System for performing far-field imaging on all-dielectric metasurface based on physical embedded machine learning aided design

Abstract

The invention provides a far-field imaging system based on a physical embedded machine learning aided design all-dielectric metasurface. The method is characterized in that the method is composed of a target image 101, a neural network 102, metasurface small unit structure parameters 103, a pre-trained neural network 104, metasurface amplitude and phase 105, a physical embedding system 106, a far-field imaging graph 107, and a mean square error 108 of the imaging graph calculated by the target image and the neural network. A far-field pattern that reproduces a transverse profile of light using complex amplitude information shows more excellent performance than a far-field pattern calculated only using amplitude or phase information. The metasurface array structure can be used for designing and calculating the arrangement mode of the metasurface units and the size information of all media so as to achieve different imaging effects of the same metasurface array structure under different frequency points, and can be widely applied to the fields of large-size intelligent structure imaging and the like.

Description

(1) Technical Field

[0001] The present invention relates to a machine learning-assisted design for terahertz band far-field imaging technology, which can be used for far-field imaging of terahertz waves and belongs to the field of terahertz communication. (2) Background Art

[0002] The field of metasurfaces has recently provided extraordinary capabilities to control electromagnetic waves as needed and has shown many interesting phenomena and promising applications, including negative and zero refractive indices, invisibility cloaks, super-resolution imaging, electromagnetic induced resonance, and topological insulators. However, for metamaterials, manufacturing challenges and material losses are usually large and inevitable, thus hindering their practical applications. Recently, reconfigurable design metasurfaces have been proposed for unprecedented control of the wavefront of electromagnetic waves. The subwavelength thickness of metasurfaces has significantly lower optical losses, along with reduced complexity of the manufacturing process. Therefore, metasurfaces have opened up an unprecedented path for shaping the wavefront of transmitted radiation and thus providing fascinating properties that may lead to revolutionary application-oriented photonic devices, such as high-resolution holograms, flat lenses, anomalous reflection and refraction, and quarter-wave plates. The application space of metasurfaces almost covers the entire electromagnetic spectrum, including the technically important terahertz range.

[0003] Conventional far-field imaging works by recording the interference pattern of light scattered by an object and coherent light beams, which contain the amplitude and phase information of the scattered light wavefront for reconstructing a 3D image. However, difficulties arise with this technique because it requires a real object and a highly time- and space-coherent light source. Subsequently, a new method for generating far-field maps has been proposed, including using numerical calculations to compute the phase distribution at the far-field map interface and encoding this information into a specific surface structure or spatial light modulator (SLM). This method is widely known as computer-generated holography (CGH). The use of metasurfaces as a new scheme for holograms has attracted much attention in the photonics and plasmonics communities. In most recent demonstrations, only the phase information has been designed while the amplitude remains constant, resulting in inevitable limitations in obtaining high-resolution perfect images.

[0004] To solve the above problems, Wang Qiu et al. proposed a broadband metasurface far-field map fabrication scheme in 2016 that simultaneously achieves five-level amplitude modulation and eight-level phase modulation in the terahertz band. The paper DOI is 10.1038 / srep32867. Using C-shaped split-ring resonators (CSRRs) as the basic unit structure, the designed far-field map has high robustness, high resolution, and broadband characteristics without conjugate images. The far-field map was experimentally characterized using a near-field scanning terahertz microscope (NSTM). However, the structures used are metal structures and there are few of them, and the search space of the design cannot be well exploited.

[0005] In view of the deficiencies of the above prior art, the present invention proposes a system for far-field imaging by physically embedding a machine learning-assisted design of all-dielectric metasurfaces. This system can be used to design the metasurface structure of all-dielectric materials. Since all-dielectric structures such as silicon have a high transmittance, they are very popular. Coupled with the physically embedded machine learning model to assist in the design of the far-field imaging of the metasurface, it can well expand the search space during design. (III) SUMMARY OF THE INVENTION

[0006] An object of the present invention is to provide a system for far-field imaging by physically embedding a machine learning-assisted design of all-dielectric metasurfaces, which has a simple and compact structure and is easy to operate and adjust.

