, predefined tasks and lack adaptability to dynamic design frameworks, and robust design strategies empowered by AI
perception objectives or changing environmental conditions. Future significantly accelerate electromagnetic (EM) simulations, overcome

14
Y. Zeng and X. Jin iOptics 1 (2025) 100004

Table 1
Summary of AI-empowered optical metasurface methodologies across different levels, highlighting AI-driven objectives, representative methods, and corresponding
key advantages.
Design Level Sub-task AI-driven Objective Representative Methods Key Advantages

Unit-Cell Surrogate modelling Replace full-wave solvers with fast, FNN, CNN, active learning Speed-up; analytic gradients
Optimization differentiable EM predictors
Inverse design Solve one-to-many mapping VAE, GAN Rapid multi-solution synthesis; few
design iterations
Robust design Embed process tolerances and material Uncertainty-aware CNN, Bayesian nets, High efficiency retained under
variability into design loop physics-guided-augmentation realistic errors
Metasurface Level Inter-unit coupling Learn near- and far-field interactions Context-aware CNN, GNN, Transformers Full-aperture accuracy without
Design modelling beyond local-phase approximation massive FDTD
Multi-objective/multi- Jointly balance Multi-objective Multi-task DNN, Pareto RL Efficient Pareto search;
physics optimization constraints multifunctional broadband devices
Real-time dynamic control Closed-loop, millisecond reconfiguration RL, RNN Autonomous wavefront tuning;
of tunable metasurfaces adaptive operation
System-Level Backend-processing Recover high-quality imaging U-Net, auto-encoder, Transformer; SNR and resolution beyond optics-
Optimization reconstruction only limits
End-to-end co- Jointly train nanostructure and post- Differentiable optical simulator+ CNN/ Hardware–software codesign
optimization processing ResNet

complex inverse problems, and ensure performance under realistic robustness—are urgently required. A promising direction is hierarchical
fabrication conditions. Progressing to the metasurface-level design, AI optimization strategies that embed these diverse physical constraints
methodologies facilitate accurate modeling of intricate inter-unit elec­ within a coherent, differentiable optimization framework. Such methods
tromagnetic interactions, enable efficient multi-objective and multi- will be essential in creating multifunctional, high-performance optical
physics optimizations, and provide real-time adaptive control capabil­ metasurfaces that simultaneously satisfy complex, real-world
ities. Finally, at the system level, end-to-end AI-driven frameworks have specifications.
emerged, integrating optical nanostructure design with computational At the system integration and intelligent control level, future de­
backend reconstruction, thus redefining the performance boundaries velopments should bridge the persistent gap between optical hardware
achievable in metasurface-based imaging and optical systems. Collec­ design and backend computational architectures, transforming meta­
tively, these developments signify a clear shift from isolated structural surfaces into truly autonomous photonic platforms. Establishing fully
design toward holistic, intelligent, and fully integrated optimization differentiable end-to-end co-design pipelines—combining metasurface
paradigms, underscoring the transformative potential of AI in advancing structure optimization, optical propagation modeling, and AI-driven
next-generation optical metasurface technologies. backend reconstruction—will significantly elevate performance, adapt­
ability, and application portability. Another vital frontier is real-time
5. Conclusion adaptive control. Integrating lightweight, high-speed neural inference
models with dynamically reconfigurable metasurface platforms can
This review has systematically summarized recent advances and enable closed-loop, self-adaptive optical systems capable of intelligently
emerging frontiers in AI-empowered optical metasurfaces, clearly adapting to changing environmental conditions. This convergence will
delineating a comprehensive design trajectory from unit-cell optimiza­ enable metasurfaces to perform real-time tasks such as autonomous
tion to full system-level integration. Artificial intelligence is funda­ aberration correction, adaptive beam steering, and dynamic environ­
mentally reshaping the optical design paradigm—not merely as an mental sensing without human intervention.
auxiliary computational tool, but as a critical enabler driving metasur­ Furthermore, it is worth highlighting that artificial intelligence is not
face research toward unprecedented performance, complexity, and only reshaping the academic landscape of optical metasurface research
autonomous intelligence. but is also playing a pivotal role in accelerating its commercialization
At the unit-cell optimization level, future efforts should prioritize and industrial adoption. Recent breakthroughs have led to the emer­
physics-informed machine learning methods that integrate fundamental gence of several successful startup companies, such as Meta-lenz, which
electromagnetic theory directly into AI frameworks. Rather than relying applies metasurface-based polarization imaging technologies to critical
on extensive training datasets, developing compact yet physically applications like smartphone facial recognition systems; NILT, special­
consistent surrogate models—possibly using transfer learning, physics- izing in nanoimprint lithography for scalable fabrication of high-
guided neural networks, and uncertainty quantification—will enhance performance metasurface optics; and IOPTEE, dedicated to integrating
the accuracy, generalizability, and data-efficiency of inverse design metasurface technology into industrial imaging and display products.
