 Review
 Open Access
 Published:
Intelligent designs in nanophotonics: from optimization towards inverse creation
PhotoniX volume 2, Article number: 22 (2021)
Abstract
Applying intelligence algorithms to conceive nanoscale metadevices becomes a flourishing and extremely active scientific topic over the past few years. Inverse design of functional nanostructures is at the heart of this topic, in which artificial intelligence (AI) furnishes various optimization toolboxes to speed up prototyping of photonic layouts with enhanced performance. In this review, we offer a systemic view on recent advancements in nanophotonic components designed by intelligence algorithms, manifesting a development trend from performance optimizations towards inverse creations of novel designs. To illustrate interplays between two fields, AI and photonics, we take metaatom spectral manipulation as a case study to introduce algorithm operational principles, and subsequently review their manifold usages among a set of popular metaelements. As arranged from levels of individual optimized piece to practical system, we discuss algorithmassisted nanophotonic designs to examine their mutual benefits. We further comment on a set of open questions including reasonable applications of advanced algorithms, expensive data issue, and algorithm benchmarking, etc. Overall, we envision mounting photonictargeted methodologies to substantially push forward functional artificial metadevices to profit both fields.
Background
With a research focus on lightmatter interactions [1, 2], the remarkable progress in nanophotonics has been made over the last two decades, leading to the fruitful innovations such as photonic crystals [3, 4], plasmonic [5, 6], and metamaterials [7]. Owing to their exceptional electromagnetic (EM) features originating from the unique geometric distributions, artificial nanostructures allow the manipulation of light at subwavelength scale and even drive fictional concepts (e.g., invisibility cloaks [8, 9]) into reality. However, the optimal structure design is an iterative and timeconsuming process, in which the humaninvolved trialanderror strategy combined with the parameter sweep method is usually adopted [10]. A typical workflow initializes from manguessed or prior verified structural layouts, and subsequently solves Maxwell’s equations or analytical formulas to update photonic geometries [11]. Until optical responses satisfy object demands, these procedures are consistently repeated, taking hours to days for a final output. Apparently, such low design efficiency hinders the rapid growth of nanooptical devices, and the limited human brainpower is characteristically weak in terms of reasoning hyperdimensional parameter space, that is commonly found in complicated structuredesign problems [12].
To tackle these challenges in nanophotonics, using artificial intelligence (AI) techniques to design the sophisticated metadevices emerges as a promising approach that could assist us to bust the mentioned bottlenecks (e.g., inefficient design process). As the term indicated, AI commonly refers to programmable computerenabled imitations of human intelligence. In short, the idea of broad AI came firstly, and machine learning (ML, a subfield of AI) blossomed later. The latest stage of deep learning (DL, a subfield of ML) drive today’s AI explosion [13, 14].
The cornerstone of AI techniques lies in varied algorithms, and here we discuss AI as different types of artificial algorithms that could help us to perform nanostructure designs. Up to date, there could be generally classified as two stages of learning algorithms applied for metastructures designs, identified as traditional optimization techniques [10, 15–17] and latest DL models [12, 18–20], respectively. Historically speaking, traditional optimizations caught public attention in the late 1990s by several initial work on wavelengthscale optical structures [21, 22] (e.g., a fibertowaveguide coupler [21]), and hereafter the notable extensions (e.g., a ‘Z’like photonic crystal waveguide [23] and a Yshaped splitter [24]) were implemented over the past two decades, enriching the library of nanodevice templates (see detailed roadmaps about photonic designs in refs. [20, 25]). In the past couple of years, there is an exponential rise to explore DL potential in artificial photonic structures, as some notable works illustrated in cases of broadband beamsplitters [26], chiral metamaterials [27], optical neural networks [28], and so on. All in all, the growing investigations mentioned here shows that two stages of learning algorithms could enhance optical performance of nanocomponents and, in particular, to inversely design nanostructures on demand.
Here we summarize the latest research on this interdisciplinary topic covering associated publications in the past three years, and we offer a systemically viewpoint on functional metacomponents to inspect their optical performances improved by learning algorithms. As shown in Fig. 1, the review is arranged into two units: learning algorithms (highlighted by two gray shadows on the left) and metacomponents (shown in the colorcoded sector on the right). On the one hand, learning algorithms include traditional optimizations (such as genetic algorithm (GA), particle swarm optimization (PSO), and topological optimization (TO)) together with neural network (NN)based DL techniques (e.g., multilayer perception (MLP), generative adversarial network (GAN), variational autoencoder (VAE) and double deep Qlearning (DQN)). Note that many works discuss that only neural networkbased ML and DL techniques are considered to be AIrelated, while other references classify traditional optimizations (e.g., GA) as AIassociated techniques. To clear this point, in this review we employ a more general concept of ‘intelligence algorithms’ to refer to all methods shown in Fig. 1. On the other hand, targeted metacompenents are listed, namely, metalens, metagrating, beam splitter, onchip coupler, optical interference unit (OIU), optical diffractive neural network (ONN), etc. To illustrate the interplay of two distinctive study fields, we present the abovementioned from the individual patterns (e.g., metagrating) capable of manipulating incident light to the complex devices (e.g., ONN) performing advanced tasks like image processing.
The Table 1 further generally illustrates modules and their characteristics related to Fig. 1. As an extension of diagram, the table summarizes important concepts and highlights references on metaatom designs. Also, it works as an outline for subsequent algorithm introductions providing certain guidelines for the choice of appropriate approaches. Due to a lack of opensource database, to reasonably characterize, compare, and benchmark various algorithms performances becomes unrealistic, and thus it is rather difficult to suggest any specific methods to conduct nanostructure designs. In general, parametersweeping is a straightforward strategy for the narrow degree of freedom (DoF) space (e.g., < 5 variables), and traditional optimizations could be considered for mediumsized design tasks (e.g., ≈ tens variables). Here, global optimizers (e.g., GA) are recommended for relatively lowdimensional situations, while local optimization method (e.g., TO) is a suitable candidate for highdimensional problem under highconstrained situations. The training of DL models requires large scale raw data (at least thousands of labels) and demands more computing resources. Hence this is suggested for the much higher DoF, in particular, for inverse design problems.
The review proceeds as follow. Over the second part, we shortly present typical algorithms and then we discuss the progress on the abovementioned metacomponents throughout the third section. After that, the current dilemmas in intelligent designs are summed up, and the future interplays between nanophotonics and AI research are given in the end. Differing from previous literature (see these reviews [10, 12, 15–20, 45–48], tutorials [49–51], and perspectives [52–55]), we deliver an intuitive instruction of GA and MLP algorithms by case studies of spectrum tunings of metaatoms (see an introduction on metaatom’s physical properties in ref. [56]), that is a wellknown example in the field of nanophotonics. Here we also extended coverage range to include two stages of intelligence algorithms and categorized metastructure designs for a fairly report on comprehensive advancements of this topic. Additionally, for the reference of newcomers, we summarize available opensource code packages with EM simulators for a quick start in this blooming direction (see Table 2 in appendix). With all these efforts, we hope to provide readers with fundamental knowledge alongside practical toolbox to involve the intensive innovations.
Intelligence algorithms with case study in metaatom design
Composed of Traditional optimization techniques, Deep learningbased techniques, Reinforcement learning and Opensource packages, this section serves to bridge study fields between AI and metastructure. How to employ intelligent algorithms to improve nanophotonic properties (e.g., spectral behaviors) is demonstrated through optimizations of metaatoms. As a fundamental building block, the plasmonic metaatom commonly refers to the subwavelength artificial particle exhibiting extraordinary optical properties arising from its material and geometries [56]. To distinguish applications varying from specific metaatoms to general metastructures, this section concentrates on working principles of design algorithms, whereas functional metadevices involving more complex geometries will be elaborately presented in the next section. Note that the wellorganized nanopatterns (e.g., diffractive optical neural networks [28]) could perform more advanced computation, while isolated simpler metaatoms fail to fulfill such intricate tasks.
In general, the primary mission of nanostructure designs is to optimize the spatial distribution of candidate types of composite materials (e.g., metallic and/or dielectric), with an aim to best assure the ondemand performances subjected to all needed constraints (e.g., nanofabrications). In pursuing this goal, two design paradigms (i.e., traditional technique and contemporary DL one) exhibit both pros and cons. A traditional algorithm repetitively updates a structurerelated object (defined by a number of variables) to search for target outputs. However, only slight enhancement or even no visible enhancement may be obtained over lengthy optimizations, with no guarantee to approach the optimal geometry.
In contrast, bigdatadriven DL methods could spare immense computational efforts needed for traditional designs. Instead, these stateoftheart strategies utilize given nanopatterns to train deep neural network (DNN) models in order to fit the inputoutput relationship of original metapatterns. Although model trainings could consume a long period of time, the learned neural networks are able to map the geometrical information of functional component to its optical response and vice versa in a second [11]. Note that despite of the mentioned benefits, the network training could be rather timely and expensive due to multiple facts such as computational resources, raw data quality, etc. And it is still believed that the trained networks are unable for structure reasoning beyond the original physical scenarios. Hence, the selection of a proper algorithm should overall consider abovediscussed strengths and weaknesses of both design strategies.
Although the instant rush on AI techniques unboxes a multitude of powerful instruments for nanostructure designs, the doubts of the rapid development are also raising recently. First of all, the performance (e.g., design efficiency) of gradient and gradientless algorithms have not been widely discussed, fairly compared, and solidly examined yet. Secondly, for traditional optimization techniques like GA, the final outputted results may be just iterated layouts with slightly properties enhancement. Especially facing hyper nonlinear design space, TO is assumed to be more efficient compared to GA (see a study in Ref. [51]). At last, several potential factors (e.g., the high cost of nanopattern generations) may fundamentally query the usage of advanced DL models. Critics have argued whether to apply DL model is in doubt under certain scenarios (e.g., high DoF) since the wellestablished TO could already better cope with such situations. These questions are left open for future studies, and we would like to comment on parts of them in the last section of the review.
Figure 2 explains the relationships between metaatoms and learning algorithms. Four types of optical particles (circle, square, triangle, and rectangle) are patterned on a substrate and the spectral modification is shown on the right. The metaatoms, separated by the dashed lines, exhibit distinctive spectral features imposed by their variations in shapes. Aided by learning algorithms, the target spectrum modifications (i.e., to sharpen and/or shift the gaussian lineshape) set solid basis for extended applications like optical filters and color generations. Note that, multiple types of spectral data, including e.g., transmittance, reflectance, absorption, and circular dichroism (CD) could be individually altered or cooptimized. Commonly, the layouts of metaatoms are processed as pixelwise image matrices or structural parameter vectors for input/output of learning algorithms. For illustration purpose, we only display necessary variables to define metaatom geometries (e.g., the length and width of rectangle atoms) and spectral information. Note that, the variables could be further extended to include such as hyperparameters (e.g., lattice constants) or higherresolution image matrix. Thereby, the increased variables dramatically improve DoF that may produce highperformance metastructures, while the parameter expansions also bring a number of problems e.g., longer time of model convergence.
In the middle of diagram 3, the widely covered intelligence algorithmsassociated techniques are presented as traditional optimizer and deep learning, respectively, where the electromagnetic simulators (highlighted in the light green) take charge for cost functions and DLmodel data preparations. As AI techniques burst, amounts of algorithms have been proposed and applied spanning past several decades. Within the scope of nanophotonics, here we spotlight two typical approaches (i.e., GA in Fig. 3 and MLP in Figs. 4 & 5) as corresponding representatives, and other algorithms (i.e., TO, VAE, and GAN) would be generally narrated as well. Two specific approaches have been consistently reported in broad designs of metastructures, and we explain key steps packed with easytounderstand schematics. Principally, GA is a metaheuristic global optimizer widely reported in property optimizations, and MLP is a learning model capable of inversely nanostructure designs. Due to the review interest, more problems occurred over algorithm usages could be looked up in these documents [57–59]. Note that, GA and MLP may fail to reach the optimal designs of functional metastructures and thus maybe not best fit for certain occasions. Especially taken high DoF into account, alternative effective methodologies (e.g., TO) could be particularly considered. Detailed stepbystep tutorials of TO can be obtained from here [51, 60, 61].
Traditional optimization techniques
Within this section, we firstly provide a detailed establishment process of genetic algorithm and then present two additional approaches (i.e., PSO and TO). In brief, usually with a starting point of wellestablished solutions, traditional optimization algorithms define an objective function or a figure of merit (FoM) to tailor the form of nanostructures in an iterative process. Note that computational costs of objective functions are closely related to the model complexity, applied constraints and design degrees of freedom. For a proper choice of global or local optimizers, these references [10, 15, 62] can be examined in depth.
Genetic algorithm
