A reverse-engineered metasurface device and its preparation and design method

Abstract

The present invention discloses a reverse-designed metasurface device and a preparation and design method thereof, comprising a transparent medium substrate and a first layer of dielectric micro-nano unit structure, a HSQ separation layer and a second layer of dielectric micro-nano unit structure arranged in sequence from bottom to top, and more layers of micro-nano unit structures can be added according to the number of layers. The reverse-designed metasurface device of the present invention obtains other structural parameters and optical responses of free-form structures after training and optimizing the network based on the existing database, providing a new design idea for high-degree-of-freedom shapes that are difficult to obtain in the forward design method. The present invention utilizes the parallel computing power of the GPU to improve computing performance, and integrates multiple algorithms to avoid problems such as local optimal solutions and difficulty in convergence, with low computing resource loss and excellent computing results. The overlay process is used for multi-layer preparation, which is simple in process and has high inter-layer alignment accuracy.

Description

Technical Field

[0001] The invention relates to a reverse-designed metasurface device and a preparation method thereof, belonging to the technical field of optical metasurfaces. Background Art

[0002] Metasurface is a two-dimensional planar electromagnetic wave control element formed by a periodic arrangement of many sub-wavelength units (hereinafter collectively referred to as micro-nano units). With the arbitrary selection and arrangement of the units, a very high degree of design freedom can be obtained, thereby realizing the flexible control of optical parameters such as the amplitude, polarization, and phase of the electromagnetic wave, which is sufficient to meet most application requirements. As a two-dimensional planar element for controlling light beams, optical metasurfaces (hereinafter collectively referred to as metasurfaces) have been widely studied due to their excellent characteristics such as ultra-lightness, multi-function, and easy integration. They have shown immeasurable application potential, and have gradually been applied to near-eye display, biological detection, and national defense science and technology in the form of optical devices represented by metalenses and metasurface holograms. However, at present, when designing metasurfaces, data-driven forward design methods are often adopted, that is, physical models are established with the help of Mie scattering, equivalent waveguide and other theories for solution. However, in the face of design conditions such as spectrum and dispersion, it is difficult to establish an intuitive numerical mapping relationship between the structural parameters of the metasurface and the required optical performance. Therefore, it is necessary to generate an optical response database of micro-nano units through a large number of simulation calculations, and then complete the design by searching for ideal micro-nano structural parameters in the database. This huge demand for computing resources is particularly exhausting when facing large-area, multi-functional application scenarios, and it is difficult to calculate the ideal optimal value among a large number of local optimal solutions, thus limiting its scope of application. Summary of the invention

[0003] The present invention provides a reverse-designed metasurface device and a preparation method thereof.

[0004] The technical solution adopted by the present invention to solve the above technical problems is:

[0005] A reverse-designed metasurface device includes a transparent medium substrate, a first layer of dielectric micro-nano unit structure, a HSQ separation layer, and a second layer of dielectric micro-nano unit structure arranged in sequence from bottom to top. More layers of micro-nano unit structure can be added depending on the number of layers.

[0006] Preferably, the first layer of dielectric micro-nano unit structure and the second layer of dielectric micro-nano unit structure can select corresponding materials according to different wavelength bands, including TiO2, HfO2, GaN, and Si.

[0007] Preferably, the height range of the first layer of dielectric micro-nano unit structures and the second layer of dielectric micro-nano unit structures is 200-1500 nm.

[0008] Preferably, the shapes of the first layer of dielectric micro-nano unit structures and the second layer of dielectric micro-nano unit structures include regular shapes and free shapes, wherein the regular shapes include rectangles, circles, and the like.

[0009] A design method for a reverse-designed metasurface device includes firstly using numerical simulation software to calculate regular shapes and establish a database in order to obtain a training data set required for reverse design, i.e., the data set required for training.

