A circular dichroism metasurface based on reinforcement learning and a design and preparation method thereof

Abstract

This invention relates to a circular dichroic metasurface based on reinforcement learning and its design and fabrication method. By combining neural networks and genetic algorithms, and employing reinforcement learning algorithms for pixel-like two-dimensional planar topologically connected structures, this invention enables rapid inverse design of specific transmission circular dichroic spectra within the infrared band, yielding the micro / nano structural units of the two-dimensional metasurface, effectively improving design efficiency. The invention utilizes a two-dimensional planar topologically connected structure, maximizing design freedom and avoiding noise in the design process. The design method provided by this invention significantly improves bandwidth and design freedom, adapting to various target designs, including broadband and narrowband applications. Furthermore, the design speed is fast; based on computer configuration, iterative optimization of the structure can be completed within one to several minutes to obtain the optimal micro / nano structural units under the target conditions. Periodic extension within a transverse two-dimensional plane yields the circular dichroic metasurface.

Description

Technical Field

[0001] This invention relates to a method for reverse design and fabrication of circular dichroic metasurfaces based on hybrid neural networks, and particularly to a method for reverse design of metasurface patterns with different circular dichroic spectra, belonging to the field of polarization optical metamaterials. Background Technology

[0002] Chiral metasurfaces are nanostructured materials with unique optical properties. They typically consist of a two-dimensional array of micro / nanostructures (superatoms). These units are designed to precisely control the phase and amplitude of incident light and to exhibit different responses to different polarization states. A key property of chiral metasurfaces is circular dichroism (CD), which is the difference in transmittance between right-handed circularly polarized (RCP) and left-handed circularly polarized (LCP) light (hereinafter uniformly defined as CD = T). RCP -T LCP The circular dichroism of chiral metasurfaces can be achieved by adjusting the geometric parameters, material properties, and external excitations of the unit cells. These unique properties give chiral metasurfaces broad application potential in various fields, including polarization detection, biomedical molecular recognition, sensing, optical communication, holographic imaging, and other fields. They can be used to fabricate efficient optical polarization controllers, optical components, optical scatterers, and optical sensors. Furthermore, due to their compact structure and designability, circular dichroic metasurfaces also offer new possibilities for the miniaturization and integration of optical devices.

[0003] Existing chiral metasurface designs often require selecting suitable, promising geometries for the superatomic unit, involving a lengthy trial-and-error process of potential superatomic structure shapes, often relying on human design experience; followed by parameter optimization one by one, both steps depending on complex and time-consuming numerical computations. It can be seen that the performance of all these previous studies is generally limited by experience-driven, pre-assigned substructures, which are typically limited to a fixed number of geometric parameters and lack precisely adjustable degrees of freedom, leading to a mismatch between the chiral optical response and the target chirality.

[0004] In recent years, the use of neural networks (NNs) for photonic device design has yielded certain results (see references: The inverse design of structural color using machine learning[J]. Nanoscale,2019, 11(45): 21748–21758; A Bidirectional Deep Neural Network for AccurateSilicon Color Design[J]. Advanced Materials, 2019, 31(51): 1905467), including forward prediction of multiple polarization spectra based on structure and reverse generation of target structures based on spectral inversion. Among them, the forward prediction neural network has shown quite accurate prediction ability; however, due to high-dimensional nonlinear electromagnetic processes and potential one-to-many mapping problems, the trained inverse neural network may not be able to accurately reproduce a series of spectral properties in physical space, resulting in a large gap between the actual spectral properties of the inversely designed structure and the expected requirements, especially for circular dichroism, a high-dimensional physical property that is manifested by the combined effect of two circular polarization transmittances. To improve the performance of inverse neural networks, researchers have also attempted to integrate neural networks with various traditional optimization algorithms, using reinforcement learning algorithms that employ neural networks as solvers for traditional optimization algorithms. These algorithms have been applied to, for example, the inverse design of terahertz reflective circular dichroic metasurfaces (see: Probabilistic Representation and Inverse Design of Metamaterials Based on a Deep Generative Model with Semi‐Supervised Learning Strategy[J]. Advanced Materials, 2019, 31(35): 1901111). However, this approach can only predict structures with simple single resonance peaks and designs of polarization-conversion metasurface units (see: Compounding Meta‐Atoms into Metamolecules with Hybrid Artificial Intelligence Techniques[J]. Advanced Materials, 2020, 32(6): 1904790). Furthermore, this design is only applicable to a single wavelength of 800 nm. Overall, neural networks specialized for circular dichroism design remain scarce and ineffective. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a reinforcement learning algorithm that combines neural networks with genetic algorithms for pixel-based two-dimensional planar topological single-connected structures. This algorithm can rapidly design and invert the micro / nano structure units of two-dimensional metasurfaces based on broadband specific transmission circular dichroism spectra within the infrared band, effectively improving design efficiency.

