A D-based 2 Terahertz phased array device for NNs and implementation method thereof

Abstract

The present invention relates to a method based on D 2 Terahertz phased array device of NNs and implementation method thereof. The phased array device includes an active modulation layer and a two-layer metasurface layer structure. The active modulation layer mainly includes patterned micro-nano metal electrodes and active functional materials, and the optical properties of the active functional materials are adjusted by applying an external voltage, so as to achieve a beam output with a specific phase distribution in the terahertz band. The metasurface layer is composed of a silicon pillar structure array located on the surface of the substrate. The two metasurface layers can finely control the complex electric field distribution of a specific diffraction plane to construct an optical diffraction neural network architecture. For each required control mode, the phase distribution of the metasurface layer is optimized using the gradient descent and back propagation algorithms in the neural network, and active control of a specific beam direction can be achieved under different voltage drives. The present invention solves the problems of terahertz phased array technology in terms of control accuracy, continuity, etc., and provides a more efficient and flexible beam control method.

Description

Technical Field

[0001] The present invention relates to a phased array technology in the terahertz band, specifically a D 2 Terahertz phased array devices and methods based on Diffractive Deep Neural Networks (NNs) are used in communications, radar, sensing and other fields, and are suitable for dynamic control and high-precision beam control. Background Art

[0002] D 2 NNs is an all-optical architecture based on deep learning. By designing a passive diffraction layer and training it using an error back-propagation algorithm, it can achieve a variety of complex machine learning tasks at nearly the speed of light.

[0003] Terahertz waves are electromagnetic waves with a frequency in the range of 0.1-10 THz. Terahertz waves have broadband characteristics and strong penetration, and have important application prospects in the fields of radar and communications. However, the key components in the above system - terahertz phased array devices are still lacking. Phased array devices can achieve functions such as directional beam emission and rapid scanning by changing the phase of the electromagnetic wave emitted by each unit. Limited by the immaturity of the material response and regulation theory system, terahertz phased array devices can only be coded and regulated at present, with disadvantages such as low modulation accuracy (angle scanning accuracy>1°) and inability to scan continuously. Summary of the invention

[0004] Aiming at the problems of control accuracy and continuity of terahertz phased array technology, a D-based 2 The terahertz phased array device of NNs and its implementation method introduces the optical diffraction neural network structure into the phased array device, which can not only solve the shortcomings of terahertz phased array technology in control accuracy and continuity, but also provide a more flexible and efficient terahertz beam control method.

[0005] The technical solution of the present invention is:

[0006] A D-based 2 The terahertz phased array device of NNs includes an active modulation layer and two metasurface layers; the two metasurface layers are located behind the active modulation layer, the first metasurface layer is close to the active modulation layer, and the two metasurface layers are separated by a certain distance; the two metasurfaces constitute an optical diffraction neural network architecture, and the phase distribution of the metasurface layer is optimized by gradient descent and back propagation algorithms to realize dynamic control of terahertz waves;

[0007] The active modulation layer includes patterned micro-nano metal electrodes and active functional materials. An external multi-channel voltage source is connected to the patterned micro-nano metal electrodes and voltage is applied to cause the active functional material to undergo phase change. Different voltages are applied to each channel to form a coded voltage pattern. The corresponding phase change pattern of the active functional material forms a code, thereby realizing active modulation of the terahertz wave phase.

[0008] The metasurface layer is composed of a metasurface unit structure, where each metasurface unit is composed of a rectangular silicon column structure located on the surface of the substrate. The rectangular silicon column structure in each unit is rotated at different angles with the center of the unit as the origin. This angle is optimized through optical diffraction neural network calculations.

[0009] Furthermore, the metasurface controls the phase of the terahertz wave by the geometric phase principle. The rotation angle of each rectangular silicon pillar structure has a specific spatial distribution. The optimization process is as follows:

[0010] Determine the light field of the target terahertz output. When a multi-channel voltage source applies different coding voltages to the active modulation layer, consider the two-layer metasurface diffraction, and use the Rayleigh-Sommerfeld diffraction algorithm to construct an optical diffraction neural network. Take multiple samples as a batch and input them into the neural network. Combined with the gradient descent and back propagation algorithms, the phase modulation parameters of the metasurface are obtained, and the corresponding rotation angle distribution is obtained using the geometric phase principle.