[0010] Secondly, the data set is divided into a training set and a test set, and then a reinforcement learning network is constructed. A supervised learning regression model is constructed for training in both the forward and reverse calculation processes. The neural network is trained by building a regression model for the training set. After the neural network training is completed, the size parameters of the micro-nano structure are obtained. Next, the test set is used to predict the optical response it can produce, and the performance of the network model is evaluated by the determination coefficient and training test time in the form of cross-validation and parameter selection. In the reverse regression model, the difference between the forward prediction result and the ideal value is compared, and the size parameters of the micro-nano structure are adjusted in a feedback manner. The process of forward training prediction-reverse optimization is carried out until the ideal result is obtained.

[0011] Finally, the micro-nano structure is divided into n×n small pixel structures. The pixel at each position is binarized to determine whether the pixel is filled with the structure, and then the pixel image and the required optical response are used as the input of the network to construct a generative adversarial network. The three convolutional neural networks of simulator, generator and discriminator are used to forward train the binary structure parameters. The weight function of the generator is not only affected by the discriminator and simulator, but also by comparing the difference between the shape of the ideal micro-nano structure and the structure generated by the generator, which is used as the loss function to reversely optimize the weight parameters of the generator, so as to continuously update the shape parameters. When the loss function reaches a certain threshold, the pixelated image is smoothed and passed through a fixed weight pre-trained simulator, that is, the network accuracy is evaluated and verified without numerical simulation. The reverse design method is used to train and optimize the network to obtain other structural parameters and the optical response of the free-form structure.

[0012] A method for preparing a reverse-designed metasurface device, the method comprising the following steps: selecting a conformal filling process for a single-layer metasurface for preparation, firstly spin-coating polymethyl methacrylate and ZEP and other resists on a transparent medium substrate, and then using photolithography technology to expose and develop and fix after baking; then using atomic layer deposition technology to fill the exposed holes, and using ion beam etching to remove the resist on the sample surface, and finally using reactive ion etching to remove the residual photoresist on the sample surface to obtain a single-layer metasurface.

[0013] Compared with the prior art, the reverse-designed metasurface device of the present invention obtains other structural parameters and optical responses of free-form structures after training and optimizing the network based on the existing database by using the reverse design method, providing a new design idea for high-degree-of-freedom shapes that are difficult to obtain in the forward design method. The present invention perfectly utilizes the parallel computing power of the GPU to improve computing performance, and integrates multiple algorithms to avoid problems such as local optimal solutions and difficulty in convergence. The computing resource loss is low and the computing results are excellent. The present invention adopts an overlay process to prepare multiple layers, which is simple in process and has high inter-layer alignment accuracy. The present invention can be applied to a variety of applications, and corresponding structural parameters can be generated according to different input optical requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the design route of the reverse-engineered metasurface device of the present invention.

[0015] Figure 2 Schematic diagram of regular-shaped micro-nano units and free-shaped micro-nano units in the reverse-designed metasurface device of the present invention.

[0016] Figure 3 This is a flow chart of the preparation process of the reverse-engineered metasurface device of the present invention.

[0017] Figure 4 This is a functional schematic diagram of the reverse-engineered metasurface device of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below with reference to the accompanying drawings.

[0019] like Figures 1 to 4 As shown, a reverse-designed metasurface device includes a transparent medium substrate, a first layer of dielectric micro-nano unit structure, a HSQ separation layer and a second layer of dielectric micro-nano unit structure arranged in sequence from bottom to top. More layers of micro-nano unit structure can be added depending on the number of layers.

[0020] Metasurfaces can be provided with different application requirements, including structural color, spectrum, dispersion requirements, etc.

[0021] The metasurface design scheme is suitable for different bands according to different application requirements, including visible light band, ultraviolet light band and infrared light band.

[0022] The first layer of dielectric micro-nano unit structure and the second layer of dielectric micro-nano unit structure can select corresponding materials according to different bands, including TiO2, HfO2, GaN, Si, etc.