processes. Additionally, robustness-aware optimization that systemati­ These examples clearly demonstrate that metasurfaces have rapidly
cally accounts for fabrication variability and material uncertainty transitioned from academic innovations to viable industrial technolo­
should become a fundamental standard rather than an afterthought. gies. In this context, artificial intelligence serves as a key enabler,
This approach will drive the translation of AI-optimized meta-atom significantly improving the efficiency of the design process, reducing
designs from ideal simulations into practically manufacturable computational costs, and enhancing robustness against fabrication
components. imperfections—thereby accelerating technology translation from proof-
At the metasurface device and array level, addressing multiscale and of-concept prototypes to scalable, market-ready products. Looking for­
multi-objective design challenges remains critical. Future research ward, the integration of advanced AI frameworks such as large-scale
should move beyond traditional local approximations and develop foundation models—including pre-trained generative transformers and
unified AI architectures—such as hierarchical graph neural networks multimodal neural networks—will further streamline the end-to-end
and transformer-based models—that rigorously capture complex near- design-manufacturing-validation cycle, substantially reducing product
field and far-field coupling phenomena across large and aperiodic development timelines, cost, and risk. Consequently, AI-driven meth­
metasurfaces. Moreover, AI techniques capable of efficiently navigating odologies promise not only to foster further entrepreneurial endeavors
multi-dimensional trade-off spaces—balancing electromagnetic perfor­ but also to firmly establish metasurfaces as a foundational technology
mance, bandwidth, thermal management, and mechanical within the broader optics industry, enabling widespread commercial

15
Y. Zeng and X. Jin iOptics 1 (2025) 100004

adoption and market success. [15] Y. Hu, X. Wang, X. Luo, et al., All-dielectric metasurfaces for polarization
manipulation: principles and emerging applications, Nanophotonics 9 (2020)
Ultimately, the full integration of artificial intelligence with optical
3755–3780.
metasurface technologies marks a profound shift toward intelligent [16] F. Ding, B. Chang, Q. Wei, et al., Versatile polarization generation and
photonics, spanning from nanoscale optimization to system-level manipulation using dielectric metasurfaces, Laser & Photon. Rev. 14 (2020)
adaptive intelligence. By systematically addressing these open challen­ 2000116.
[17] J. Hong, H. Son, C. Kim, et al., Absorptive metasurface color filters based on
ges—leveraging physics-informed, data-efficient unit-cell models, uni­ hyperbolic metamaterials for a cmos image sensor, Opt. Express 29 (2021)
fied multi-scale device optimization frameworks, and end-to-end 3643–3658.
intelligent system architectures, as well as capitalizing on recent ad­ [18] F.T. Faul, L. Cronier, A. Alhulaymi, et al., Agile free-form signal filtering and
routing with a chaotic-cavity-backed non-local programmable metasurface, Adv.
vances in large-scale foundational AI models—we pave the way for Sci. (2025) 2500796.
transformative applications and unprecedented capabilities beyond the [19] L. Li, Z. Liu, X. Ren, et al., Metalens-array–based high-dimensional and
reach of conventional optical design methods. Future metasurface sys­ multiphoton quantum source, Science 368 (2020) 1487–1490.
[20] T. Santiago-Cruz, S.D. Gennaro, O. Mitrofanov, et al., Resonant metasurfaces for
tems will not only dynamically adapt and optimize but also autono­ generating complex quantum states, Science 377 (2022) 991–995.
mously learn and evolve, heralding a new era of self-evolving photonic [21] T. Li, X. Xu, B. Fu, et al., Integrating the optical tweezers and spanner onto an
systems and intelligent optical integration. individual single-layer metasurface, Photon. Res. 9 (2021) 1062–1068.
[22] T. Li, J.J. Kingsley-Smith, Y. Hu, et al., Reversible lateral optical force on phase-
gradient metasurfaces for full control of metavehicles, Opt. Lett 48 (2023)
Author contributions 255–258.
[23] H. Xu, T. Li, G. Gao, et al., Metasurface-integrated pattern-preserved fiber mode
separator, Adv. Funct. Mater (2025) 2505081.
The review proposal was initiated by Y.Z., with contributions from X. [24] Q. Zhang, Z. He, Z. Xie, et al., Diffractive optical elements 75 years on: from
J. The primary responsibility for the thesis’s main body lies with Y.Z., micro-optics to metasurfaces, Photon. Insights 2 (2023).
while X.J. serve as the primary supervisors. [25] N.A. Rubin, Z. Shi, F. Capasso, Polarization in diffractive optics and metasurfaces,
Adv. Opt. Photon. 13 (2022) 836–970.