Genetic algorithm is a gradientfree optimization method mimicking natural concept of selection and genetics [15]. A typical operational diagram of the GA is set out in Fig. 3 (a).
The first step is to create initial populations containing underlying physics of metaatoms. Here the metaatom is firstly recorded as a binary N ×N pixelwise image and additionally represented by a N ×N dimensional matrix (defined as an individual of X with unity of 1 and void of 0). To illustrate, the transmission spectrum T (as a function of individual X) of the metaatoms is computed by EM simulators. Hence the fitness function F can be evaluated as the mean square error (MSE) between calculated value T(X) and target spectral intensity G. i.e., F=MSE(T(X),G).
After the fitness assessment, GA repeats steps of selection, crossover and mutation until the termination criteria are fulfilled. In principle, the GA picks up these ‘elite’ individuals (e.g., higher fitness scores) to construct the next generation by operators of crossover and mutation. Note that the mutation procedure is likely to maintain the population diversity and prevent premature convergence. The algorithm can be ended once the fitness scores reach an appropriate threshold (e.g., the MSE is less than the demanded value) and/or the finite iterations run out.
Figure 3 (b) shows a GA methodology for broadband absorption optimization of a single metaatom [29]. Organized in a fixed lattice, a gold nanodisk was placed on a gold substrate coated with a SiO_{2} middlelayer. The unit cell pitch, the disk height alongside diameter, and the thickness of middlelayer were those participating variables. In the reference [29], the fitness function was defined as \(F= 1\bar {A}\), with \(\bar {A}\) being the averaged absorption between spectral range of 350nm and 800nm. An increasing trend of \(\bar {A}\) was clearly observed for both best and average value of absorption, and the best value of \(\bar {A}\) (red diamonds) reached a plateau after about twenty generations. Two illustrative metaatom shapes (upper panel) together with their absorptances (lower panel) at a generation of 0 and 25 denoted prominent value enhancements from about 60% (red dashed line) to around 80% (black dashed line). In addition, the authors surveyed how to apply GA to enlarge the color coverage and to generate structural colors by increasing the pitch size and adding more metaatoms in one cell [29].
So far, various studies on spectral regulations by GA have been conducted among cases of alldielectric magnetic nanoantennas [30], infrared binarypattern absorbers [31], reconfigurable metaatoms [32] and strong circular dichroism of chiral patterns [33], etc. These works substantially extend the usage of GAs for metaatoms modifications and pave the way for further investigations in GAassisted photonic structure design.
In spite of abovementioned work, further critics question the efficiency of GA dealing with largescale design variables. Firstly, GA requires a set of solutions. It is unfriendly for brandnew usages since the prior verified layouts are exceptionally rare. Secondly, over numerous function evaluations, GA is computationally expensive and timeconsuming. The high cost also denotes that GAembedded parameters (e.g., mutation rate) need huge efforts to be finely tuned. Last but not least, as we previously discussed, there is no guarantee to identify global maxima. In extreme situation, the final outputs maybe merely the updated patterns, which lack substantial property enhancement. Hereby, alternative approaches (e.g., TO) are more advisable for high DoF circumstance.
Particle swarm optimization
Inspired by the animal social behaviors akin to bird flocking and fish schooling, PSO is an evolutionary computation method which is strongly linked to GA and evolutionary programming [63, 64]. A standard working flow of PSO starts from a set of randomly distributed particles. Until the system converges (or a certain termination criterion is satisfied), these particles keep searching for better positions by remembering and sharing their best location achieved so far [15]. An example of PSO is the study showcased by Forestiere [34] to optimize an array of 55×55 plasmonic nanospheres with the goal of creating field enhancement across the 400900 nm spectral range. The process of binary PSO was eventually stopped after 100 iterations, upgrading the maximum field improvements to a value of 35.9 (at 500 nm). While the associated values for periodic particles and single particle both approximated to 6, being almost 6 times smaller than the PSOamended arrays. This behavior comparison implies PSO as an effective method for more general layoutdevelopment problems within functional nanostructures.
Topology optimization
TO is often taken as an inverse design technique [49] expansively studied in engineering fields within fine mechanics, aerospace, and building architectures, etc. By combining with gradient and/or nongradient algorithms, it is a highly efficient tool to address largescale structure design problems with potentially up to millions of DoF [17]. From the functionality’s aspects (i.e., forward prediction and inverse design), a main difference compared to the undermentioned DLmodels is that TO cannot perform the forward prediction functionality (i.e., to directly project nanostructural optical properties with no need of any simulators). However, TO exhibits enormous advantages and high working efficiency especially dealing with freeform nanophotonics under high DoF situations (potentially billions of DoF [35]). For instance, a metalensbased study reveals that the computational cost by TOidentified local optimum structure required merely ≈1.0% in comparison to counterparts via GA calculations [51].
Overall, with an initial guess, TO seeks to maximize the geometryassociated related FoM (Φ) over a series of iterations. Assuming that the distribution of metaatom is reasonably arranged to obtain high transmission T at wavelengths of λ, the FoM could be defined as \(\underset {\epsilon }{\max }\: \Phi (T(\lambda), \epsilon)\), where ε denotes design variables (e.g., material distributions) within the assigned field [65]. Commonly utilizing the adjoint variables method, the gradients of the FoM Φ provide a guide for iteratively modifying these design variables ε in order to search for a local minimum (or maximum) of FoM [10]. Note that the last outputs may be significantly influenced by the initial guess if the FoM value fails to be dramatically improved. For stepbystep tutorials on TO, we would suggest these materials [35, 51, 60, 66, 67] for indepth examinations.
Recently TO for design and optimization of nanophotonic components becomes a soaring trend. Up to date, there exist a large number of reports to investigate TO applications in nanophotonics due to its significant benefits discussed above. Confined to metaatoms, a good illustration of TO for the metaatom adjustments could be found in Christiansen work [35], where TOmodified metallic resonators showed over two order’s improvement in surfaceenhanced Raman scattering (SERS) efficiency compared to a parameteroptimized bowtie antenna. Here the physical model was numerically simulated in COMSOL Multiphysics and the TO problem was executed by a builtin algorithm titled Globally Convergent Method of Moving Asymptotes.
Deep learningbased techniques
As arranged in a consequence of ‘multilayer perceptron’, ‘autoencoder and variational autoencoder’ and ‘generative adversarial network’, this section focuses on DL models to examine how the advanced algorithms could inversely generate functional nanopatterns on demand. At first, we would like to introduce fundamentals on neural network learning model as illustrated in Fig. 4. The diagram of (a) shows a simple NN consisted of an input layer (‘X’, green circles on the right), an output layer (‘Y’, green circles on the left) and two hidden layers (X^{(l−1)}& X^{(l)}, dark blue circles in the middle). In principle, the neural network could perform a nonlinear transformation from the input X to the output Y. Given training data, deep learning is used to learn network parameters (e.g., weights (indicated by connected black strips) and biases) that characterize the transformation. Specifically in lower panel of Fig. 4b, the artificial neuron layer transforms its input X^{(l−1)} to the output X^{(l)} by X^{(l)}=f(X^{(l−1)}W^{(l−1)}), where f is the nonlinear activation function (e.g., sigmoid) and W^{(l−1)} is the transformation matrix of weights. For optical component designs, the transformation from input to output is an also a function of a given metastructure S so that the output could be interpreted as a function of X^{(l)}=f(X^{(l−1)}W^{(l−1)}S).
A general training procedure of NN for metaatom spectral tuning is provided in Fig. 4 (c), where four numbered steps start from raw data, training and validation, model test, to new applications. At beginning, raw data are prepared including nanostructure geometry (represented by an image matrix or a parameter vector) and its spectral distribution (calculated by simulators). The obtained labels would be further divided into three groups for training, validation, and test usages, respectively. Secondly, a NN model with initial parameters of weights W is established and thus learning procedure is conducted so as to obtain an optimal matrix W. Here, a common loss function L(W) is defined as L(W)=MSE(f(Y_{training})−Y_{ground_truth}), where f(Y_{training}) and \(Y_{ground\_truth}\) denote current model outputs and original value, respectively. To search for the minimum value of loss function, the gradient stochastic descent (\(\frac {\partial L}{\partial W}\)) is usually adopted. Here, an ideal lineshape of L(W) (presented as a dark green line in second frame of Fig. 4 c) gradually declines as training epochs, reaching a minimum value where the model converges. Furthermore, the learned model should be examined by the validation data in order to give an unbiased estimate of NN performance (see the ideal light green line). After the first two steps, the trained NN will be finally checked using test data. Once it passes the task, the NN could be utilized as a highly efficient tool for new structure designs.
Multilayer perceptron
As illustrated by an exemplary case [11] shown in Fig. 5 (a), the bidirectional multilayer perceptron model [19] (originally discussed as DNN model) was composed of two subnetworks, namely, a spectrumpredictingnetwork (SPN) and a geometrypredictingnetwork (GPN), respectively. The former could anticipate the nanoparticle’s spectrum without any assistance of simulators, and the latter was assumed to execute inverse design functionality (i.e., outputting the nanostructure geometry given the input spectrum). The construction of such a bidirectional model typically goes as follows:
(A) Building physical models of metaatoms. This step is aimed to set physics behind lightnanostructure interactions under diverse experiments. Each experiment encompassed, but not limited to, a plasmonic metaatom with parameterized geometry, the environmental permittivity and the incident light field. Here various Hshaped gold antennas (see a sketch in Fig. 5 (b)) could be described by eight values: three continuous parameters (L_{0},L_{1} and rotation angle ϕ) and five binary ones (leg 1 to leg 5). Some randomly generated populations are presented within the inner box of step A. Alternatively, these metaatoms could be taken as pixelwise images for more general treatments (see the illustration in Fig. 3 (a)).
(B) Obtaining spectral information of metaatoms. For such Hlike metaatoms, the calculations of broadband spectra require wellestablished EM simulators to conduct fullwave numerical simulations. In particular, this stage could be extremely timeconsuming for a largescale data (at least several thousand spectra). For the choices of EM simulators and DL model training approaches, we will discuss the potential solutions and sum up available opensourced ones in appendix.
(C) Training DL models using the dataset prepared by the simulator. Overall, the MLP was based on fully connected NNs, with one submodel GPN being made of three fully connected layers and another submodel SPN employing eight connected layers. The input of GPN was composed three groups of data: two vectors of 43 samples (two polarizations spectra) and a vector of 25 parameters (material’s properties). While the output of eight neurons in GPN encoded the predicted geometry. For the SPN, the input layer received the eight output parameters from GPN, additionally with the materials’ properties and a flag (indicating the polarization). And the output layer of SPN was constituted by 43 neurons representing wavelength data points.
It was proposed to train both GPN and SPN simultaneously, so that two submodels could coadapt to each other. The complete dataset generated from previous two steps was categorized into three parts: 80% for model training, 5% for validation, and 15% for model testing. By a gradient descent optimization algorithm (i.e., Adadelta), the networks were trained to minimize the MSE between the predicted spectra and geometry to their ground truths. The process took around two hours and terminated at about 3000 epochs with a final MSE of 0.16.
(D) Metaatom design using the trained model. The authors employed the model to design a gold plasmonic structure targeted for the organic molecule dichloromethane. For this type of molecule, the required transmissions should vary with incident polarizations (e.g., vertical and horizontal directions). The guessed plasmonic structure and an elaborate configuration table are displayed at bottom of Fig. 5 (b). The spectral evidence exhibited a fine data match among desired response, DLestimated value and numerical simulated results, indicating the strong capability of neural network model to address various targeted resonances. Up to date, deep neural network models have been substantially explored for chiral metamaterials [27, 36, 37], accurate silicon color design [38], plasmonic spectral sensors [39], etc. These groundworks allow the ondemand design of optical responses of nanostructures and enable metasurfacelike structures for considerable utilization.
Autoencoder and variational autoencoder
Autoencoder (AE) is often regarded as an unsupervised learning neural network constituted of an encoder and a decoder [12]. Basically, the encoder compresses the input data as ‘code’ (alternatively referred as latent variables or latent representations), and the decoder maps the ‘code’ to reconstruct data that imitates original inputs. An applied case of AE for metaatom design can be examined in Hemmatyar’s research [40], in which Fanotype resonant HfO_{2} nanopillars (prime variables: height h, diameter d, and a squarelattice with periodicity p) were adjusted to generate a largescale highquality reflective color gamut. Here the AE transforms data in the input space of dimensionality to a lower dimensional latent space of dimensionality, reducing space dimensionality of the prepared reflectance data. And the output from AE was further fed into a pseudoencoder to extract the importance of each design parameter (h, d, or p).
Variational autoencoder (VAE) is a productive DL scheme with autoencoderlike architecture, while the mathematical basis of VAEs owns loose connections with classical AEs [68]. In fact, as a major kind of generative models, VAEs provide a probabilistic manner to describe an observation in latent space. A breakthrough using VAE for metaatoms is carried out by Ma et al. [41], who introduced a probabilistic graphic model by encoding metamaterial geographies and optical response (in total > 20000 data of reflectance and CD spectra) into a 20dimentional latent space. The network was able to answer both the forward and inverse problems simultaneously, and the clear separation of three shape groups (cross, split ring, and hshape metaatoms) could be discovered in a visualization of the latent space.
Generative adversarial network