[0023] The height range of the first layer dielectric micro-nano unit structure and the second layer dielectric micro-nano unit structure is 200-1500nm, and the shapes of the first layer dielectric micro-nano unit structure and the second layer dielectric micro-nano unit structure include regular shapes and free shapes, wherein the regular shapes include rectangles, circles, etc.

[0024] A reverse design method for metasurface devices: When designing a metasurface, it is necessary to input the required spectrum, dispersion and other response curve parameters and the corresponding data sets, and construct forward training and reverse optimization models at the same time. During the forward training process, the optical response of given structural parameters is predicted, and the difference between the ideal result and the predicted result is fed back to the structural parameters through reverse optimization. After several iterations, the design of the metasurface is completed, and the accuracy of the design results is verified by electromagnetic field simulation software and other methods.

[0025] A method for preparing a metasurface:

[0026] A conformal filling process is selected for the preparation of a single-layer metasurface. First, polymethyl methacrylate (PMMA) and ZEP and other resists are spin-coated on a transparent medium substrate. After baking, they are exposed and developed using photolithography technology (electron beam exposure, nanoimprinting, etc.). Subsequently, atomic layer deposition technology is used to fill the exposed holes, and ion beam etching is used to remove the resist on the sample surface. Finally, reactive ion etching is used to remove the residual photoresist on the sample surface to obtain a single-layer metasurface.

[0027] An overlay process is selected for the preparation of the multi-layer metasurface. After the first layer of the metasurface is prepared, an alignment mark is additionally exposed, and then HSQ and other resists are spin-coated to space and protect different metasurface layers. Next, the single-layer metasurface preparation process is repeated based on the alignment mark to prepare a multi-layer structure that can be precisely aligned.

[0028] First, in order to obtain the training data set required for reverse design, it is necessary to explore the propagation mechanism of electromagnetic waves in micro-nano units based on Mie scattering and equivalent waveguide theories, and then use numerical simulation software to calculate regular shapes (such as isotropic rectangles, circles and anisotropic shapes such as Figure 2 The optical responses such as amplitude and phase shown are used to establish a database, that is, the data set required for training.

[0029] For the reverse design of ordinary single micro-nano structure, the present invention divides the data set into a training set and a test set, and then constructs a reinforcement learning network, and constructs a supervised learning regression model for training in both the forward and reverse calculation processes. By establishing a Gaussian process regression model, a decision tree regression model, a multi-layer perceptron and other regression models to train the neural network, the size parameters of the micro-nano structure are obtained after the network training is completed, and then the test set is used to predict the optical response that it can produce, and the performance of the network model is evaluated by the determination coefficient and the training test time in the form of cross-validation and parameter selection. In the reverse regression model, the difference between the result of the forward prediction and the ideal value is compared, and the size parameters of the micro-nano structure are adjusted in a feedback manner, and the process of forward training prediction-reverse optimization is repeated until the ideal result is obtained. Among them, the use of reinforcement learning in the reverse regression model can effectively overcome the defects of non-unique solutions and slow convergence speed in traditional supervised learning.

[0030] In order to improve the degree of freedom of the shape of the micro-nano structure, the present invention divides the micro-nano structure into n×n small pixel structures, and determines whether the pixel is filled with the structure by binarizing the pixel at each position (0 / 1), and then taking the pixel image and the required optical response as the input of the network to construct a generative adversarial network. The three convolutional neural networks of the simulator, the generator and the discriminator are used to forward train the binarized structure parameters, wherein the weight function of the generator is not only affected by the discriminator and the simulator, but also by comparing the difference between the shape of the ideal micro-nano structure and the structure generated by the generator, and using this as the loss function to reversely optimize the weight parameters of the generator, so as to continuously update the shape parameters. When the loss function reaches a certain threshold, the pixelated image is smoothed, and after a pre-trained simulator with fixed weights, the network accuracy can be evaluated and verified without numerical simulation.