[0006] The technical solution to achieve the objective of this invention is to provide a design method for circular dichroic metasurfaces based on reinforcement learning, which employs reverse design and includes the following steps: (1) The metasurface unit structure to be designed is a pixel-type topological planar two-dimensional structure with a structure height of 200-500 nm and a design wavelength range of 1300-1700 nm in the near-infrared band. (2) Construct an electromagnetic simulation model between a randomly generated pixel-based topological planar two-dimensional metasurface and a designed transmission spectrum within a wavelength range, generating a dataset containing a large number of samples; the dataset includes: Randomly generated pixel-like patterns representing the shapes of different micro / nano structures; Target circular dichroism spectral data of metasurfaces with corresponding micro / nano structure shapes within the designed wavelength range; (3) The neural network is trained using the dataset constructed in step (2) to obtain two neural networks for left-handed circularly polarized light and right-handed circularly polarized light, respectively; (4) Using the neural network trained in step (3) as a fast computational unit for processing the structure-spectral response relationship, a genetic algorithm is used to construct a reverse design process from the target circular dichroism spectrum to the required metasurface unit structure. (5) Set the target circular dichroism spectrum and use the reverse design process of step (4) to obtain the optimal micro-nano structure unit under the target conditions; (6) The micro-nano structure unit designed in step (5) is periodically extended in a transverse two-dimensional plane to obtain a circular dichroic metasurface.

[0007] The metasurface unit structure described in this invention contains N×N pixels in a single period and has rotational symmetry; in the design, the unit structure is represented by a binary matrix pixel pattern and stored as an N×N matrix; N ∈ 18~100.

[0008] The target circular dichroism spectrum described in this invention consists of a left-handed circularly polarized spectrum and a right-handed circularly polarized spectrum, which are stored as M×1 vectors, where M ∈ 101~1001.

[0009] The technical solution of the present invention also includes a circular dichroic metasurface based on reinforcement learning obtained according to the above design method.

[0010] The preparation method of the circular dichroic metasurface based on reinforcement learning described in this invention includes the following steps: (1) Amorphous silicon was deposited on a cleaned silica substrate using plasma-enhanced chemical vapor deposition; (2) The pattern of the circular dichroic metasurface obtained based on reinforcement learning design is exposed to the sample by electron beam lithography, and then developed and fixed to obtain a photoresist mask of the circular dichroic metasurface pattern. (3) An inductively coupled plasma etching was performed using SF6 and CHF3 gases to obtain a circular dichroic metasurface designed based on reinforcement learning.

[0011] The optimization target of the technical solution of this invention is a two-dimensional planar topological single-connected structure, thereby maximizing the degree of freedom in the design and avoiding noise in the design (too many noises will form a mosaic-like pattern. Due to the existing fabrication process, it is not practically possible to use metasurface micro-nano devices in the infrared and visible light bands, so it needs to be avoided in the design).

[0012] Compared with existing technologies, the advantages of this invention are as follows: Traditional design methods relying on optimization algorithms are limited by complex and time-consuming numerical calculations, making them unable to design quickly or use a large population size, thus making it difficult to achieve optimal design results. Compared with other existing technologies that use inverse neural networks for reverse design, this invention has significantly improved bandwidth and design freedom, can adapt to various target designs including broadband and narrowband, and has a faster design speed, providing iterative optimization structures within one to several minutes based on computer configuration. Attached Figure Description