[0007] The object of the present invention is achieved as follows:

[0008] A system for far-field imaging by physically embedding a machine learning-assisted design of all-dielectric metasurfaces. Its characteristics are: it consists of a target image 101, a neural network 102, the structural parameters of the metasurface small unit 103, a pre-trained neural network 104, the amplitude and phase of the metasurface 105, a physical embedding system 106, a far-field imaging diagram 107, the mean square error between the target image and the imaging diagram calculated by the neural network 108, an all-dielectric column 201, and a substrate 202. The forward neural network in the system, that is, the pre-trained neural network, is used to learn the relationship between the geometric parameters of the metasurface unit structure and the amplitude and phase, so as to ensure the correctness of the geometric parameters of the metasurface unit structure predicted by the physically embedded neural network in the subsequent inverse design. The physical embedding module is used to calculate the far-field imaging according to the predicted amplitude and phase during inverse design. The inverse neural network is used for machine learning to perform inverse design of the metasurface through the target imaging diagram. Then, the metasurface unit structure is arranged and simulated according to the predicted parameters to achieve the imaging effect.

[0009] In one embodiment, the neural network includes a physically embedded neural network model 6, that is, the Rayleigh-Sommerfeld formula is added.

[0010] Where U(r0) and U(r1) represent the electric fields at the point r0 on the metasurface and the point r1 on the image plane, respectively; λ is the wavelength in vacuum; n is the vector perpendicular to the image plane in the indicated orientation; r 01 is the distance between r0 and r1; and cos<n,r 01 > is the inclination factor. We discretize the virtual object into a grid to convert the integral into the superposition of multiple point sources.

[0011] In one embodiment of the terahertz metasurface unit structure, the main feature is that the material of the substrate 202 includes at least one of silicon, quartz, and sapphire.

[0012] In one embodiment of the terahertz metasurface unit structure, the main feature is that for the dimensions of the all-dielectric column 201, the height H is 150 μm, the width W is between 10 μm and 50 μm, the length L is between 80 μm and 120 μm, and its rotation angle θ is between 0° and 180°. The thickness t of the substrate 202 is 2 mm, the unit period p is 130 μm, and its material includes at least one of silicon, quartz, and sapphire.

[0013] In one embodiment of the arrangement of the terahertz metasurface unit structure, it is required that the arrangement array is a 100*100 array structure.

[0014] According to the terahertz metasurface unit structure described in claim 1, its operating frequencies are 0.8 THz and 1.0 THz, and the imaging distance is 6 mm.

[0015] The above-mentioned system for far-field imaging based on physically embedded machine learning-assisted design of all-dielectric metasurfaces adopts a forward neural network to learn the relationship between the geometric parameters 103 and the amplitude and phase 105 of the all-dielectric column 201 of the metasurface unit structure, thereby ensuring the correctness of the geometric parameters 103 of the metasurface unit structure predicted by the physically embedded neural network 106 in subsequent inverse design. Then, the physically embedded module 106 constrains the calculation of the far-field imaging 107 according to the predicted amplitude and phase 105 during inverse design. Finally, the mean square error between the target image 101 and the imaging map 107 is fed back to the neural network 102 for repeated training to achieve the prediction effect. Furthermore, the metasurface unit structure is arranged and simulated according to the predicted parameters 103 to achieve the imaging effect. (IV) BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the framework of the system for far-field imaging based on physically embedded machine learning-assisted design of all-dielectric metasurfaces of the present invention. It consists of the target image 101, the neural network 102, the metasurface small unit structure parameters 103, the pre-trained neural network 104, the amplitude and phase 105 of the metasurface, the physically embedded system 106, the far-field imaging map 107, and the mean square error 108 between the target map and the imaging map calculated by the neural network. The above-mentioned 101-108 are all Figure 1 serial numbers in

[0017] Figure 2It is the structural diagram of a meta - surface unit of the system for far - field imaging based on physically - embedded machine - learning - assisted design of all - dielectric metasurfaces. It consists of an all - dielectric column 201 and a substrate 202. H represents the height of the all - dielectric column 201, L, W, and θ represent the length, width, and rotation angle of the all - dielectric column 201 respectively, t represents the thickness of the substrate 202, and p is the period of the unit structure. Both 201 and 202 mentioned above are Figure 2 the numbers in

[0018] Figure 3 It is the layout diagram (partial) of the metasurface of the system for far - field imaging based on physically - embedded machine - learning - assisted design of all - dielectric metasurfaces. It consists of all - dielectric columns 301 and a substrate 302 constructed according to the parameter information predicted by machine learning. Both 301 and 302 mentioned above are Figure 3 the numbers in (V) Specific implementation manners

[0019] The present invention will be further elaborated below in conjunction with specific embodiments.