[0031] The reverse-designed metasurface device of the present invention uses a reverse design method based on an existing database to obtain other structural parameters and optical responses of free-form structures after training and optimizing the network, providing a new design idea for high-degree-of-freedom shapes that are difficult to obtain in forward design methods. The present invention perfectly utilizes the parallel computing power of the GPU to improve computing performance, and integrates multiple algorithms to avoid problems such as local optimal solutions and difficulty in convergence. The computing resource loss is low and the computing results are excellent. The present invention adopts an overlay process to prepare multiple layers, which is simple in process and has high inter-layer alignment accuracy. The present invention can be applied to a variety of applications, and corresponding structural parameters can be generated according to different input optical requirements.

[0032] Finally, it should be noted that the above embodiments only illustrate the technical solutions of the present invention, rather than limiting it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A design method for a reverse-engineered metasurface device, characterized in that: Firstly, in order to obtain the training data set required for reverse design, numerical simulation software is used to calculate regular shapes and establish a database, i.e., the data set required for training; Secondly, the data set is divided into a training set and a test set, and then a reinforcement learning network is constructed. A supervised learning regression model is constructed for training in both the forward and reverse calculation processes. The neural network is trained by building a regression model on the training set. After the neural network training is completed, the size parameters of the micro-nano structure are obtained. Next, the test set is used to predict the optical response that it can produce, and the performance of the network model is evaluated by the determination coefficient and the training test time in a cross-validation and parameter selection manner. In the reverse regression model, the difference between the forward prediction result and the ideal value is compared, and the size parameters of the micro-nano structure are adjusted in a feedback manner, and the process of forward training prediction-reverse optimization is carried out until the ideal result is obtained. Finally, the micro-nano structure is divided into n×n small pixel structures. The pixel at each position is binarized to determine whether the pixel is filled with the structure, and then the pixel image and the required optical response are used as the input of the network to construct a generative adversarial network. The three convolutional neural networks of the simulator, generator and discriminator are used to forward train the binary structure parameters. The weight function of the generator is not only affected by the discriminator and the simulator, but also by comparing the difference between the shape of the ideal micro-nano structure and the structure generated by the generator, which is used as the loss function to reversely optimize the weight parameters of the generator, so as to continuously update the shape parameters. When the loss function reaches a certain threshold, the pixelated image is smoothed and passed through a pre-trained simulator with fixed weights, that is, the network accuracy is evaluated and verified without numerical simulation. The inverse design method is used to train and optimize the network to obtain other structural parameters and the optical response of the free-form structure.

2. A metasurface device using the design method of a metasurface device reversely designed according to claim 1, characterized in that: It includes a transparent medium substrate and a first layer of dielectric micro-nano unit structure, a HSQ separation layer and a second layer of dielectric micro-nano unit structure arranged in sequence from bottom to top. More layers of micro-nano unit structure can be added according to the number of layers. When designing a metasurface device, it is necessary to input the required spectrum, dispersion response curve parameters and corresponding data sets, and construct a forward training and reverse optimization model at the same time. During the forward training process, the optical response of the given structural parameters is predicted, and the difference between the ideal result and the predicted result is fed back to the structural parameters through reverse optimization. After several iterations, the design of the metasurface is completed.

3. The metasurface device according to claim 2, characterized in that: The first layer of dielectric micro-nano unit structure and the second layer of dielectric micro-nano unit structure select corresponding materials according to different wavebands, including TiO2, HfO2, GaN, and Si.

4. The metasurface device according to claim 2, characterized in that: The height range of the first layer of dielectric micro-nano unit structure and the second layer of dielectric micro-nano unit structure is 200-1500 nm.

5. The metasurface device according to claim 2, characterized in that: The shapes of the first layer of dielectric micro-nano unit structures and the second layer of dielectric micro-nano unit structures include regular shapes and free shapes, wherein the regular shapes include rectangles and circles.

Images