[26] A. Ueno, J. Hu, S. An, AI for optical metasurface, npj Nanophoton. 1 (2024) 36.
Funding [27] S. An, B. Zheng, M.Y. Shalaginov, et al., Deep convolutional neural networks to
predict mutual coupling effects in metasurfaces, Adv. Opt. Mater 10 (2022)
2102113.
This work was supported by Shenzhen Science and Technology [28] V.E. Babicheva, A.B. Evlyukhin, Metasurfaces with electric quadrupole and
Program (JCYJ20241202123921029), National Natural Science Foun­ magnetic dipole resonant coupling, ACS Photon. 5 (2018) 2022–2033.
dation of China (62205178), Major Key Project of PCL (PCL2023A10–3). [29] H. He, Y. Zhang, Y. Shao, et al., Meta-attention network based spectral
reconstruction with snapshot near-infrared metasurface, Adv. Mater 36 (2024)
2313357.
[30] D. Zhu, Z. Liu, L. Raju, et al., Building multifunctional metasystems via
Declaration of Competing Interest
algorithmic construction, ACS nano 15 (2021) 2318–2326.
[31] W. Ma, Z. Liu, Z.A. Kudyshev, et al., Deep learning for the design of photonic
The authors declare no competing interests. structures, Nat. Photon. 15 (2021) 77–90.
[32] M.K. Chen, X. Liu, Y. Sun, D.P. Tsai, Artificial intelligence in meta-optics, Chem.
Rev. 122 (2022) 15356–15413.
Acknowledgments [33] N. Wu, Y. Sun, J. Hu, et al., Intelligent nanophotonics: when machine learning
sheds light, eLight 5 (2025) 5.
[34] W. Ji, J. Chang, H.-X. Xu, et al., Recent advances in metasurface design and
This work is supported in part by Shenzhen Science and Technology quantum optics applications with machine learning, physicsinformed neural
Program under Grant JCYJ20241202123921029; in part by Natural networks, and topology optimization methods, Light. Sci. & Appl 12 (2023) 169.
Science Foundation of China under Grant 62131011 and in part by the [35] N. Yu, P. Genevet, M.A. Kats, et al., Light propagation with phase discontinuities:
generalized laws of refiection and refraction, science 334 (2011) 333–337.
Major Key Project of PCL under Grant PCL2023A10–3.
[36] D. Lin, P. Fan, E. Hasman, M.L. Brongersma, Dielectric gradient metasurface
optical elements, science 345 (2014) 298–302.
References [37] M. Khorasaninejad, W.T. Chen, R.C. Devlin, et al., Metalenses at visible
wavelengths: diffraction-limited focusing and subwavelength resolution imaging,
Science 352 (2016) 1190–1194.
[1] N. Yu, F. Capasso, Flat optics with designer metasurfaces, Nat. Mater. 13 (2014)
[38] S. Wang, P.C. Wu, V.-C. Su, et al., A broadband achromatic metalens in the
139–150.
visible, Nat. Nanotechnol. 13 (2018) 227–232.
[2] M. Khorasaninejad, F. Capasso, Metalenses: versatile multifunctional photonic
[39] R.J. Lin, V.-C. Su, S. Wang, et al., Achromatic metalens array for full-colour light-
components, Science 358 (2017) eaam8100.
field imaging, Nat. Nanotechnol. 14 (2019) 227–231.
[3] A.I. Kuznetsov, M.L. Brongersma, J. Yao, et al., Roadmap for optical
[40] T. Li, Y. Chen, B. Fu, et al., Spin-selective trifunctional metasurfaces for
metasurfaces, ACS photonics 11 (2024) 816–865.
deforming versatile nondiffractive beams along the optical trajectory, Laser
[4] A. Arbabi, A. Faraon, Advances in optical metalenses, Nat. Photonics 17 (2023)
Photon. Rev. 18 (2024) 2301372.
16–25.
[41] W. Liu, Y. Wan, H. Yu, et al., Miniature snapshot mid-infrared spectrometer based
[5] H. Zhou, C. Zhao, C. He, et al., Optical computing metasurfaces: applications and
on metal-insulator-metal metasurface, J. Opt. 26 (2024) 085003.
advances, Nanophotonics 13 (2024) 419–441.
[42] F. Yesilkoy, E.R. Arvelo, Y. Jahani, et al., Ultrasensitive hyperspectral imaging
[6] D. Conteduca, I. Barth, G. Pitruzzello, et al., Dielectric nanohole array
and biodetection enabled by dielectric metasurfaces, Nat. Photon. 13 (2019)
metasurface for high-resolution near-field sensing and imaging, Nat.