[0013] Figure 1 This is a flowchart of reinforcement learning for the reverse design of circular dichroic metasurfaces based on hybrid neural networks, provided in an embodiment of the present invention. Figure 2 This is a neural network structure layout diagram provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the micro / nano structure unit of the chiral metasurface designed in reverse according to an embodiment of the present invention; Wherein, 1 is a two-dimensional amorphous silicon structure layer; 2 is a silicon dioxide substrate; and 3 is a pixel pattern of a matching structure. Figure 4 This invention provides a chiral metasurface designed for broadband circular dichroism targets, wherein: Figure a is a schematic diagram of the structure; Figure b is a 45° oblique view; Figure c is a scanning electron microscope image of the top view; and Figure d is a scanning electron microscope image of the 45° oblique view. Figure 5This invention provides a chiral metasurface designed for a peak-shaped narrow-band circular dichroism target, wherein: Figure a is a schematic diagram of the structure; Figure b is a 45° oblique view; Figure c is a scanning electron microscope image of the top view; and Figure d is a scanning electron microscope image of the 45° oblique view. Figure 6 This invention provides a chiral metasurface designed for a valley-shaped narrow-band circular dichroism target, wherein: Figure a is a schematic diagram of the structure; Figure b is a 45° oblique view; Figure c is a scanning electron microscope image of the top view; and Figure d is a scanning electron microscope image of the 45° oblique view. Detailed Implementation

[0014] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. Example 1

[0015] This embodiment provides a reinforcement learning algorithm that combines neural networks and genetic algorithms for designing micro / nano structural units with a target circular dichroism spectrum or that are closest to the target circular dichroism spectrum within the design range. The algorithm includes the following steps: (1) The metasurface unit structure to be designed is a pixel-type topological planar two-dimensional structure with a structure height of 200-500 nm and a design wavelength range of 1300-1700 nm in the near-infrared band. (2) Construct an electromagnetic simulation model between a randomly generated pixel-based topological planar two-dimensional metasurface and a designed transmission spectrum within a wavelength range, generating a dataset containing a large number of samples; the dataset includes: Randomly generated pixel-like patterns representing the shapes of different micro / nano structures; Target circular dichroism spectral data of metasurfaces with corresponding micro / nano structure shapes within the designed wavelength range; (3) The neural network is trained using the dataset constructed in step (2) to obtain two neural networks for left-handed circularly polarized light and right-handed circularly polarized light, respectively; (4) Using the neural network trained in step (3) as a fast computational unit for processing the structure-spectral response relationship, a genetic algorithm is used to construct a reverse design process from the target circular dichroism spectrum to the required metasurface unit structure. (5) Set the target circular dichroism spectrum and use the reverse design process of step (4) to obtain the optimal micro-nano structure unit under the target conditions.

[0016] See appendix Figure 1 and 2 , Figure 1 The genetic algorithm shown serves as the main framework, and will be based on Figure 2The neural network trained on the structural layout serves as a fast computational tool for circular dichroism spectroscopy, enabling rapid genetic iteration optimization of the patterns corresponding to the required two-dimensional metasurface micro / nano structural units for specific circular dichroism spectra.

[0017] The genetic algorithm includes evolutionary processes such as selection, crossover, and mutation. The fitness evaluation is given by the correlation between the current structure's circular dichroism prediction value and the target's circular dichroism. Based on different targets, multiple estimation modes can be selected, including maximum value, global root mean square error (MSE), and similarity evaluation.

[0018] The neural network dataset is based on physical computational models established using algorithms such as Finite-Difference Time-Domain (FDTD), Finite Element Method (FEM), and Rigorous Coupled-Wave Analysis (RCWA). A dataset containing numerous pixel-based patterns and the circularly polarized spectra of the target is created and divided into training, testing, and validation sets in a 7:2:1 ratio. The neural network employs a residual neural network, comprising an image input layer, a Fourier Transform (FFT) coding layer, convolutional layers, pooling layers, activation layers (using Leaky-ReLU), flattening layers, linear layers, and residual blocks composed of these layers. The final output layer consists of three linear layers with output sizes of 3000, 1000, and 201 (the final output size matches the selected spectral data vector length M), and a gated recurrent unit layer is added to further reduce training error. In the hyperparameters, the learning rate is set to 1e-4, and the loss function is defined as the root mean square error (MSE). The process is repeated 1000 times or until the convergence condition is met, i.e., the MSE is less than 1e-5 on the training set and less than 1e-4 on the test set. Finally, the accuracy of the neural network prediction is verified by random sampling on the validation set.