Composed of a generator and a discriminator, the GAN is a recently developed unsupervised DL framework [12]. The generator is trained to produce more plausible images, while the discriminator learns to distinguish the generated fake graphs or real ones. As an outcoming, a welltrained GAN could yield numerous images resembling true data in a second. For a very first demonstration, Liu et al. [42] applied a modified GAN to inversely outline gold metaatom geometries with targeted optical responses ranging from 500 nm to 1800 nm. Consisted of three parts (namely a simulator, a generator, and a critic), the network (trained by 6500 sets data) granted generations of arbitrary patterns of periodical structures. To solve the instability problem of GANs, So et al. [43] inquired a conditional deep convolutional generative adversarial network (cDCGAN) to devise nanophotonic antennae, where the 500 nm × 500 nm physical domain was represented as a 64 × 64 pixel binary image (resulting in a 2 ^{64×64} DoF). The spectral mean absolute error of 12 test samples between a networkgenerated geometry and original one was only 0.0322, reflecting that the cDCGAN could archetype proper nanostructures with the desired reflection spectra.
Reinforcement learning
Reinforcement learning (RL) is the one of three machine learning paradigms and its selflearning mechanism is to teach AI agents how to behave in an uncertain environment [12]. To achieve a goal (e.g., the expected benefits are sought maximized), the agent interacts with the environment and gets either rewards or penalties for every action. By this trialanderror way, the agent obtains the socalled strategy policy and further applies it to determine the next movement. A remarkable benefit of RLs lies in the low amount requirements of training dataset since the agent could search the parameter environment independently, which is of vast interest for metaelements devising. However, unlike other types of DL methods, the RLassisted optical structure designs are less reported by far, among which the deep Qlearning network (DQN) manifests as an exemplary model. The DQNrelated articles probe its usages from color generations [44], ultrabroadband perfect absorbers [69] (motheye structures) to highly efficient metasurface holograms [70]. For more detailed illustrations on RL, we would like to draw readers’ attention to corresponding references.
Opensource packages
The high entry threshold and expensive learning costs of computer sciences are unfriendly to many newcomers from the optics community, particularly considering the broad knowledge gap between nanophotonic and artificial intelligence. Therefore, we summarize parts of available opensource code packages to help those longing for a rapid start (see the attached Table 2 in appendix). To employ these readytouse tools, more detailed tutorials (e.g., Ref. [19]) could be consulted and online selflearning platforms (such as Coursera) are also recommended for practical coding.
Functional metacomponents and devices
In this section, we will state the current trend in intelligent designs for applicationoriented optical components and practical devices. We roughly categorize these metastructures into subgroups of metalens, metagratings, beam splitters, onchip couplers, optical interference units, optical diffractive neural networks, and other applications (see Fig. 6 to Fig. 12). In each division, we are intended to arrange the traditional optimization algorithms firstly and then introduce the advanced DL schemes. Here we discuss the mutual benefits between AI and nanophotonics. On the one hand, as versatile tools, intelligence algorithms could evidently improve design efficiency of photonic system. On the other hand, wellarranged photonics systems can be employed as optical hardware to accelerate AIrelated tasks (e.g., object classifier empowered by ONN). After all, the intensive interplay of two field could generate novel photonic designs that are designed by AI and potentially used for AI.
Metalens
Capable of manipulating the phase distribution, metalens is increasingly applied in the practical applications [75]. Here we highlight some recent advances in Fig. 6. As illustrated in (a), Liang et al. [71] exploited a hybrid optimization algorithm (combining differential evolution, GA, PSO, and adaptive simulated annealing) to shape highperformance cSi nanobricksbased metalenses. The fused algorithm gradually converged after 43946 iterations, reaching a final FoM of 0.93. The high numerical aperture (NA with a value of 1.48) metalens (oil immersion) exhibited a 207 nm full width at half maximum (FWHM) of a beam spot with an operational efficiency of 48%, representing one of highest NA of any metalens by that time. Other types of traditional algorithms such as GA and binary search techniques are also demonstrated in, e.g., PancharatnamBerry type metalens [76] as well as multilevel diffractive lenses [77–79], respectively.
To fully modify large area metalens requires numerous computation power, and thus a set of approximate approaches are developed in order to decrease the calculation complexity. For instance, Phan et al. [72] proposed to approximate desired phase profile by a series of linear segments and further formed an aperiodic Fourier modal method for largescale metasurface lenses. Figure 6 (b) displays a linearizationapproachdesigned metalens achieving an NA of 0.8, where the experimental evidence coincided well with simulations, both showing concentrated beam waists of about 340 nm.
Another brand of widely employed algorithms for largearea metalens is topological optimizations [80]. As depicted in Fig. 6 (c), Christiansen et al. investigated a group of tunable multilayer metalens [73]. By switching the refractive index (n = 3.2 or 4.6+0.01i) of stacked phase change materials (GST41T1), the focused beam profile exported from the tenlayer metalens could be altered accordingly. More TOdesigned metalens can be checked in the references [17, 81, 82]. In addition, the general adjointbased optimizations [83, 84] have been reported by Mansouree et al. [74] for designing a doublewavelength nanopost metalens as shown in Fig. 6 (d). Here the interpost and interlayer coupling were comprehensively taken into consideration. The systemically designed dual λ metalens showed the neardiffractionlimited focusing ability and the intensity distributions captured two split spots with dimensions of 1.33 μm (wavelength: 780 nm) and 1.54 μm (wavelength: 915 nm), respectively.
Moreover, there is increasing attention on the DLassisted metalens design [84]. A specific example can be found from the Pestourie’s report [85], where they presented an activelearning algorithm to reduce at least one order of magnitude of the training time for the surrogate model. The demonstrated tenlayer metastructure (with 100 unitcells of period 400 nm) could converge light at three wavelengths (405 nm, 540 nm, and 810 nm) into three different focal spots, respectively. This surrogate evaluation is believed to be further exploited to exceptionally accelerate largescale engineering optimization.
Metagrating
The metagrating [90] could deflect incident light into desired diffraction order, bending light propagating direction at will. Among the traditional optimizations, evolutionary strategies [91] and TO [72] are two representative algorithms for metagrating designs. The former can be found in AfarZanjani studies [92], where an adaptive GA was adopted for a leakywaveantenna type metagrating, and the latter is illustrated in Fig. 7 (a). In this example, David Sell et al. [86] applied an adjointbased TO for high efficiency silicon gratings under a physical configuration that a metagrating deflected normal incident light into an refraction angle of 75°. The efficiency plot showed its value gradually exceeded 80% after 300 iterations and corresponding experimental statistics of a fabricated grating (see a SEM image in Fig. 7 (a)) characterized its working efficiency over 80%. Subsequently, more reports investigated several types of DL models [93, 94] (such as generative networks [95, 96]) to generate highperformance metagratings. An example of GAN network is presented in Fig. 7 (b), where the GAN learned geometric features from a set of metagrating images and thus was able to yield equalquality layouts [87]. After additional topology optimizations, these ‘fake’ structures were fed back into the neural network for retraining and GAN refinement. In this manner, GAN could be employed to facilitate the production of nearoptimal device at a desired deflection angle and wavelength.
Recently growing attention has focused on the hybrid tactics to customize metagratings [97], e.g., to combine NN with conventional optimizers. Besides the abovenoted contributions [87], another seminal work of Liu et al. [88] proposed a fusion scheme connecting compositional patternproducing network and a cooperative coevolution to tailor moleculeresembled metasurfaces (see the algorithm diagram in Fig. 7 (c)). The simulated electric field of E _{y} component emerging from the metasurface confirmed the accurate phase gradient as demanded. Furthermore, for a systemically design of tunable metagratings depicted in Fig. 7 (d), Chuang [89] developed a combination framework comprised of the adjointbased localoptimization with a globaloptimization search (i.e., PSO). Here, the liquid crystal (LC) was aligned perpendicular to TEmode electric fields once the voltage turned on, while the LC became parallel to TEmode on the condition of voltageoff, thereby maximizing the effective refractiveindex change of the LC. The optimal device exhibited a broad deflection angle span (from 12° to 144°) and a high switching efficiency (> 80%), both being 6 times improvements to the stateoftheart reports.
Onchip waveguidebased coupler
The vertical couplers tend to deliver the light from the optical fibers to planar waveguide/chips/photonic circuits. For the purpose of elevating coupling efficiencies, Su et al. [102] proposed a twostage gradientbased optimization algorithm for 1D uniform grating couplers, suppressing the insertion loss of a blazedgratingbased coupler below 0.2 dB. Thereafter, the broadband (> 100 nm) grating couplers devised by the gradientbased strategy achieved 3 dB bandwidths while maintaining central coupling efficiencies ranging from 3.0 dB to 5.4 dB [103]. A recent work on TO was undertaken by Dory et al. [98] who demonstrated several types of diamondbased vertical couplers. The efficiency plot of a representative structure is presented in Fig. 8 (a) with an eventual value stabilized at about 25%. The footprint of such a compact coupler was only 1.0 × 1.0 μm^{2} (see the SEM image). Together with other nanophotonic interfaces, the inversedesigned diamond platforms represent a critical advancement toward integrated diamond quantum optical circuits.
Regarding DL schemes, Gostimirovic et al. [99] discussed on how to use ANN to accelerate preparations of polarizationinsensitive grating couplers (see the geometrical configuration in Fig. 8 (b)). Here the training of ANN model terminated when the validation error reached a value of 6.8%. The optimal operating wavelength for polarizationinsensitive coupler was characterized at 1.596 μm, where two transmittance lines (represented as TM_{0} and TE_{0} polarized incident light) intersected, obtaining an identical amplitude of 0.31. The performance comparison between ANN and FDTD simulations proved that the former was 1830 times faster than the latter while the accuracy was slightly lower (with a 93.2% accuracy). In addition, to quest the overall design potential of vertical grating couplers, Melati et al. [100] demonstrated a MLbased methodology as described in Fig. 8 (c). Following by the procedures of design collection, dimensionality reduction and exhaustive mapping, this pattern recognition method could increase couplertofiber efficiencies > 0.74 at 1550 nm. By dimensionality reduction technique, the ML methodology was expected to navigate and comprehend a wide range of highdimensional design spaces.
Besides the abovementioned, optimizations on other types of onchip couplers have been reported in several studies [104]. For instance, Jin et al. [101] utilized a gradient adjointvariable TO approach to intuitively fashion waveguidecavity couplers served for nonlinear frequency conversion as well as frequency comb generation. The total (or near total) critical coupling between multimode ring resonators and waveguides was achieved at all relevant wavelengths (up to six widely separated wavelengths spanning the 5601500 nm domain), and the ultimate structure distributions are portrayed in Fig. 8 (d).
Beam spiltter
Planar beam splitter (or routers [107, 108]) has been intensively discussed in these publications [25] and here we select recent highlights in this field. Among these advancements, some pioneering breakthroughs were accomplished by Jelena group. For instance, Piggott et al. [26] employed an inverse design algorithm combined the socalled objective first method and the steepest descent approach to formulate the smallest dielectric wavelength splitter at that time (a footprint of 2.8 × 2.8 μm^{2}, see a fabricated SEM image in Fig. 9 (a)). The entire discussion on this gradientbased adjoint photonic optimization could be examined in the relevant documents [109, 110]. A consequential work by Hughes et al. [105] extended the adjoint method to devise nonlinear beam switch by introducing Kerr nonlinearity (induced by the chalcogenide glass, see Fig. 9 (b)). The light was guided to either the right port (linear regime with low power) or the bottom port (nonlinear regime with high power) depending on the refractive index shift imposed by beam intensities. As for DL approach, Tahersima et al. [106] developed a deep residual network model for designs of highly efficient (maximum value > 90%) power splitter (see Fig. 9 (c)). The DNN was trained with nearly 20000 simulation data and could project the nanostructured geometry (i.e., to arrange positions of 20 × 20 etch holes) in a fraction of second. This approach paves the way for speedy design of integrated photonic components with complex layouts.
Optical interference unit
Lately, there is a rising attention towards utilizing optical computing [55, 111–113] to carry out specific machine learning tasks, thanks to the inherent advantages of nanooptics like parallel computation, low power consumption and propagating at the speed of light. Among the proposed optical hardware schemes, the sonamed artificial intelligence interference [52] may offer opportunities to establish photonic systems for visual computing applications. In the next two subsections, we would brief optical interferencebased neural network [52] and diffractive neural networks [114], respectively, with priorities given to network elements and the training methods.
A pioneering work (Fig. 10 (a)) by Shen et al. [115] experimentally demonstrated ONNs using a cascaded array of programmable MachZehnder interferometers (MZIs) and further showed its utility for vowel identifications. Overall, the preprocessed signals (e.g., vowels of ‘ABCD’) were encoded in optical pulse and transported through the main body of ONN, that is, photonic integrated circuits (i.e., builtin nlayer architecture in the sketch of Fig. 10 (a)). As the key component of an ONN, optical interference units (OIUs), relying on MZIlike structures (containing a phase shifter, directional couplers, followed by another phase shifter) can implement any realvalued optical matrix multiplication. The voicerecognition matrixes showed that the electrical and optical hardware were both good at classifying vowels ‘A’ and ‘B’, while the computer surpassed ONN in terms of ‘C’ and ‘D’. After that, intensive studies have profoundly explored interferencelike photonic processors [52]. In fact, the implementations of ONNs depend upon the models simulated in electrical computers due to a lack of efficient training protocols for these optical networks [116]. Hence, for in situ tutoring of an optical neural network, Hughes et al. [116] employed adjoint variable methods to derive the photonic analogue of the backpropagation algorithm. As an example, they numerically demonstrated the instruction of a network (consisting of two 3 × 3 unitary OIUs) to form an XOR gate (see the sketch of neural networks in Fig. 10 (b)), where the MSE between predictions and targets slowly declined to a sufficient small value. The property comparisons of before and after training proved that the ONN had successfully learned the XOR function with projections (blue circles) coinciding well with the targets (black crosses) in the posttraining diagram.