390–396.
communications 12 (2021) 3293.
[43] A. Zaidi, N.A. Rubin, M.L. Meretska, et al., Metasurface-enabled single-shot and
[7] H. Kim, H. Yun, S. Jeong, et al., Optical metasurfaces for biomedical imaging and
complete mueller matrix imaging, Nat. Photon. (2024) 1–9.
sensing, ACS nano (2025).
[44] B. Fu, X. Zhou, T. Li, et al., Miniaturized high-efficiency snapshot polarimetric
[8] X. Liao, L. Gui, S. Bi, et al., Multi-objective optimization of chiral metasurface for
stereoscopic imaging, Optica 12 (2025) 391–398.
sensing based on a distributed algorithm, IEEE Photonics J 16 (2023) 1–6.
[45] Z. Liu, D. Zhu, S.P. Rodrigues, et al., Generative model for the inverse design of
[9] J. Hu, Z. Guo, J. Shi, et al., A metasurface-based full-color circular auto-focusing
metasurfaces, Nano letters 18 (2018) 6570–6576.
airy beam transmitter for stable high-speed underwater wireless optical
[46] J. Jiang, J.A. Fan, Global optimization of dielectric metasurfaces using a physics-
communications, Nat. Commun 15 (2024) 2944.
driven neural network, Nano letters 19 (2019) 5366–5372.
[10] D. Xu, W. Xu, Q. Yang, et al., All-optical object identification and three-
[47] C. Qian, B. Zheng, Y. Shen, et al., Deep-learning-enabled self-adaptive microwave
dimensional reconstruction based on optical computing metasurface, Opto-
cloak without human intervention, Nat. Photon. 14 (2020) 383–390.
Electron. Adv. 6 (2023) 230120.
[48] Q. Fan, W. Xu, X. Hu, et al., Trilobite-inspired neural nanophotonic light-field
[11] Y. Yin, Q. Jiang, H. Wang, L. Huang, Color holographic display based on complex-
camera with extreme depth-of-field, Nat. Commun. 13 (2022) 2130.
amplitude metasurface, Laser & Photon. Rev. (2024) 2400884.
[49] J. Seo, J. Jo, J. Kim, et al., Deep-learning-driven end-to-end metalens imaging,
[12] Z. Liu, H. Gao, T. Ma, et al., Broadband spin and angle co-multiplexed waveguide-
Adv. Photon. 6 (2024).
based metasurface for six-channel crosstalk-free holographic projection, eLight 4
[50] A. Tafiove, S.C. Hagness, M. Piket-May, Computational electromagnetics: the
(2024) 7.
finite-difference time-domain method, Electr. Eng. Handb 3 (2005) 15.
[13] G. Jiang, Y. Wang, Z. Zhang, et al., Abnormal beam steering with kirigami
[51] M. Moharam, T.K. Gaylord, Rigorous coupled-wave analysis of planar-grating
reconfigurable metasurfaces, Nat. Commun. 16 (2025) 1660.
diffraction, J. Opt. Soc. Am 71 (1981) 811–818.
[14] P.C. Wu, R.A. Pala, G. Kafaie Shirmanesh, et al., Dynamic beam steering with all-
[52] S. So, J. Mun, J. Park, J. Rho, Revisiting the design strategies for metasurfaces:
dielectric electro-optic iii–v multiple-quantum-well metasurfaces, Nat.
fundamental physics, optimization, and beyond, Adv. Mater 35 (2023) 2206399.
communications 10 (2019) 3654.

16
Y. Zeng and X. Jin iOptics 1 (2025) 100004

[53] D. Lee, W. Chen, L. Wang, et al., Data-driven design for metamaterials and [86] Y. Zeng, H. Cao, and X. Jin, “Anchor-controlled generative adversarial network
multiscale systems: a review, Adv. Mater 36 (2024) 2305254. for high-fidelity electromagnetic and structurally diverse metasurface design,”
[54] A. Razi, A. Safdar, R. Irfan, Enhancing tandem solar cell’s efficiency through arXiv preprint arXiv:2408.16231 (2024).
convolutional neural network-based optimization of metasurfaces, Mater. & Des. [87] V. Medvedev, A. Erdmann, A. Rosskopf, Physics-informed deep learning for 3d
236 (2023) 112475. modeling of light diffraction from optical metasurfaces, Opt. Express 33 (2025)
[55] O. Khatib, S. Ren, J. Malof, W.J. Padilla, Learning the physics of all-dielectric 1371–1384.