[0019] The pixel-like patterns and corresponding micro / nano structure units provided by this invention are illustrated in the following diagram. Figure 3 As shown, the micro / nano structure unit includes: a two-dimensional amorphous silicon structure layer 1 with a period P of 900 nm and a thickness of... h 2. 330 nm, silicon dioxide (glass) substrate; 3. Pixel pattern 3 matching micro / nano structural units, with 36×36 pixels.

[0020] The designed micro / nano structural units are periodically extended in a transverse two-dimensional plane to obtain a circular dichroic metasurface. Example 2

[0021] This embodiment provides a method for designing and preparing chiral metasurfaces based on reinforcement learning.

[0022] This embodiment adopts a topological pixel-based planar two-dimensional structure, which includes a pixel-based planar topological single-connected structure and a silicon dioxide substrate in sequence along the light incident direction; the pixel-based planar topological single-connected structure can be made of various infrared band materials, including dielectric materials such as titanium dioxide, aluminum oxide, and silicon, and metallic materials such as silver and gold; in this embodiment of the invention, amorphous silicon material is used.

[0023] The period of the pixel-based planar topological single-connected structure can be selected within the high resonant region of the subwavelength range of the design wavelength. In this embodiment, the preferred period is 900 nm and the height is 330 nm. Since the structure is uniform in the height direction, it can be characterized by a two-dimensional black and white pixel pattern, thereby simplifying the design process. Pixels are defined as 1 (Si, white) and 0 (etched area, black). A single period contains N×N pixels (N ∈ 18-100), which can be determined according to the specific operating wavelength and the required precision of the micro-nano fabrication process. In this embodiment, N = 36, and the size of each pixel is 25 nm.

[0024] This embodiment establishes an electromagnetic simulation model between a pixel-like planar two-dimensional structure periodic model and a transmission spectrum. A dataset containing a large number of samples is generated using computer calculations. The dataset includes: randomly generated pixel patterns representing different micro / nano structure shapes (stored as an N×N matrix); and the target circular dichroism spectrum of the metasurface with the micro / nano structure shape for the required wavelength range (the two required circularly polarized spectral data are stored as M×1 vectors). The size of M depends on the size of the required wavelength range and the required wavelength sampling precision, typically ranging from 101 to 1001. In this embodiment, the required wavelength range is 1300–1700 nm, and the sampling precision is 2 nm; therefore, M = 201.

[0025] Based on the above dataset, two neural networks were trained for left-handed and right-handed circularly polarized light, respectively. The trained neural networks were then combined with a genetic algorithm to construct a complete reverse design scheme from the target circular dichroism spectrum to the required metasurface unit structure, resulting in the pattern of the circular dichroism metasurface, which is used to prepare chiral metasurfaces.

[0026] The method for preparing a chiral metasurface using the reverse design scheme of this embodiment to obtain the desired metasurface unit structure includes the following steps: (1) Place the silicon dioxide substrate in acetone solution and alcohol solution in sequence and sonicate for 10 minutes to remove dirt, then clean with flowing deionized water and blow dry with high pressure nitrogen.

[0027] (2) Amorphous silicon with a thickness of 330 nm was deposited on a cleaned silica substrate using plasma-enhanced chemical vapor deposition (PECVD).

[0028] (3) Spin-coating a 360 nm thick ZEP 520A film as an electron beam photoresist.

[0029] (4) Spin-coating AR-PC 5090 thin film as a conductive layer to avoid charge accumulation effect.

[0030] (5) The pattern of the circular dichroic metasurface obtained by reinforcement learning design in this embodiment is exposed to the sample by electron beam lithography, and then developed by ZED-N50 and fixed with isopropanol to obtain a photoresist mask of the circular dichroic metasurface pattern.

[0031] (6) Etching is performed at room temperature using inductively coupled plasma etching process, with SF6 and CHF3 gases introduced for etching until the etching is complete.

[0032] (7) Remove the residual photoresist with acetone and anhydrous ethanol, and then dry it with high-pressure nitrogen to obtain a circular dichroic metasurface designed based on reinforcement learning. Example 3

[0033] Following the reinforcement learning design method provided in Example 1, a circular dichroism metasurface is designed; the optimization target is selected as having a circular dichroism of 1 (T0) in a broadband range of 1400–1700 nm. RCP For 1, T LCP The chiral metasurface sample was prepared according to the preparation method provided in Example 2 (where the value is 0).