Besides OIUs, a sequence of functional computing elements [119–122] have been proposed and thus occupied as fundamentals to set up ONNs. For example, as displayed in Fig. 10 (c), Qu et al. [117] introduced a novel integrated ONN framework based on optical scattering units (OSUs). In principle, resting on multimode interferences, OSUs could directly manipulate light intensity to run arbitrary matrix multiplications (see the sketch on the top panel of Fig. 10 (c)). In order to realize a highprecision multiplication matrix, the relative permittivity distribution of a specific 4 × 4 OSU was optimized by an adjointbased training process (i.e., Adam optimizer), minimizing the eventual MSE < 10^{−4} (the working performances are presented in Fig. 10 (c)).
Another variant demonstration of OIUs can be found from Khoram paper [118], where an OIUlike media (namely nanophotonic neural medium, NNM) accomplished computer vision tasks such as image recognition of handwritten digits. The input wave recorded the image (20 × 20 pixels) as the intensity distribution and the optical energy yielded from NNM concentrated to corresponding locations according to the image’s label. The iterative training cycle (using minibatch stochastic gradient descent) is displayed in Fig. 10 (d) with an illustrative classifier for an image of number ‘8’ showing a final identification accuracy of 77.3%. In addition, the recognition accuracy could be even more improved to 84% by extending NNM from the flat structure to threedimensional medium.
Optical diffractive neural network
As discussed, optical diffractive neural network could execute various ML tasks as well [28, 126]. An exploratory milestone of diffractive neural networks is displayed in Fig. 11 (a), in which Lin et al. [28] developed an alloptical diffractive deep neural network (D^{2}NN) comprised of multiple passive cascaded layers patterned with complexvalued transmission (or reflection) coefficients. The parameters embedded in diffractive layers were iteratively tuned throughout the error backpropagation learning process. As a result, the classification accuracy of optical D^{2}NN reached 93.39% (seven layers) for digits dataset (Modified National Institute of Standards and Technology, MNIST) and 86.60% (ten layers) for fashion products images (FashionMNIST class 5), respectively. In followup studies, a set of groups continued to improve the system performance of D^{2}NN by several meanings [127–130]. For acceleration of D^{2}NN training speed, Zhou et al. [123] conducted the backpropagation algorithm for in situ training of both linear and nonlinear optical networks. As a sketch illustrated in Fig. 11 (b), the vital instruction procedures involved forward propagation, error calculation, backward propagation, and gradient update. This optical learning architecture achieved a high accuracy on the specific tasks (e.g., object classification and matrixvector multiplication) close to in silico training on an electronic computer.
Furthermore, there exist several studies on Fourier space image processing by optical neural network [124, 131]. For instance, Yan et al. [124] set up a Fourier space D^{2}NN by placing an optical nonlinear activation function (introduced by ferroelectric thin films) in an 2f system (see Fig. 11 (c)). The Fourier D^{2}NN exhibited a classification accuracy of 97.0% based on MNIST database and the hybrid D^{2}NN (diffractive layers embedded in both Fourier space and real space) further lifted the accuracy to 98.1%.
Wider NNlike diffractive optics has been analyzed in bibliographies of diffraction grating based neural network [132], neuromorphic metasurface [133], azimuthal multiplexing 3D diffractive neural network[134], and vortex beams possessing [135], etc. In particular, a noted work [125] of the NNenabled logic gates is conceptually illustrated in Fig. 11 (d). Here, the incident wave was firstly spatially encrypted by a specific logic operator at the input layer and the composite metasurfaces (built on gradient descent training) directionally scattered the encoded light into corresponding designated regions (marked as numbers of ‘0’ and ‘1’). The experimental demonstration of three basic logic operations (i.e., NOT, OR, and AND) exhibited an intensity contrast ratio over 9.6 dB (at a work frequency of 17 GHz), suggesting a strong functional reliability of three gates. Applying extra training on metasurfaces, all seven logic operations (additionally NOR, XNOR, NAND and XOR) were numerically demonstrated via the same optical diffractive neural network.
Other applications
AIalgorithmaided design methodologies are fast becoming fundamental toolboxes to develop highquality photonic structures. To date, the soaring number of nanoscale applications [139–141] have been appreciably benefited from the artificial intelligence such as photonic crystals [142], Fano resonators [143], photon extractors [144], topological insulators [145], and particle accelerators [146], etc. in addition to the aforesaid metadevices. Since it is impractical to cover every aspect, here we would additionally provide three more demonstrations in Fig. 12, listed as holography, nonlinear optical fibers, and optical information storage. Note that, even though they are slightly off the review focus, three cases here show tightly connections between AI and nanophotonic, and further indicate there are plenty room that we can explore.
Optical holography permits the fully restoring and reconstruction amplitude and phase information of object targets [147]. In a recent work by Ren et al. [136], the authors made use of multilayer perceptron ANNs to inversely create a floatingdisplayed holography (see the schematic in Fig. 12 (a)). The trained ANN could instantly output a 2D vector field used for the generation of any desired 3D vectorial field targets, enlarging holographic viewing angle up to 94° with the diffraction efficiency of 78%.
The optical fibers have been intensively studied for more than half a century and so far, still typify as a major scientific field within the photonics community. Recently, Sitawarin et al. [137] applied topological and shape optimization to simulate chalcogenide/polyethersulfone heterostructure fibers to boost thirdharmonic generation at desired wavelength. As can be seen from Fig. 12 (b), the phasematching condition was satisfied at k_{opt}=1.4×2π/λ, improving nonlinear overlap factor almost 4 orders of magnitude larger than standard plain fibers.
Recording information in the scattering spectra of plasmonic nanostructures is perceived as a potential approach for highdensity information storage. While a major problem lies in robust information retrieval in which dissimilar nanostructure geometries may lead to indistinguishable optical responses. To address this issue, Wiecha et al. [138] experimentally demonstrated ANNaided quasierrorfree readout spectral message of silicon nanopatterns loaded up to 9 bits data (in Fig. 12 (c)), contributing to a digitstorage capacity of around 40% higher than that of a Bluray disk.
Perspective and conclusion
Overall, this review contributes to the emerging role of AI learning algorithms in the context of functional metadevices, and spotlights the exponential growth in this crossdisciplinary topic over past three years. The intelligent algorithms, classified as traditional optimization methods and latest DL models, generate productive design approaches for high DoF metastructures, extending the toolkits for reliable photonic devices and liberating us from the heavy workloads over development routines. In addition, the immense innovations of AI may shed new insights into understandings of lightmatter interactions, and hereby help us to further explore the fundamental optics studies.
Despite of the explosive progress achieved so far, the studies of the nanophotonic intelligent designs are still limited in terms of, e.g., design object scopes (wavelengthscale optimizations and independent physical fields) and DL dataset issue (small volume of raw data and distributions). Here we would like to comment on current difficulties with possible solutions, and discuss future metaelement applications (photonic computation scheme) in short.
For the majority of metadevices covered in this review, the dimensional scope of target objects concentrate on the wavelength scale. The largemagnitude and high DoF structures are commonly decomposed into periodical arrays, thereby optimized to enhance the performance of individuals with advanced algorithms. And the separately improved elements would additionally be ‘stitched’ together [17, 72]. Obviously when confronting the practical devices such as achromatic metalens [75], the nanoscale ‘misalignment’ may lead to the degeneration of overall functionalities. Therefore, the efficient optimizers (such as topological optimizations [49, 60, 61, 148]) for systemic and large scale designs [62, 149] should be given higher priorities. Recent attempts on TO studies [150, 151] make promising progress to expedite photonic devices from nanoscale elements toward feasible applications.
A second limitation within object scopes lies in the solo EM simulator. Nowadays, nanophotonics turns to be a much general scientific area and several reports on metacomponents optimizations are expanded to multiphysical backgrounds, e.g., acoustic metamaterials [152] and thermal emitters [153]. Under such multiphysics scenes, merely EM simulator can not offer comprehensive information. Consequently, it is advised to cooperate with multiphysics simulation platforms that encompass diverse solvers, in which we can access software modules by the interfaces and fulfil optimization tasks on demand.
Data is always crucial for deep learnings. The highquality and sufficient information leads to high performance DL models, while a poor data representation is likely to reduce qualities even for most performant algorithms. However, to acquire massive dataset is expensive and sometimes unrealistic regarding the economic and labor costs. To release data generation pressures, generative models (e.g., GANs and VAE) should be much welcomed, since they require a relatively small amount of labels to produce equalquality dataset. Moreover, other advanced AI techniques such as transfer learning and RL models could be particularly considered. For instance, Qu et al. established a DNN architecture with transfer learning ability that can significantly improve the performance of physical problems albeit that the original data sets only included 500 examples [154]. To some extent, these approaches could elucidate behavior of nanophotonic structures [155], enable exploration on design feasibility [100] and potentially discover the underlying physical rules. [156, 157].
Another problem of tackling dataset is how to reasonably collect and coordinate them. On the one hand, many algorithms and nanostructure layouts remain unshared so far. On the other hand, there is a lack of appropriate and reliable platforms, which could properly maintain and effectively distribute metaelement dataset. Apart from some professional sites (e.g., Github) and open platforms (e.g., Metanet[158]), here we would like to call for more neutral databases to share design templates as well as source codes for extra usage like benchmarking algorithm performances.
In addition to the previous issue of data generations and collections, couples of more debates are ongoing centered around the usages of varied algorithms. In comparison to the longestablished approaches (e.g., TO), there exist uncertainties that may further hinder the DL applications. First of all, the comprehensive evaluations of traditional techniques and DL models for metastructure designs are growing in importance, especially considering key aspects such as costperformance ratio. The comparative studies could offer feasible guidance on appropriate selections for various metastructures. Secondly, AIdesigned metastructures ought to be crucially examined particularly including experimental evidence. The emerging metacomponents serves for practical applications all along. Without explicit and solid verification, all designs are empty talks. Lastly, the open questions like performance limits of nanostructures, underlying physics, etc. remain unexplored yet. We should pay more efforts towards fundamental mechanism that cannot be directly answered by DL models. In total, the abovementioned queries are merely the tip of the iceberg, and these would be critically investigated in future studies.
Regardless of object scopes and dataset issue, the AI learning algorithms certainly reduce design effects to develop photonic elements. Among the reviewed photonicelements, one of key application could be anticipated in optical computing. Apart from the aforementioned optical hardware (e.g., OIU and D^{2}NN), numerous efforts including schemes of photonic neuromorphic computing [112, 159–162], optical analog computing [111, 120, 163], photonic integrated circuits [115, 164, 165], and programmable metasurface [166–168] have been persistently forged to drive progressive innovations of photonic computing to realworld usages, such as image processing [169] and equation solvers [170]. These proposed fulloptical or electricoptics coprocessor [171] could be developed as faster, energyefficient, and more powerful computation tools [122, 172] to prompt the AI community as well.
In conclusion, we summarize the AIassisted designs of metacomponents and system from the simple to the complex. We believe that further strengthening the connection between nanophotonics and intelligence algorithms will be of great importance to promote the upcoming breakthroughs in both fields. We warmly welcome the advent of intelligence designed photonic systems that are destined to revolutionize our digital society and remarkably benefit this information era.
Appendix
Opensource packages
Inverse design approaches have reshaped the landscape of nanophotonic structures and here goes to the summary of opensource code packages for inversely designed metadevices illustrated in the main content. Note that, a variety of solutions have been developed and maintained by individual research groups worldwide. Unfortunately, only finite proportions could be presented here due to the limited space. To execute these lines, we strongly endorse powerful CPU, GPU resources and adequate RAM capacity to diminish time spent on date preparations and model trainings. Besides the attached, there exist wellorganized online platforms (e.g., Metanet) collecting and distributing codes together with deduced structure geometries. We would like to recommend readers to consult the relevant documents and websites for more information.
The design and optimization of metastructures require intensive computational efforts due to the heavily dependency on the simulator. These EM solvers are constructed based on various numerical methods like finiteelement method (FEM) and finitedifference timedomain (FDTD) or alternatively using an analytic or a semianalytic model (e.g., rigorous coupled wave analysis (RCWA)). Here, we recap part of wellfounded interpreters into appendix Table 2 (see the content below the dashed line). Except for these listed, another type of highly suggested EM simulators is commercial products including, without limitation, COMSOL Multiphysics, Lumerical, CST Studio, and JCMwave, etc. The numerical software could imitate true 3D/2D physical scenarios to obtain multidimensional optical properties, providing accessible application interface for customized programming.
Availability of data and materials
Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
Abbreviations
 AI:

Artificial intelligence
 ANN:

Artificial neural network
 AE:

Autoencoder
 CD:

Circular dichroism
 cDCGAN:

Conditional deep convolutional generative adversarial network
 DL:

Deep learning
 DNN:

Deep neural network
 DQN:

Deep Q learning network
 DoF:

Design degree of freedom
 D2NN:

Diffractive deep neural network
 EM:

Electromagnetic
 FoM:

Figure of merit
 FF:

Fill factor
 FDTD:

Finitedifference timedomain
 FEM:

Finiteelement method FD^{2}NN: Fourierspace diffractive neural network
 FWHM:

Full width at half maximum
 GAN:

Generative adversarial network
 GA:

Genetic algorithm
 GPN:

Geometry prediction network
 LC:

Liquid crystal
 ML:

Machine learning
 MZI:

Mach–Zehnder interferometer
 MSE:

Mean square root
 MNIST:

Modified National Institute of Standards and Technology
 MLP:

Multilayer perception
 NNM:

Nanophotonic neural medium
 NN:

Neural network
 NA:

Numerical aperture
 N.A.:

Not applicable
 OF:

Objective function
 OIU:

Optical interference unit
 ONN:

Optical neural network
 OSU:

Optical scattering unit
 PSO:

Particle swarm optimization
 RL:

Reinforcement learning
 RCWA:

Rigorous coupled wave analysis
 SEM:

Scanning electron microscopy
 SHG:

Second harmonic generation
 SPN:

Spectrum predicting network
 SFG:

Sumfrequency generation
 SERS:

Surface enhanced Raman scattering
 TO:

Topological optimization
 VAE:

Variational autoencoder
References
 1
Novotny L, Hecht B. Principles of Nanooptics. Cambridge: Cambridge University Press; 2012.
 2
Maier SA. Plasmonics: Fundamentals and Applications. Berlin: Springer; 2007.
 3
Lin Sy, Fleming J, Hetherington D, Smith B, Biswas R, Ho K, Sigalas M, Zubrzycki W, Kurtz S, Bur J. A threedimensional photonic crystal operating at infrared wavelengths. Nature. 1998; 394(6690):251–53.
 4
Russell P. Photonic crystal fibers. Science. 2003; 299(5605):358–62.
 5
Maier SA, Brongersma ML, Kik PG, Meltzer S, Requicha AA, Atwater HA. Plasmonics  a route to nanoscale optical devices. Adv Mater. 2001; 13(19):1501–05.
 6
Barnes WL, Dereux A, Ebbesen TW. Surface plasmon subwavelength optics. Nature. 2003; 424(6950):824–30.
 7
Smith DR, Padilla WJ, Vier D, NematNasser SC, Schultz S. Composite medium with simultaneously negative permeability and permittivity. Phys Rev Lett. 2000; 84(18):4184.
 8
Yan M, Ruan Z, Qiu M. Scattering characteristics of simplified cylindrical invisibility cloaks. Opt Express. 2007; 15(26):17772–82.
 9
Zhang B, Luo Y, Liu X, Barbastathis G. Macroscopic invisibility cloak for visible light. Phys Rev Lett. 2011; 106(3):033901.
 10
Campbell SD, Sell D, Jenkins RP, Whiting EB, Fan JA, Werner DH. Review of numerical optimization techniques for metadevice design. Opt Mater Express. 2019; 9(4):1842–63.
 11
Malkiel I, Mrejen M, Nagler A, Arieli U, Wolf L, Suchowski H. Plasmonic nanostructure design and characterization via deep learning. Light: Sci Appl. 2018; 7(1):1–8.
 12
So S, Badloe T, Noh J, Rho J, BravoAbad J. Deep learning enabled inverse design in nanophotonics. Nanophotonics. 2020; 9(5):1041–57.
 13
Russell S, Norvig P. Artificial Intelligence: a Modern Approach, 4th Ed. Hoboken: Pearson Education, Inc.; 2020.
 14
corporation N. Difference Between AI, Machine Learning, and Deep Learning. 2021. https://blogs.nvidia.com/blog/2016/07/29/whatsdifferenceartificialintelligencemachinelearningdeeplearningai/. Accessed 28 June 2021.
 15
Elsawy MM, Lanteri S, Duvigneau R, Fan JA, Genevet P. Numerical optimization methods for metasurfaces. Laser Photon Rev. 2020; 14(10):1900445.
 16
Rho J, Fan JA. Freeform metasurface design based on topology optimization. MRS Bull. 2020; 45(3):196–201.
 17
Lin Z, Liu V, Pestourie R, Johnson SG. Topology optimization of freeform largearea metasurfaces. Opt Express. 2019; 27(11):15765–75.
 18
Zhang Q, Yu H, Barbiero M, Wang B, Gu M. Artificial neural networks enabled by nanophotonics. Light: Sci Appl. 2019; 8(1):1–14.
 19
Ma W, Liu Z, Kudyshev ZA, Boltasseva A, Cai W, Liu Y. Deep learning for the design of photonic structures. Nat Photon. 2021; 15(2):77–90.
 20
Jiang J, Chen M, Fan JA. Deep neural networks for the evaluation and design of photonic devices. Nat Rev Mater. 2020:1–22.
 21
Offrein BJ, Bona GL, Germann R, Massarek I, Erni D, et al. A very short planar silica spotsize converter using a nonperiodic segmented waveguide. J Lightwave Technol. 1998; 16(9):1680.
 22
Dobson DC, Cox SJ. Maximizing band gaps in twodimensional photonic crystals. SIAM J Appl Math. 1999; 59(6):2108–20.
 23
Borel PI, Harpøth A, Frandsen LH, Kristensen M, Shi P, Jensen JS, Sigmund O. Topology optimization and fabrication of photonic crystal structures. Opt Express. 2004; 12(9):1996–2001.
 24
LalauKeraly CM, Bhargava S, Miller OD, Yablonovitch E. Adjoint shape optimization applied to electromagnetic design. Opt Express. 2013; 21(18):21693–701.
 25
Molesky S, Lin Z, Piggott AY, Jin W, Vucković J, Rodriguez AW. Inverse design in nanophotonics. Nat Photonics. 2018; 12(11):659–70.
 26
Piggott AY, Lu J, Lagoudakis KG, Petykiewicz J, Babinec TM, Vučković J. Inverse design and demonstration of a compact and broadband onchip wavelength demultiplexer. Nat Photonics. 2015; 9(6):374–77.
 27
Ma W, Cheng F, Liu Y. Deeplearningenabled ondemand design of chiral metamaterials. ACS Nano. 2018; 12(6):6326–34.
 28
Lin X, Rivenson Y, Yardimci NT, Veli M, Luo Y, Jarrahi M, Ozcan A. Alloptical machine learning using diffractive deep neural networks. Science. 2018; 361(6406):1004–08.
 29
Liu C, Maier SA, Li G. Geneticalgorithmaided metaatom multiplication for improved absorption and coloration in nanophotonics. ACS Photonics. 2020; 7(7):1716–22.
 30
Bonod N, Bidault S, Burr GW, Mivelle M. Evolutionary optimization of alldielectric magnetic nanoantennas. Adv Opt Mater. 2019; 7(10):1900121.
 31
Li Z, Stan L, Czaplewski DA, Yang X, Gao J. Broadband infrared binarypattern metasurface absorbers with microgenetic algorithm optimization. Opt Lett. 2019; 44(1):114–17.
 32
Pogrebnyakov AV, Bossard JA, Turpin JP, Musgraves JD, Shin HJ, RiveroBaleine C, Podraza N, Richardson KA, Werner DH, Mayer TS. Reconfigurable nearir metasurface based on ge 2 sb 2 te 5 phasechange material. Opt Mater Express. 2018; 8(8):2264–75.
 33
Li Z, Rosenmann D, Czaplewski DA, Yang X, Gao J. Strong circular dichroism in chiral plasmonic metasurfaces optimized by microgenetic algorithm. Opt Express. 2019; 27(20):28313–23.
 34
Forestiere C, Donelli M, Walsh GF, Zeni E, Miano G, Dal Negro L. Particleswarm optimization of broadband nanoplasmonic arrays. Opt Lett. 2010; 35(2):133–35.
 35
Christiansen RE, Michon J, Benzaouia M, Sigmund O, Johnson SG. Inverse design of nanoparticles for enhanced raman scattering. Opt Express. 2020; 28(4):4444–62.
 36
Li Y, Xu Y, Jiang M, Li B, Han T, Chi C, Lin F, Shen B, Zhu X, Lai L, et al. Selflearning perfect optical chirality via a deep neural network. Phys Rev Lett. 2019; 123(21):213902.
 37
Tao Z, You J, Zhang J, Zheng X, Liu H, Jiang T. Optical circular dichroism engineering in chiral metamaterials utilizing a deep learning network. Opt Lett. 2020; 45(6):1403–06.
 38
Gao L, Li X, Liu D, Wang L, Yu Z. A bidirectional deep neural network for accurate silicon color design. Adv Mater. 2019; 31(51):1905467.
 39
Li X, Shu J, Gu W, Gao L. Deep neural network for plasmonic sensor modeling. Opt Mater Express. 2019; 9(9):3857–62.
 40
Hemmatyar O, Abdollahramezani S, Kiarashinejad Y, Zandehshahvar M, Adibi A. Full color generation with fanotype resonant hfo 2 nanopillars designed by a deeplearning approach. Nanoscale. 2019; 11(44):21266–74.
 41
Ma W, Cheng F, Xu Y, Wen Q, Liu Y. Probabilistic representation and inverse design of metamaterials based on a deep generative model with semisupervised learning strategy. Adv Mater. 2019; 31(35):1901111.
 42
Liu Z, Zhu D, Rodrigues SP, Lee KT, Cai W. Generative model for the inverse design of metasurfaces. Nano Lett. 2018; 18(10):6570–76.
 43
So S, Rho J. Designing nanophotonic structures using conditional deep convolutional generative adversarial networks. Nanophotonics. 2019; 8(7):1255–61.
 44
Sajedian I, Badloe T, Rho J. Optimisation of colour generation from dielectric nanostructures using reinforcement learning. Opt Express. 2019; 27(4):5874–83.
 45
Yao K, Unni R, Zheng Y. Intelligent nanophotonics: merging photonics and artificial intelligence at the nanoscale. Nanophotonics. 2019; 8(3):339–66.
 46
Liu Z, Zhu D, Raju L, Cai W. Tackling Photonic Inverse Design with Machine Learning. Adv Sci. 2021; 8(5):2002923.
 47
Xu Y, Zhang X, Fu Y, Liu Y. Interfacing photonics with artificial intelligence: an innovative design strategy for photonic structures and devices based on artificial neural networks. Photon Res. 2021; 9(4):135–52.
 48
Wiecha PR, Arbouet A, Girard C, Muskens OL. Deep learning in nanophotonics: inverse design and beyond. Photon Res. 2021; 9(5):182–200.
 49
Christiansen RE, Sigmund O. A tutorial for inverse design in photonics by topology optimization. arXiv preprint arXiv:2008.11816. 2020.
 50
Ma L, Li J, Liu Z, Zhang Y, Zhang N, Zheng S, Lu C. Intelligent algorithms: new avenues for designing nanophotonic devices. Chin Opt Lett. 2021; 19(1):011301.
 51
Christiansen RE, Sigmund O. Compact 200 line matlab code for inverse design in photonics by topology optimization: tutorial. J Opt Soc Am B. 2021; 38(2):510–20.
 52
Wetzstein G, Ozcan A, Gigan S, Fan S, Englund D, Soljačić M, Denz C, Miller DA, Psaltis D. Inference in artificial intelligence with deep optics and photonics. Nature. 2020; 588(7836):39–47.
 53
Zhou J, Huang B, Yan Z, Bünzli JCG. Emerging role of machine learning in lightmatter interaction. Light: Sci Appl. 2019; 8(1):1–7.
 54
Brown KA, Brittman S, Maccaferri N, Jariwala D, Celano U. Machine learning in nanoscience: Big data at small scales. Nano Lett. 2019; 20(1):2–10.
 55
Shastri BJ, Tait AN, de Lima TF, Pernice WH, Bhaskaran H, Wright CD, Prucnal PR. Photonics for artificial intelligence and neuromorphic computing. Nat Photon. 2021; 15(2):102–14.
 56
Meinzer N, Barnes WL, Hooper IR. Plasmonic metaatoms and metasurfaces. Nat Photon. 2014; 8(12):889.
 57
Weile DS, Michielssen E. Genetic algorithm optimization applied to electromagnetics: A review. IEEE Trans Antennas Propag. 1997; 45(3):343–53.
 58
LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015; 521(7553):436–44.
 59
Goodfellow I, Bengio Y, Courville A, Bengio Y. Deep Learning vol. 1: MIT press Cambridge; 2016.
 60
Bendsoe MP, Sigmund O. Topology Optimization: Theory, Methods, and Applications. Berlin: Springer; 2013.
 61
Jensen JS, Sigmund O. Topology optimization for nanophotonics. Laser Photonics Rev. 2011; 5(2):308–21.
 62
Schneider PI, Garcia Santiago X, Soltwisch V, Hammerschmidt M, Burger S, Rockstuhl C. Benchmarking five global optimization approaches for nanooptical shape optimization and parameter reconstruction. ACS Photonics. 2019; 6(11):2726–33.
 63