[0011] Furthermore, the incident light uses circularly polarized light, and its phase distribution changes after passing through the active modulation layer, and then passes through two metasurface layers, thereby realizing the emission of a specific light beam.

[0012] Preferably, for the metasurface unit, each rectangular silicon column structure has a length of 145 microns, a width of 65 microns, and a height of 450 microns; the silicon substrate thickness is 600 microns; the metasurface unit can be equivalent to a half-wave plate and can operate at a specified frequency.

[0013] Preferably, the metasurface unit structure is designed and optimized, and at 0.6 THz, the metasurface unit structure is a half-wave plate with excellent performance, a transmittance greater than 80%, and a polarization conversion rate greater than >95%; according to the principle of geometric phase, the incident circularly polarized terahertz wave will be converted into another circularly polarized light orthogonal to the incident circularly polarized terahertz wave, and the outgoing light will obtain a rotation angle with the rectangular columnar structure. θ Related Phase 2 θ ; Therefore, the desired phase modulation can be obtained by simply rotating each rectangular columnar structure in the metasurface.

[0014] Preferably, the active functional material in the active modulation layer is a material having different electromagnetic properties to the terahertz wave under different voltages, and may be liquid crystal, VO2, or GaN.

[0015] A D-based 2 The implementation method of terahertz phased array device of NNs is based on D 2 The terahertz phased array device of NNs is realized by the following specific steps:

[0016] Step 1: Construction process of optical diffraction neural network: Assume that the incident circularly polarized terahertz wave passes through the active modulation layer, two metasurface layers and is output to the detection plane. Use the Rayleigh-Sommerfeld diffraction algorithm to simulate the diffraction process of the circularly polarized terahertz wave passing through the active modulation layer, the metasurface layer and the spatial propagation. Model the above process using programming language, and construct the optical diffraction neural network using the deep learning neural network technology framework.

[0017] Step 2, optimization method: divide the active modulation layer into columns, apply different phase conditions to different columns to form different coding modes, form an incident light field with a coding mode, and set and determine the form of the target light field to be emitted; calculate the diffraction process of the incident light passing through the double-layer metasurface layer by using the Rayleigh-Sommerfeld diffraction algorithm, solve the diffraction light field to be emitted, calculate the error between the emitted diffraction light field and the set target light field, use the gradient descent and back propagation algorithms to form the effect of beam deflection in different directions according to the diffraction spot position of the target light field, and obtain the optimized phase distribution of the metasurface in the above process;

[0018] Step 3: Reverse design of the metasurface layer: By applying the geometric phase principle, the metasurface layer selects a rectangular silicon column structure with half-wave plate properties; each rectangular silicon column structure can adjust the rotation angle according to the above optimization results to achieve different beam control requirements;

[0019] Step 4: Apply a coding voltage to the active modulation layer through a multi-channel voltage source to form incident light fields with different phase codes under different coding voltages. After passing through the metasurface layer, the diffracted light beam can be deflected to the set position and focused according to the set diffraction process. By switching the applied voltage coding mode, the focusing position or deflection direction of the light beam can be dynamically controlled, so that the device can achieve dynamic beam deflection and precise adjustment of focusing; the specific implementation method includes: the incident terahertz wave is a circularly polarized Gaussian beam or a plane wave, and a multi-channel power supply is used to apply voltage to the active modulation layer to dynamically encode and control the terahertz wave phase; the modulated terahertz wave will continue to pass through the double-layer metasurface, thereby further diffracting and modulating the encoded terahertz beam to achieve a specific terahertz beam output; wherein, the metasurface structure is optimized by substituting the output target light field into the neural network, and the control effect of the terahertz wave can be accurately controlled.