metamaterials with deep lorentz neural networks, Adv. Opt. Mater 10 (2022) [88] C. Yeung, R. Tsai, B. Pham, et al., Global inverse design across multiple photonic
2200097. structure classes using generative deep learning, Adv. Opt. Mater 9 (2021)
[56] S. An, B. Zheng, M.Y. Shalaginov, et al., Deep learning modeling approach for 2100548.
metasurfaces with high degrees of freedom, Opt. Express 28 (2020) [89] A. Ueno, H.-I. Lin, F. Yang, et al., Dual-band optical collimator based on deep-
31932–31942. learning designed, fabrication-friendly metasurfaces, Nanophotonics 12 (2023)
[57] Z. Sun, T. Li, S. Kuang, et al., “On-demand quick metasurface design with 3491–3499.
neighborhood attention transformer,” arXiv preprint arXiv:2412.08405 (2024). [90] R.P. Jenkins, S.D. Campbell, D.H. Werner, Establishing exhaustive metasurface
[58] X. Han, Z. Fan, Z. Liu, et al., Inverse design of metasurface optical filters using robustness against fabrication uncertainties through deep learning,
deep neural network with high degrees of freedom, InfoMat 3 (2021) 432–442. Nanophotonics 10 (2021) 4497–4509.
[59] R. Pestourie, Y. Mroueh, T.V. Nguyen, et al., Active learning of deep surrogates [91] J. Zhang, C. Qian, J. Chen, et al., Uncertainty qualification for metasurface design
for pdes: application to metasurface design, npj Comput. Mater 6 (2020) 164. with amendatory bayesian network, Laser Photon. Rev. 17 (2023) 2200807.
[60] L. Huang, Z. Han, A. Wirth-Singh, et al., Broadband thermal imaging using meta- [92] J. Yue, T. Gao, W. Zhang, et al., Velocity-adaptive electrohydrodynamic printing
optics, Nat. Commun. 15 (2024) 1662. for microscale conformal circuits on freeform curved surfaces, ACS Appl. Mater.
[61] W. Ma, Y. Xu, B. Xiong, et al., Pushing the limits of functionalitymultiplexing Interfaces (2025).
capability in metasurface design based on statistical machine learning, Adv. [93] A. Baucour, M. Kim, J. Shin, Data-driven concurrent nanostructure optimization
Mater 34 (2022) 2110022. based on conditional generative adversarial networks, Nanophotonics 11 (2022)
[62] F. Wang, G. Geng, X. Wang, et al., Visible achromatic metalens design based on 2865–2873.
artificial neural network, Adv. Opt. Mater 10 (2022) 2101842. [94] R. Paniagua-Dominguez, Y.F. Yu, E. Khaidarov, et al., A metalens with a near-
[63] C. Yeung, D. Ho, B. Pham, et al., Enhancing adjoint optimizationbased photonic unity numerical aperture, Nano letters 18 (2018) 2124–2132.
inverse design with explainable machine learning, Acs Photonics 9 (2022) [95] H. Iyer, Y. Vlasov, Ultra-high numerical aperture waveguideintegrated meta
1577–1585. beam shaper, Optica 11 (2024) 1750–1761.
[64] M. Choi, J. Park, J. Shin, et al., Realization of high-performance optical [96] Y. Zhou, C. Mao, E. Gershnabel, et al., Large-area, high-numericalaperture,
metasurfaces over a large area: a review from a design perspective, npj freeform metasurfaces, Laser Photon. Rev. 18 (2024) 2300988.
Nanophotonics 1 (2024) 31. [97] A. Wirth-Singh, J.E. Fröch, F. Yang, et al., Wide field of view large aperture meta-
[65] J. Skarda, R. Trivedi, L. Su, et al., Low-overhead distribution strategy for doublet eyepiece, Light. Sci. Appl 14 (2025) 17.
simulation and optimization of large-area metasurfaces, npj Comput. Mater 8 [98] H. Zheng, F. Yang, H.-I. Lin, et al., Compound metalens enabling distortion-free
(2022) 78. imaging, Engineering 45 (2025) 52–58.
[66] C. Kim, B. Lee, Torcwa: Gpu-accelerated fourier modal method and gradient- [99] Y. Liu, W.-D. Li, K.-Y. Xin, et al., Ultra-wide fov meta-camera with transformer-
based optimization for metasurface design, Comput. Phys. Commun 282 (2023) neural-network color imaging methodology, Adv. Photon. 6 (2024).
108552. [100] C. Yeung, J.-M. Tsai, B. King, et al., Multiplexed supercell metasurface design and
[67] G. Yoon, J. Rho, Maxim: Metasurfaces-oriented electromagnetic wave simulation optimization with tandem residual networks, Nanophotonics 10 (2021)
software with intuitive graphical user interfaces, Comput. Phys. Commun 264 1133–1143.