Kennedy J, Eberhart R. Particle swarm optimization. In: Proceedings of ICNN’95international Conference on Neural Networks, vol. 4. IEEE: 1995. p. 1942–48.
 64
Shi Y, et al. Particle swarm optimization: developments, applications and resources. In: Proceedings of the 2001 Congress on Evolutionary Computation (IEEE Cat. No. 01TH8546), vol. 1. IEEE: 2001. p. 81–86.
 65
Christiansen RE, Sigmund O. Inverse design in photonics by topology optimization: tutorial. J Opt Soc Am B. 2021; 38(2):496–509.
 66
Sigmund O. A 99 line topology optimization code written in matlab. Struct Multidiscip Optim. 2001; 21(2):120–27.
 67
Ferrari F, Sigmund O. A new generation 99 line matlab code for compliance topology optimization and its extension to 3d. Struct Multidiscip Optim. 2020; 62(4):2211–28.
 68
Doersch C. Tutorial on variational autoencoders. arXiv preprint arXiv:1606.05908. 2016.
 69
Badloe T, Kim I, Rho J. Biomimetic ultrabroadband perfect absorbers optimised with reinforcement learning. Phys Chem Chem Phys. 2020; 22(4):2337–42.
 70
Sajedian I, Lee H, Rho J. Doubledeep qlearning to increase the efficiency of metasurface holograms. Sci Rep. 2019; 9(1):1–8.
 71
Liang H, Lin Q, Xie X, Sun Q, Wang Y, Zhou L, Liu L, Yu X, Zhou J, Krauss TF, et al. Ultrahigh numerical aperture metalens at visible wavelengths. Nano Lett. 2018; 18(7):4460–66.
 72
Phan T, Sell D, Wang EW, Doshay S, Edee K, Yang J, Fan JA. Highefficiency, largearea, topologyoptimized metasurfaces. Light: Sci Appl. 2019; 8(1):1–9.
 73
Christiansen RE, Lin Z, RoquesCarmes C, Salamin Y, Kooi SE, Joannopoulos JD, Soljačić M, Johnson SG. Fullwave maxwell inverse design of axisymmetric, tunable, and multiscale multiwavelength metalenses. Opt Express. 2020; 28(23):33854–68.
 74
Mansouree M, Kwon H, Arbabi E, McClung A, Faraon A, Arbabi A. Multifunctional 2.5 d metastructures enabled by adjoint optimization. Optica. 2020; 7(1):77–84.
 75
Zou X, Zheng G, Yuan Q, Zang W, Chen R, Li T, Li L, Wang S, Wang Z, Zhu S. Imaging based on metalenses. PhotoniX. 2020; 1(1):1–24.
 76
Fan Y, Xu Y, Qiu M, Jin W, Zhang L, Lam EY, Tsai DP, Lei D. Phasecontrolled metasurface design via optimized genetic algorithm. Nanophotonics. 2020; 9(12):3931–9.
 77
Meem M, Banerji S, Pies C, Oberbiermann T, Majumder A, SensaleRodriguez B, Menon R. Largearea, highnumericalaperture multilevel diffractive lens via inverse design. Optica. 2020; 7(3):252–53.
 78
Banerji S, Meem M, Majumder A, SensaleRodriguez B, Menon R. Imaging over an unlimited bandwidth with a single diffractive surface. arXiv preprint arXiv:1907.06251. 2019.
 79
Banerji S, Meem M, Majumder A, SensaleRodriguez B, Menon R. Extremedepthoffocus imaging with a flat lens. Optica. 2020; 7(3):214–17.
 80
Chung H, Miller OD. Highna achromatic metalenses by inverse design. Opt Express. 2020; 28(5):6945–65.
 81
Lin Z, Johnson SG. Overlapping domains for topology optimization of largearea metasurfaces. Opt Express. 2019; 27(22):32445–53.
 82
Lin Z, Johnson SG. Topologyoptimized nanostructures for highna lensing optics. In: 2020 International Applied Computational Electromagnetics Society Symposium (ACES). IEEE: 2020. p. 1–2.
 83
Bayati E, Pestourie R, Colburn S, Lin Z, Johnson SG, Majumdar A. Inverse designed metalenses with extended depth of focus. ACS Photon. 2020; 7(4):873–78.
 84
Mansouree M, McClung A, Samudrala S, Arbabi A. Largescale parametrized metasurface design using adjoint optimization. ACS Photon. 2021; 8(2):455–63.
 85
Pestourie R, Mroueh Y, Nguyen TV, Das P, Johnson SG. Active learning of deep surrogates for pdes: Application to metasurface design. npj Comput Mater. 2020; 6(1):1–7.
 86
Sell D, Yang J, Doshay S, Yang R, Fan JA. Largeangle, multifunctional metagratings based on freeform multimode geometries. Nano Lett. 2017; 17(6):3752–57.
 87
Jiang J, Sell D, Hoyer S, Hickey J, Yang J, Fan JA. Freeform diffractive metagrating design based on generative adversarial networks. ACS Nano. 2019; 13(8):8872–78.
 88
Liu Z, Zhu D, Lee KT, Kim AS, Raju L, Cai W. Compounding metaatoms into metamolecules with hybrid artificial intelligence techniques. Adv Mater. 2020; 32(6):1904790.
 89
Chung H, Miller OD. Tunable metasurface inverse design for 80% switching efficiencies and 144 angular deflection. ACS Photon. 2020; 7(8):2236–43.
 90
Khorasaninejad M, Capasso F. Broadband multifunctional efficient metagratings based on dielectric waveguide phase shifters. Nano Lett. 2015; 15(10):6709–15.
 91
Elsawy MM, Lanteri S, Duvigneau R, Brière G, Mohamed MS, Genevet P. Global optimization of metasurface designs using statistical learning methods. Sci Rep. 2019; 9(1):1–15.
 92
JafarZanjani S, Inampudi S, Mosallaei H. Adaptive genetic algorithm for optical metasurfaces design. Sci Rep. 2018; 8(1):1–16.
 93
Inampudi S, Mosallaei H. Neural network based design of metagratings. Appl Phys Lett. 2018; 112(24):241102.
 94
Jiang J, Fan JA. Global optimization of dielectric metasurfaces using a physicsdriven neural network. Nano Lett. 2019; 19(8):5366–72.
 95
Wen F, Jiang J, Fan JA. Robust freeform metasurface design based on progressively growing generative networks. ACS Photon. 2020; 7(8):2098–104.
 96
Jiang J, Fan JA. Simulatorbased training of generative neural networks for the inverse design of metasurfaces. Nanophotonics. 2020; 9(5):1059–69.
 97
Chen M, Jiang J, Fan JA. Design space reparameterization enforces hard geometric constraints in inversedesigned nanophotonic devices. ACS Photon. 2020; 7(11):3141–51.
 98
Dory C, Vercruysse D, Yang KY, Sapra NV, Rugar AE, Sun S, Lukin DM, Piggott AY, Zhang JL, Radulaski M, et al.Inversedesigned diamond photonics. Nat Commun. 2019; 10(1):1–7.
 99
Gostimirovic D, Winnie NY. An opensource artificial neural network model for polarizationinsensitive silicononinsulator subwavelength grating couplers. IEEE J Sel Top Quantum Electron. 2018; 25(3):1–5.
 100
Melati D, Grinberg Y, Dezfouli MK, Janz S, Cheben P, Schmid JH, SánchezPostigo A, Xu DX. Mapping the global design space of nanophotonic components using machine learning pattern recognition. Nat Commun. 2019; 10(1):1–9.
 101
Jin W, Molesky S, Lin Z, Fu KMC, Rodriguez AW. Inverse design of compact multimode cavity couplers. Opt Express. 2018; 26(20):26713–21.
 102
Su L, Trivedi R, Sapra NV, Piggott AY, Vercruysse D, Vučković J. Fullyautomated optimization of grating couplers. Opt Express. 2018; 26(4):4023–34.
 103
Sapra NV, Vercruysse D, Su L, Yang KY, Skarda J, Piggott AY, Vučković J. Inverse design and demonstration of broadband grating couplers. IEEE J Sel Top Quantum Electron. 2019; 25(3):1–7.
 104
Jin W, Li W, Orenstein M, Fan S. Inverse design of lightweight broadband reflector for relativistic lightsail propulsion. ACS Photon. 2020; 7(9):2350–55.
 105
Hughes TW, Minkov M, Williamson IA, Fan S. Adjoint method and inverse design for nonlinear nanophotonic devices. ACS Photon. 2018; 5(12):4781–87.
 106
Tahersima MH, Kojima K, KoikeAkino T, Jha D, Wang B, Lin C, Parsons K. Deep neural network inverse design of integrated photonic power splitters. Sci Rep. 2019; 9(1):1–9.
 107
Liu Z, Liu X, Xiao Z, Lu C, Wang HQ, Wu Y, Hu X, Liu YC, Zhang H, Zhang X. Integrated nanophotonic wavelength router based on an intelligent algorithm. Optica. 2019; 6(10):1367–73.
 108
Lu C, Liu Z, Wu Y, Xiao Z, Yu D, Zhang H, Wang C, Hu X, Liu YC, Liu X, et al. Nanophotonic polarization routers based on an intelligent algorithm. Adv Opt Mater. 2020; 8(10):1902018.
 109
Su L, Vercruysse D, Skarda J, Sapra NV, Petykiewicz JA, Vučković J. Nanophotonic inverse design with spins: Software architecture and practical considerations. Appl Phys Rev. 2020; 7(1):011407.
 110
Piggott AY, Petykiewicz J, Su L, Vučković J. Fabricationconstrained nanophotonic inverse design. Sci Rep. 2017; 7(1):1–7.
 111
ZangenehNejad F, Sounas DL, Alù A, Fleury R. Analogue computing with metamaterials. Nat Rev Mater. 2020:1–19.
 112
Marković D, Mizrahi A, Querlioz D, Grollier J. Physics for neuromorphic computing. Nat Rev Phys. 2020; 2(9):499–510.
 113
Xu R, Lv P, Xu F, Shi Y. A survey of approaches for implementing optical neural networks. Opt Laser Technol. 2021; 136:106787.
 114
Liu J, Wu Q, Sui X, Chen Q, Gu G, Wang L, Li S. Research progress in optical neural networks: theory, applications and developments. PhotoniX. 2021; 2(1):1–39.
 115
Shen Y, Harris NC, Skirlo S, Prabhu M, BaehrJones T, Hochberg M, Sun X, Zhao S, Larochelle H, Englund D, et al.Deep learning with coherent nanophotonic circuits. Nat Photon. 2017; 11(7):441.
 116
Hughes TW, Minkov M, Shi Y, Fan S. Training of photonic neural networks through in situ backpropagation and gradient measurement. Optica. 2018; 5(7):864–71.
 117
Qu Y, Zhu H, Shen Y, Zhang J, Tao C, Ghosh P, Qiu M. Inverse design of an integratednanophotonics optical neural network. Sci Bull. 2020; 65(14):1177–83.
 118
Khoram E, Chen A, Liu D, Ying L, Wang Q, Yuan M, Yu Z. Nanophotonic media for artificial neural inference. Photon Res. 2019; 7(8):823–27.
 119
Harris NC, Carolan J, Bunandar D, Prabhu M, Hochberg M, BaehrJones T, Fanto ML, Smith AM, Tison CC, Alsing PM, et al.Linear programmable nanophotonic processors. Optica. 2018; 5(12):1623–31.
 120
Estakhri NM, Edwards B, Engheta N. Inversedesigned metastructures that solve equations. Science. 2019; 363(6433):1333–38.
 121
Hamerly R, Bernstein L, Sludds A, Soljačić M, Englund D. Largescale optical neural networks based on photoelectric multiplication. Phys Rev X. 2019; 9(2):021032.
 122
Miscuglio M, Sorger VJ. Photonic tensor cores for machine learning. Appl Phys Rev. 2020; 7(3):031404.
 123