[0020] Furthermore, by introducing a diffraction neural network to simulate the diffraction process of terahertz waves, and combining the gradient descent and back propagation algorithms, the phase distribution of the metasurface layer is optimized to ensure efficient dynamic control of the light beam under different voltage conditions; the metasurface layer is composed of a rectangular silicon pillar structure array. The design is based on the geometric phase principle, and a rectangular silicon pillar structure with a half-wave plate function is selected. The deflection control of the terahertz wave is achieved according to the optimized metasurface layer phase distribution arrangement; specifically, when a beam of light passes through the metasurface structure, each metasurface unit can independently control the incident light, which can be expressed as:

[0021] (1)

[0022] Represents the complex amplitude of the light field, describing the light wave at position ( x , y , z ) and the electric field intensity and phase at wavelength λ, is the amplitude of the light field, expressed in ( x , y , z ) is the light field intensity at position and wavelength λ, is the phase distribution of the light field, which contains the phase information of the light wave. describes the phase factor of the light field, where i is an imaginary unit;

[0023] According to Huygens' principle, the light emitted from the metasurface can be equivalent to a secondary wavelet source. When the propagation distance is r When , the electric field generated by the secondary sub-wave source at this position is:

[0024] (2)

[0025] is the propagation factor of the secondary wave source, which contains information about distance and wavelength, is the equivalent area of ​​the hypersurface unit, r Describes the spatial distribution of the secondary wave source and the target location, the distance from the emission point to the observation point, is the phase factor; therefore, when the spatial position of the double-layer metasurface plane in the diffraction process and the position distribution of the unit structure on the metasurface are determined, the modulation of the light by the metasurface can be regarded as the weight that needs to be trained. By optimizing the phase or amplitude modulation coefficient through back propagation, a specific light field distribution can be obtained; if the output light fields of different phased array units are controlled into different diffraction patterns, then the total light field distribution on a specific plane will be the superposition of the light fields generated by the output light fields of these units on different planes, that is E total = EA + E B + E C + E D ; At this time, if the phase or amplitude modulation is performed on the phased array unit α , the total outgoing field becomes E total = αE A + E B + E C + E D , and so on:

[0026] (3)

[0027] E total is the sum of the light fields generated by the phased array units in all coding modes on the same plane, E A , E B , E C , E D is the independent output light field of each phased array unit, α is the modulation coefficient of the phased array unit, represents the phase adjustment parameter, and represents the different coding modes formed. E 1, E 2, E 3, E 4 is the total light field intensity under different modulation coefficient conditions, that is, different input coding modes;

[0028] By using metasurfaces to construct special diffraction light fields, the control mode of the phased array can be converted into the eigensolution of the matrix. To use this characteristic for light field control, it is only necessary to design the metasurface to meet the corresponding output light field, thereby greatly improving the performance of phased array devices in modulation accuracy and continuity.

[0029] The beneficial effects of the present invention are:

[0030] The technical means of the present invention can solve the problems of low control accuracy and difficulty in achieving continuous scanning in the current technology, and provide a new solution with high flexibility and efficient dynamic control capabilities for the fields of terahertz communication, imaging, detection, etc.

[0031] The present invention discloses a method based on D 2NNs' terahertz phased array device proposes a new terahertz phased array device, which realizes efficient, accurate and dynamic control of terahertz beams in complex environments. This invention has important application prospects in the future 6G communications, radar, security detection and other fields.

[0032] The present invention discloses a method based on D 2 NNs' terahertz phased array devices are combined with metasurfaces and phased array devices, making use of the metasurface's refined light field control capabilities to achieve higher modulation accuracy (angle scanning accuracy <0.3°) and better functionality (extended to focusing and holographic control).

[0033] The present invention discloses a method based on D 2 NNs' terahertz phased array device uses neural networks to optimize the phase distribution of the metasurface layer and combines it with the geometric phase principle to achieve efficient control of the terahertz beam while ensuring accuracy, including beam deflection and focusing.

[0034] The present invention discloses a method based on D 2 The terahertz phased array device of NNs introduces metasurfaces to construct optical diffraction neural networks, which greatly reduces the demand for active functional materials and processing of phased array devices. Through the diffraction control ability of the metasurface and the optimization ability of the neural network, good modulation effects can be achieved even in active modulation layers with low phase modulation depth and large unit size.