(2021) 107846. [101] C. Spägele, M. Tamagnone, D. Kazakov, et al., Multifunctional wideangle optics
[68] J.L. Su, Z.X. Cai, Y. Mao, et al., Multi-dimensional multiplexed metasurface for and lasing based on supercell metasurfaces, Nat. Commun. 12 (2021) 3787.
multifunctional near-field modulation by physics-driven intelligent design, Adv. [102] T. Wu, X. Zhang, Q. Xu, et al., Heterogeneous-gradient supercell metasurfaces for
Sci. (2025) 2503899. independent complex amplitude control over multiple diffraction channels, Small
[69] Y. Deng, K. Fan, B. Jin, et al., Physics-informed learning in artificial 21 (2025) 2407303.
electromagnetic materials, Appl. Phys. Rev. 12 (2025). [103] M. Chen, R. Lupoiu, C. Mao, et al., High speed simulation and freeform
[70] R. Peng, S. Ren, J. Malof, W.J. Padilla, Transfer learning for metamaterial design optimization of nanophotonic devices with physics-augmented deep learning,
and simulation, Nanophoton. 13 (2024) 2323–2334. ACS Photon. 9 (2022) 3110–3123.
[71] R. Zhu, T. Qiu, J. Wang, et al., Phase-to-pattern inverse design paradigm for fast [104] S.S. Panda, R.S. Hegde, A learning based approach for designing extended unit
realization of functional metasurfaces via transfer learning, Nat. Commun. 12 cell metagratings, Nanophotonics 11 (2022) 345–358.
(2021) 2974. [105] Y. Jia, C. Qian, Z. Fan, et al., In situ customized illusion enabled by global
[72] T. Lookman, P.V. Balachandran, D. Xue, R. Yuan, Active learning in materials metasurface reconstruction, Adv. Funct. Mater 32 (2022) 2109331.
science with emphasis on adaptive sampling using uncertainties for targeted [106] D. Seo, D.W. Nam, J. Park, et al., Structural optimization of a one-dimensional
design, npj Comput. Mater 5 (2019) 21. freeform metagrating deflector via deep reinforcement learning, Acs Photon. 9
[73] H. Wei, X. He, W. Cao, Spin-multiplexed metasurface inverse design based on a (2021) 452–458.
bi-directional deep neural network for terahertz wavefront control, Optica 12 [107] Y. Ha, Y. Luo, M. Pu, et al., Physics-data-driven intelligent optimization for large-
(2025) 505–517. aperture metalenses, Opto-Electr. Adv. 6 (2023).
[74] C. Park, Y. Jeon, J. Rho, 36-channel spin and wavelength co-multiplexed [108] T.B. Kanmaz, E. Ozturk, H.V. Demir, C. Gunduz-Demir, Deep-learning-enabled
metasurface holography by phase-gradient inverse design, Adv. Sci. (2025) electromagnetic near-field prediction and inverse design of metasurfaces, Optica
2504634. 10 (2023) 1373–1382.
[75] W.-L. Hsu, C.-Y. Yu, Y.-S. Huang, et al., Simplest but efficient design of a color [109] O. Wu, C. Qian, Z. Fan, et al., General characterization of intelligent metasurfaces
router optimized by genetic algorithms, ACS Photon. (2025). with graph coupling network, Laser Photon. Rev. 19 (2025) 2400979.
[76] H. Gao, X. Huang, Z. Wang, et al., Two-dimensional anomalous reflection with [110] D. Sell, J. Yang, S. Doshay, et al., Large-angle, multifunctional metagratings based
high efficiency and arbitrary direction based on a low-profile wideband on freeform multimode geometries, Nano letters 17 (2017) 3752–3757.
metasurface, Photon. Res. 13 (2025) 1165–1171. [111] J. Kim, J. An, W. Kim, et al., Large-area floating display with wafer-scale
[77] C. Kang, D. Seo, S.V. Boriskina, H. Chung, Adjoint method in machine learning: a manufactured metalens arrays, Laser Photon. Rev (2024) 2401425.
pathway to efficient inverse design of photonic devices, Mater. & Des. 239 (2024) [112] Z.-B. Fan, Y.-F. Cheng, Z.-M. Chen, et al., Integral imaging near-eye 3D display
112737. using a nanoimprint metalens array, eLight 4 (2024) 3.
[78] M. Mansouree, A. McClung, S. Samudrala, A. Arbabi, Largescale parametrized [113] Y. Wang, H. Luo, Y. Shao, et al., Detection and anti-detection with microwave-
metasurface design using adjoint optimization, Acs Photon. 8 (2021) 455–463. infrared compatible camouflage using asymmetric composite metasurface, Adv.