[0034] like Figure 4 As shown, Figure a is the optimized two-dimensional pattern and the corresponding structural schematic diagram; Figure c is a scanning electron microscope image of the top view of the structure; Figure d is a scanning electron microscope image of the structure viewed at a 45° angle; Figure b is the transmittance of the simulation (lines) and the experiment (marked as circles). Example 4

[0035] Following the reinforcement learning design method provided in Example 1, a circular dichroic metasurface is designed. The optimization target is a narrow-band circular dichroic peak (T0) with a peak shape in the range of 1400–1700 nm. RCP T is 1 at a certain wavelength and 0 at all other wavelengths. LCP (Always 0); according to the preparation method provided in Example 2, the corresponding chiral metasurface sample was prepared. Figure 5 As shown, Figure a is the optimized two-dimensional pattern and the corresponding structural schematic diagram; Figure c is a scanning electron microscope image of the top view of the structure; Figure d is a scanning electron microscope image of the structure viewed at a 45° angle; Figure b is the transmittance of the simulation (line) and the experiment (circle mark). Example 5

[0036] Following the reinforcement learning design method provided in Example 1, a circular dichroic metasurface is designed. The optimization target is a narrow-band circular dichroic peak with a valley shape (T0) existing in the range of 1400–1700 nm. RCP Always 1, T LCP A chiral metasurface sample was prepared using the preparation method provided in Example 2 (with a value of 0 at a certain wavelength and 1 at all other wavelengths). Figure 6 As shown, Figure a is the optimized two-dimensional pattern and the corresponding structural schematic diagram; Figure c is a scanning electron microscope image of the top view of the structure; Figure d is a scanning electron microscope image of the structure viewed at a 45° angle; Figure b is the transmittance of the simulation (lines) and the experiment (marked as circles).

Claims

1. A method for designing a circular dichroism metasurface based on reinforcement learning, adopting inverse design, characterized in that Includes the following steps: (1) The metasurface unit structure to be designed is a pixel-type topological planar two-dimensional structure with a structure height of 200-500 nm and a design wavelength range of 1300-1700 nm in the near-infrared band. (2) Construct an electromagnetic simulation model between a randomly generated pixel-based topological planar two-dimensional metasurface and a designed transmission spectrum within a wavelength range, generating a dataset containing a large number of samples; the dataset includes: Randomly generated pixel-like patterns representing the shapes of different micro / nano structures; Target circular dichroism spectral data of metasurfaces with corresponding micro / nano structure shapes within the designed wavelength range; (3) The neural network is trained using the dataset constructed in step (2) to obtain two neural networks for left-handed circularly polarized light and right-handed circularly polarized light, respectively; (4) Using the neural network trained in step (3) as a fast computational unit for processing the structure-spectral response relationship, a genetic algorithm is used to construct a reverse design process from the target circular dichroism spectrum to the required metasurface unit structure. (5) Set the target circular dichroism spectrum and use the reverse design process of step (4) to obtain the optimal micro-nano structure unit under the target conditions; (6) The micro-nano structure unit designed in step (5) is periodically extended in a transverse two-dimensional plane to obtain a circular dichroic metasurface.

2. The design method for a circular dichroic metasurface based on reinforcement learning according to claim 1, characterized in that: The metasurface unit structure described herein contains N×N pixels in a single period and exhibits rotational symmetry; In the design, the unit structure is represented by a binary matrix pixel pattern and stored as an N×N matrix; N ∈ 18～100。 3. The design method for a circular dichroic metasurface based on reinforcement learning according to claim 1, characterized in that: The target circular dichroism spectrum consists of a left-handed circularly polarized spectrum and a right-handed circularly polarized spectrum, which are stored as M×1 vectors, where M ∈ 101~1001.

4. A circular dichroic metasurface based on reinforcement learning, obtained by the design method of claim 1.

5. The method of claim 4, wherein the circular dichroism metasurface is prepared based on reinforcement learning. Includes the following steps: (1) Amorphous silicon was deposited on a cleaned silica substrate using plasma-enhanced chemical vapor deposition; (2) The pattern of the circular dichroic metasurface obtained based on reinforcement learning design is exposed to the sample by electron beam lithography, and then developed and fixed to obtain a photoresist mask of the circular dichroic metasurface pattern. (3) An inductively coupled plasma etching was performed using SF6 and CHF3 gases to obtain a circular dichroic metasurface designed based on reinforcement learning.

Images