[0020] In the description of the present invention, it should be noted that terms such as "length", "width", "thickness", "left", "right", "angle" and similar expressions are for illustrative purposes only and are not the only implementation manners, so they should not be construed as limitations to the present invention.

[0021] Please refer to Figure 1 and Figure 2 , which are the schematic diagram of the framework and the structural diagram of the meta - surface unit of the system for far - field imaging based on physically - embedded machine - learning - assisted design of all - dielectric metasurfaces in an embodiment. The system for far - field imaging based on physically - embedded machine - learning - assisted design of all - dielectric metasurfaces includes a target image 101, a neural network 102, meta - surface unit structure parameters 103, a pre - trained neural network 104, the amplitude and phase of the meta - surface 105, a physical embedding system 106, a far - field imaging graph 107, the mean square error between the target graph and the imaging graph calculated by the neural network 108, all - dielectric columns 201, and a substrate 202.

[0022] Please refer to Figure 1 and Figure 2 , which are the schematic diagram of the framework and the structural diagram of the meta - surface unit of the system for far - field imaging based on physically - embedded machine - learning - assisted design of all - dielectric metasurfaces in an embodiment. Figure 2The structural parameters L, W, and θ of the all-dielectric column 201 are used to train the relationship between the structural parameters L, W, and θ of the metasurface unit cell and the amplitude and phase through a pre-trained neural network 104. After training, it is placed into the system architecture. The structural parameters L, W, and θ of the metasurface unit cell are predicted from the target image 101, and then the parameters are brought into the pre-trained network 104 to obtain the amplitude and phase 105. The information of the amplitude and phase 105 is put into the physically embedded neural network model 106 to obtain the far-field imaging diagram 107. The mean square error 108 between the far-field imaging diagram 107 and the target image 101 is fed back into the neural network 102 for repeated training to make the far-field imaging diagram 107 match the target image 101. Thus, the prediction effect is achieved.

[0023] Please refer to Figure 3 , which is a partial layout diagram of the metasurface of a system for far-field imaging based on physically embedded machine learning-assisted design of an all-dielectric metasurface in an embodiment. According to the structural parameters L, W, and θ predicted by the network model, a 100*100 metasurface is arranged and then simulated. Thus, the effect of far-field imaging is achieved.

Claims

1. A system for far-field imaging based on physically embedded machine learning-assisted design of all-medium metasurfaces. Its characteristics are: It consists of a target image 101, a neural network 102, a small unit structure parameter of the metasurface 103, a pre-trained neural network 104, an amplitude and phase of the metasurface 105, a physical embedding system 106, a far-field imaging map 107, a mean square error 108 between the target map and the imaging map calculated by the neural network, a full dielectric column 201, and a substrate 202. The forward neural network in the system is a pre-trained neural network, which is used to learn the relationship between the geometric parameters and the amplitude phase of the metasurface unit structure, thereby ensuring the correctness of the geometric parameters of the metasurface unit structure predicted by the physically embedded neural network in the subsequent reverse design. The physical embedding module is used to calculate the far-field imaging according to the predicted amplitude phase during the reverse design. The reverse neural network is used for machine learning to reverse design the metasurface through the target imaging map. Then, the unit structure of the metasurface is arranged and simulated according to the predicted parameters to achieve the imaging effect.

2. The neural network according to claim 1 comprises a physically embedded neural network model 106, i.e., the Rayleigh-Somfit formula is added.

3. The terahertz metasurface unit structure according to claim 1 is mainly characterized in that the material of the substrate 202 includes at least one of silicon, quartz, and sapphire.

4. The terahertz metasurface unit structure according to claim 1 is mainly characterized in that the dimensions of the all-dielectric column 201, wherein the height H is 150 μm, the width W is between 10 μm and 50 μm, the length L is between 80 μm and 120 μm, and the rotation angle θ is between 0° and 180°. The thickness t of the substrate 202 is 2 mm, the unit period p is 130 μm, and the material thereof comprises at least one of silicon, quartz, and sapphire.

5. According to the arrangement of the terahertz metasurface unit structure according to claim 1, the arrangement array is required to be a 100*100 array structure.

6. The terahertz metasurface unit structure according to claim 1 has an operating frequency of 0.8 THz and 1.0 THz, and an imaging distance of 6 mm.

Images