[79] W. Ma, F. Cheng, Y. Xu, et al., Probabilistic representation and inverse design of Sci. (2024) 2410364.
metamaterials based on a deep generative model with semi-supervised learning [114] C. Dai, T. Liu, D. Wang, L. Zhou, Multiplexing near-and far-field functionalities
strategy, Adv. Mater 31 (2019) 1901111. with high-efficiency bi-channel metasurfaces, PhotoniX 5 (2024) 11.
[80] Z.A. Kudyshev, A.V. Kildishev, V.M. Shalaev, A. Boltasseva, Machine-learning- [115] L. Chen, Y. Yu, X. Zhang, Imaging spectropolarimeter using a multifunctional
assisted metasurface design for high-efficiency thermal emitter optimization, metasurface, Nano Lett (2024).
Appl. Phys. Rev 7 (2020). [116] S. Molesky, Z. Lin, A.Y. Piggott, et al., Inverse design in nanophotonics, Nat.
[81] H. Yang, C. Yang, H. Li, Enhancing high-degree-of-freedom metaatom design Photon. 12 (2018) 659–670.
precision and speed with a tandem generative network, ACS Photon. (2025). [117] J. Jiang, M. Chen, J.A. Fan, Deep neural networks for the evaluation and design of
[82] Z. Wang, C. Qian, Y. Saifullah, et al., Robust large-scale metasurface inverse photonic devices, Nat. Rev. Mater 6 (2021) 679–700.
design using phaseless data with invertible neural network, IEEE Trans. on [118] P. Dainese, L. Marra, D. Cassara, et al., Shape optimization for high efficiency
Antennas Propag (2024). metasurfaces: theory and implementation, Light. Sci. & Appl 13 (2024) 300.
[83] Q. Zhang, D. Guo, C. Shen, et al., Customizing the electric field of metalens with [119] S. An, C. Fowler, B. Zheng, et al., A deep learning approach for objective-driven
high degrees of freedom based on neural network, Nanophotonics (2025). all-dielectric metasurface design, Acs Photon. 6 (2019) 3196–3207.
[84] F. Wen, J. Jiang, J.A. Fan, Robust freeform metasurface design based on [120] R. Zhu, J. Wang, X. Fu, et al., Deep-learning-empowered holographic metasurface
progressively growing generative networks, Acs Photon. 7 (2020) 2098–2104. with simultaneously customized phase and amplitude, ACS Appl. Mater. &
[85] S. An, B. Zheng, H. Tang, et al., Multifunctional metasurface design with a Interfaces 14 (2022) 48303–48310.
generative adversarial network, Adv. Opt. Mater 9 (2021) 2001433.

17
Y. Zeng and X. Jin iOptics 1 (2025) 100004

[121] W. Zhou, S. Wang, Q. Wu, et al., An inverse design paradigm of multifunctional [139] J. Zhou, F. Tian, J. Hu, et al., Eagle-eye inspired meta-device for phase imaging,
elastic metasurface via data-driven machine learning, Mater. & Des. 226 (2023) Adv. Mater 36 (2024) 2402751.
111560. [140] J. Zhou, A. Li, M. Lei, et al., Advanced quantitative phase microscopy achieved
[122] Y. Bao, H. Shi, R. Wei, et al., Efficient gradient-based metasurface optimization with spatial multiplexing and a metasurface, Nano Lett (2025).
toward the limits of wavelength-polarization multiplexing, Nano Lett (2025). [141] M.K. Chen, X. Liu, Y. Wu, et al., A meta-device for intelligent depth perception,
[123] E. Zhu, Z. Zong, E. Li, et al., Frequency transfer and inverse design for metasurface Adv. Mater 35 (2023) 2107465.
under multi-physics coupling by euler latent dynamic and data-analytical [142] J. Wei, H. Huang, X. Zhang, et al., Neural-network-enhanced metalens camera for
regularizations, Nat. Commun. 16 (2025) 2251. high-definition, dynamic imaging in the long-wave infrared spectrum, ACS
[124] L. Li, Y. Shuang, Q. Ma, et al., Intelligent metasurface imager and recognizer, Photon. (2025).