Zhou T, Fang L, Yan T, Wu J, Li Y, Fan J, Wu H, Lin X, Dai Q. In situ optical backpropagation training of diffractive optical neural networks. Photon Res. 2020; 8(6):940–53.
 124
Yan T, Wu J, Zhou T, Xie H, Xu F, Fan J, Fang L, Lin X, Dai Q. Fourierspace diffractive deep neural network. Phys Rev Lett. 2019; 123(2):023901.
 125
Qian C, Lin X, Lin X, Xu J, Sun Y, Li E, Zhang B, Chen H. Performing optical logic operations by a diffractive neural network. Light: Sci Appl. 2020; 9(1):1–7.
 126
Sui X, Wu Q, Liu J, Chen Q, Gu G. A review of optical neural networks. IEEE Access. 2020; 8:70773–83.
 127
Li J, Mengu D, Luo Y, Rivenson Y, Ozcan A. Classspecific differential detection in diffractive optical neural networks improves inference accuracy. Adv Photon. 2019; 1(4):046001.
 128
Luo Y, Mengu D, Yardimci NT, Rivenson Y, Veli M, Jarrahi M, Ozcan A. Design of taskspecific optical systems using broadband diffractive neural networks. Light: Sci Appl. 2019; 8(1):1–14.
 129
Dou H, Deng Y, Yan T, Wu H, Lin X, Dai Q. Residual d 2 nn: training diffractive deep neural networks via learnable light shortcuts. Opt Lett. 2020; 45(10):2688–91.
 130
Rahman MSS, Li J, Mengu D, Rivenson Y, Ozcan A. Ensemble learning of diffractive optical networks. Light: Sci Appl. 2021; 10(1):1–13.
 131
Muminov B, Vuong LT. Fourier optical preprocessing in lieu of deep learning. Optica. 2020; 7(9):1079–88.
 132
Lu L, Zhu L, Zhang Q, Zhu B, Yao Q, Yu M, Niu H, Dong M, Zhong G, Zeng Z. Miniaturized diffraction grating design and processing for deep neural network. IEEE Photon Technol Lett. 2019; 31(24):1952–55.
 133
Wu Z, Zhou M, Khoram E, Liu B, Yu Z. Neuromorphic metasurface. Photon Res. 2020; 8(1):46–50.
 134
Wang H, Piestun R. Azimuthal multiplexing 3d diffractive optics. Sci Rep. 2020; 10(1):1–9.
 135
Huang Z, Wang P, Liu J, Xiong W, He Y, Xiao J, Ye H, Li Y, Chen S, Fan D. Alloptical signal processing of vortex beams with diffractive deep neural networks. Phys Rev Appl. 2021; 15(1):014037.
 136
Ren H, Shao W, Li Y, Salim F, Gu M. Threedimensional vectorial holography based on machine learning inverse design. Sci Adv. 2020; 6(16):4261.
 137
Sitawarin C, Jin W, Lin Z, Rodriguez AW. Inversedesigned photonic fibers and metasurfaces for nonlinear frequency conversion. Photon Res. 2018; 6(5):82–89.
 138
Wiecha PR, Lecestre A, Mallet N, Larrieu G. Pushing the limits of optical information storage using deep learning. Nat Nanotechnol. 2019; 14(3):237–44.
 139
Li L, Shuang Y, Ma Q, Li H, Zhao H, Wei M, Liu C, Hao C, Qiu CW, Cui TJ. Intelligent metasurface imager and recognizer. Light: Sci Appl. 2019; 8(1):1–9.
 140
Piggott AY, Ma EY, Su L, Ahn GH, Sapra NV, Vercruysse D, Netherton AM, Khope AS, Bowers JE, Vuckovic J. Inversedesigned photonics for semiconductor foundries. ACS Photon. 2020; 7(3):569–75.
 141
Xie Z, Lei T, Qiu H, Zhang Z, Wang H, Yuan X. Broadband onchip photonic spin hall element via inverse design. Photon Res. 2020; 8(2):121–26.
 142
Minkov M, Williamson IA, Andreani LC, Gerace D, Lou B, Song AY, Hughes TW, Fan S. Inverse design of photonic crystals through automatic differentiation. ACS Photon. 2020; 7(7):1729–41.
 143
Yang KY, Skarda J, Cotrufo M, Dutt A, Ahn GH, Sawaby M, Vercruysse D, Arbabian A, Fan S, Alù A, et al.Inversedesigned nonreciprocal pulse router for chipbased lidar. Nat Photon. 2020; 14(6):369–74.
 144
Chakravarthi S, Chao P, Pederson C, Molesky S, Ivanov A, Hestroffer K, Hatami F, Rodriguez AW, Fu KMC. Inversedesigned photon extractors for optically addressable defect qubits. Optica. 2020; 7(12):1805–11.
 145
Christiansen RE, Wang F, Sigmund O. Topological insulators by topology optimization. Phys Rev Lett. 2019; 122(23):234502.
 146
Sapra NV, Yang KY, Vercruysse D, Leedle KJ, Black DS, England RJ, Su L, Trivedi R, Miao Y, Solgaard O, et al.Onchip integrated laserdriven particle accelerator. Science. 2020; 367(6473):79–83.
 147
Zhao R, Huang L, Wang Y. Recent advances in multidimensional metasurfaces holographic technologies. PhotoniX. 2020; 1(1):1–24.
 148
Liu J, Ma Y. A survey of manufacturing oriented topology optimization methods. Adv Eng Softw. 2016; 100:161–75.
 149
Zhan A, Gibson R, Whitehead J, Smith E, Hendrickson JR, Majumdar A. Controlling threedimensional optical fields via inverse mie scattering. Sci Adv. 2019; 5(10):4769.
 150
Augenstein Y, Rockstuhl C. Inverse design of nanophotonic devices with structural integrity. ACS Photon. 2020; 7(8):2190–96.
 151
Lin Z, RoquesCarmes C, Christiansen RE, Soljačić M, Johnson SG. Computational inverse design for ultracompact singlepiece metalenses free of chromatic and angular aberration. Appl Phys Lett. 2021; 118(4):041104.
 152
Zheng B, Yang J, Liang B, Cheng Jc. Inverse design of acoustic metamaterials based on machine learning using a gauss–bayesian model. J Appl Phys. 2020; 128(13):134902.
 153
Kudyshev ZA, Kildishev AV, Shalaev VM, Boltasseva A. Machinelearningassisted metasurface design for highefficiency thermal emitter optimization. Appl Phys Rev. 2020; 7(2):021407.
 154
Qu Y, Jing L, Shen Y, Qiu M, Soljacic M. Migrating knowledge between physical scenarios based on artificial neural networks. ACS Photon. 2019; 6(5):1168–74.
 155
Yeung C, Tsai JM, King B, Kawagoe Y, Ho D, Knight MW, Raman AP. Elucidating the behavior of nanophotonic structures through explainable machine learning algorithms. ACS Photon. 2020; 7(8):2309–18.
 156
Kiarashinejad Y, Zandehshahvar M, Abdollahramezani S, Hemmatyar O, Pourabolghasem R, Adibi A. Knowledge discovery in nanophotonics using geometric deep learning. Adv Intell Syst. 2020; 2(2):1900132.
 157
Wiecha PR, Arbouet A, Girard C, Muskens OL. Deep learning in nanophotonics: inverse design and beyond. arXiv preprint arXiv:2011.12603. 2020.
 158
Jiang J, Lupoiu R, Wang EW, Sell D, Hugonin JP, Lalanne P, Fan JA. Metanet: a new paradigm for data sharing in photonics research. Opt Express. 2020; 28(9):13670–81.
 159
Goi E, Zhang Q, Chen X, Luan H, Gu M. Perspective on photonic memristive neuromorphic computing. PhotoniX. 2020; 1(1):1–26.
 160
Lugnan A, Katumba A, Laporte F, Freiberger M, Sackesyn S, Ma C, Gooskens E, Dambre J, Bienstman P. Photonic neuromorphic information processing and reservoir computing. APL Photon. 2020; 5(2):020901.
 161
Ballarini D, Gianfrate A, Panico R, Opala A, Ghosh S, Dominici L, Ardizzone V, De Giorgi M, Lerario G, Gigli G, et al. Polaritonic neuromorphic computing outperforms linear classifiers. Nano Lett. 2020; 20(5):3506–12.
 162
de Lima TF, Tait AN, Mehrabian A, Nahmias MA, Huang C, Peng HT, Marquez BA, Miscuglio M, ElGhazawi T, Sorger VJ, et al.Primer on silicon neuromorphic photonic processors: architecture and compiler. Nanophotonics. 2020; 9(13):4055–73.
 163
Abdollahramezani S, Hemmatyar O, Adibi A. Metaoptics for spatial optical analog computing. Nanophotonics. 2020; 9(13):4075–95.
 164
Norman JC, Jung D, Wan Y, Bowers JE. Perspective: The future of quantum dot photonic integrated circuits. APL Photon. 2018; 3(3):030901.
 165
Peng HT, Nahmias MA, De Lima TF, Tait AN, Shastri BJ. Neuromorphic photonic integrated circuits. IEEE J Sel Top Quantum Electron. 2018; 24(6):1–15.
 166
Zhang XG, Jiang WX, Jiang HL, Wang Q, Tian HW, Bai L, Luo ZJ, Sun S, Luo Y, Qiu CW, et al.An optically driven digital metasurface for programming electromagnetic functions. Nat Electron. 2020; 3(3):165–71.
 167
Tsilipakos O, Tasolamprou AC, Pitilakis A, Liu F, Wang X, Mirmoosa MS, Tzarouchis DC, Abadal S, Taghvaee H, Liaskos C, et al.Toward intelligent metasurfaces: The progress from globally tunable metasurfaces to softwaredefined metasurfaces with an embedded network of controllers. Adv Opt Mater. 2020; 8(17):2000783.
 168
Ma Q, Cui TJ. Information metamaterials: bridging the physical world and digital world. PhotoniX. 2020; 1(1):1–32.
 169
Xu X, Tan M, Corcoran B, Wu J, Boes A, Nguyen TG, Chu ST, Little BE, Hicks DG, Morandotti R, et al.11 tops photonic convolutional accelerator for optical neural networks. Nature. 2021; 589(7840):44–51.
 170
Camacho M, Edwards B, Engheta N. A single inversedesigned photonic structure that performs parallel computing. arXiv preprint arXiv:2009.01187. 2020.
 171
Zhou T, Lin X, Wu J, Chen Y, Xie H, Li Y, Fan J, Wu H, Fang L, Dai Q. Largescale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit. Nat Photon. 2021:1–7.
 172
Zuo Y, Li B, Zhao Y, Jiang Y, Chen YC, Chen P, Jo GB, Liu J, Du S. Alloptical neural network with nonlinear activation functions. Optica. 2019; 6(9):1132–37.
Acknowledgements
The authors gratefully acknowledge support from the National Natural Science Foundation of China (No. 62005224, 61927820) and the National Key Research and Development Program of China (2017YFA0205700). N.W. specially thanks for Dr. Tim Jiangbo Zhao for his proofreading of the manuscript.
Funding
National Natural Science Foundation of China (No. 62005224,61927820), National Key Research and Development Program of China (2017YFA0205700)
Author information
Affiliations
Contributions
NW and MQ conceived the idea. NW prepared the manuscript. WY, YQ, SM, SL, and MQ revised the manuscript. All authors read and approved the final manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors declare that they have no competing interests
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Wang, N., Yan, W., Qu, Y. et al. Intelligent designs in nanophotonics: from optimization towards inverse creation. PhotoniX 2, 22 (2021). https://doi.org/10.1186/s4307402100044y
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s4307402100044y
Keywords
 Artificial intelligence
 Nanophotonics
 Metasurface
 Optical neural network
 Intelligent designs