Light. Sci. App. 8 (2019) 97. [143] Y. Zhang, L. Zhang, Y. Zhu, et al., High-spatial and spectral resolution snapshot
[125] Z. Fan, C. Qian, Y. Jia, et al., Homeostatic neuro-metasurfaces for dynamic spectral imaging based on nearfield coupled metasurfaces, ACS Photon. (2025).
wireless channel management, Sci. Adv. 8 (2022) eabn7905. [144] J.E. Fröch, P. Chakravarthula, J. Sun, et al., Beating spectral bandwidth limits for
[126] C. Liu, Q. Ma, Z.J. Luo, et al., A programmable diffractive deep neural network large aperture broadband nano-optics, Nat. Commun. 16 (2025) 3025.
based on a digital-coding metasurface array, Nat. Electron 5 (2022) 113–122. [145] Y. Park, Y. Kim, C. Kim, et al., End-to-end optimization of metalens for broadband
[127] B. Wu, T. Liu, G. Wang, et al., Integrated electromagnetic sensing system based on and wide-angle imaging, Adv. Opt. Mater 13 (2025) 2402853.
a deep-neural-network-intervened genetic algorithm, Photon. Res. 13 (2025) [146] S. Fisher, G. Arya, A. Majumdar, et al., End-to-end metasurface design for
387–394. temperature imaging via broadband planck-radiation regression, Adv. Opt. Mater
[128] F. Li, T. Pan, W. Li, et al., Flexible intelligent microwave metasurface with shape- (2025) 2402498.
guided adaptive programming, Nat. Commun. 16 (2025) 3161. [147] H. Chi, Y. Hu, X. Ou, et al., Neural network-assisted end-to-end design for full
[129] F. Lan, L. Wang, H. Zeng, et al., Real-time programmable metasurface for light field control of meta-optics, Adv. Mater (2025) 2419621.
terahertz multifunctional wave front engineering, Light. Sci. Appl 12 (2023) 191. [148] Y. Yin, Q. Jiang, H. Wang, et al., Multi-dimensional multiplexed metasurface
[130] S. Sun, W. Yang, C. Zhang, et al., Real-time tunable colors from microfiuidic holography by inverse design, Adv. Mater 36 (2024) 2312303.
reconfigurable all-dielectric metasurfaces, ACS nano 12 (2018) 2151–2159. [149] H. Xia, B. Chen, C. Zhang, et al., Joint optimization of coded aperture metasurface
[131] X.G. Zhang, Q. Yu, W.X. Jiang, et al., Polarization-controlled dual-programmable and residual self-attention network for snapshot full-stokes imaging, Opt. Express
metasurfaces, Adv. Sci. 7 (2020) 1903382. 32 (2024) 29609–29619.
[132] X.G. Zhang, Y.L. Sun, B. Zhu, et al., Wireless microwave-to-optical conversion via [150] J. Xi, J. Shen, M.T. Chow, et al., Deep-learning assisted polarization holograms,
programmable metasurface without dc supply, Nat. Commun 16 (2025) 528. Adv. Opt. Mater 12 (2024) 2202663.
[133] H. Wu, R. Shao, Z. Xu, et al., A programmable metasurface antenna that [151] G. Arya, W.F. Li, C. Roques-Carmes, et al., End-to-end optimization of
approaches the wireless information mapping limit, Nat. Electron (2025) 1–13. metasurfaces for imaging with compressed sensing, ACS Photon. 11 (2024)
[134] L. Li, H. Ruan, C. Liu, et al., Machine-learning reprogrammable metasurface 2077–2087.
imager, Nat. Commun. 10 (2019) 1082. [152] R. Zhu, J. Wang, T. Qiu, et al., Direct field-to-pattern monolithic design of
[135] H. Shi, Y. Sun, Z. Liang, et al., Feature-enhanced x-ray imaging using fused neural holographic metasurface via residual encoder-decoder convolutional neural
network strategy with designable metasurface, Nanophotonics 12 (2023) network, Opto-Electronic Adv 6 (2023).
3793–3805. [153] E. Tseng, S. Colburn, J. Whitehead, et al., Neural nano-optics for high-quality thin
[136] S. Tan, F. Yang, V. Boominathan, et al., 3D imaging using extreme dispersion in lens imaging, Nat. Commun. 12 (2021) 6493.
optical metasurfaces, ACS Photon. 8 (2021) 1421–1429. [154] P. Chakravarthula, J. Sun, X. Li, et al., Thin on-sensor nanophotonic array
[137] X. Li, B. Shi, Q. Jia, et al., Diffractive metasurface-enabled single-shot cameras, ACM Trans. on Graph. ((TOG)) 42 (2023) 1–18.
characterization of hybrid-order poincaré beams, Adv. Opt. Mater. (2025) [155] Z. Long, Y. Zeng, X. Jin, End-to-end design of the metalens array imaging system
2403405. with extended depth of field, Optics Express 33 (13) (2025) 26869–26886.
[138] H. Kwon, E. Arbabi, S.M. Kamali, et al., Computational complex optical field
imaging using a designed metasurface diffuser, Optica 5 (2018) 924–931.

18