[0035] The invention discloses a method based on D 2 Compared with the traditional static control method, the active functional materials of NNs' terahertz phased array devices can respond quickly under the action of an external electric field, realize real-time beam control, greatly improve the control efficiency of the system, and are suitable for scenarios that require rapid beam adjustment. It can be widely used in terahertz communication, imaging, detection and other fields, and can provide more efficient technical solutions in wireless communication, security inspection, medical imaging and other aspects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is an embodiment of the present invention based on D 2 3D schematic diagram of the terahertz phased array device of NNs;

[0037] Figure 2 This is an embodiment of the present invention based on D 2 Schematic diagram of the terahertz beam control principle of the terahertz phased array device of NNs;

[0038] Figure 3 This is an embodiment of the present invention based on D 2 Schematic diagram of the rectangular silicon pillar structure and frequency response diagram of the geometric phase design of the terahertz phased array device of NNs;

[0039] Figure 4 This is an embodiment of the present invention based on D 2 The spatial distribution diagram of the rotation angle of the rectangular silicon pillar structure in the metasurface optimized for terahertz focus spot control and far-field control by NNs’ terahertz phased array device;

[0040] Figure 5 This is an embodiment of the present invention based on D 2 The simulation results of NNs' terahertz phased array device for terahertz focus spot control and far-field control;

[0041] Figure 6 This is an embodiment of the present invention based on D 2 Far-field angle scanning diagram of the terahertz phased array device of NNs.

[0042] Figure ID:

[0043] 1. Active modulation layer; 2. Optical diffraction neural network; 3. Metasurface; 4. Multi-channel voltage source; 5. Metasurface unit; 6. Rectangular silicon pillar structure; 7. Dielectric substrate. DETAILED DESCRIPTION

[0044] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] Metasurfaces can perform fine diffraction modulation on electromagnetic waves. The rapid development of this technology has provided new possibilities for the optimization and control of terahertz phased arrays. In particular, the phase control of traditional phased arrays relies on complex algorithm design. By using a double-layer metasurface to construct a diffraction neural network and then using the gradient descent method to optimize and solve each output mode, the demand for active modulation devices in phased array devices can be reduced, and high-precision, continuous and fast scanning effects can be achieved under simple phase modulation.

[0046] The terahertz phased array device of the present invention includes an active modulation layer and two super surface layers. The active modulation layer includes patterned micro-nano metal electrodes and active functional materials. An external multi-channel voltage source is connected to the patterned micro-nano metal electrodes and voltage is applied to cause the active functional material to undergo phase change. Different voltages are applied to each channel to form a coded voltage mode. The corresponding phase change mode of the active functional material forms a code, thereby realizing active modulation of the terahertz wave phase. The super surface layer is composed of a super surface unit structure, wherein each super surface unit is composed of a rectangular silicon column structure located on the surface of the substrate, and the rectangular silicon column structure in each unit is rotated at different angles with the unit center as the origin, and this angle is calculated and optimized by an optical diffraction neural network.

[0047] Device design mainly includes the following processes:

[0048] Step 1: Construction process of optical diffraction neural network: Assume that the incident circularly polarized terahertz wave passes through the active modulation layer, two metasurface layers and is output to the detection plane. Use the Rayleigh-Sommerfeld diffraction algorithm to simulate the diffraction process of the circularly polarized terahertz wave passing through the active modulation layer, metasurface layer and spatial propagation. Model the above process using programming language, and construct the optical diffraction neural network using the deep learning neural network technology framework.

[0049] Step 2, optimization method: divide the active modulation layer into columns, apply different phase conditions to different columns to form different coding modes, form an incident light field with a coding mode, and set the output target light field form. The Rayleigh-Sommerfeld diffraction algorithm is used to calculate the diffraction process of the incident light passing through the double-layer metasurface layer, solve the output diffraction light field, calculate the error between the output diffraction light field and the set target light field, and use the gradient descent and back propagation algorithms to form the effect of beam deflection in different directions according to the diffraction spot position of the target light field. In the above process, the optimized phase distribution of the metasurface is obtained.

[0050] Step 3: Reverse design of the metasurface layer: By applying the geometric phase principle, the metasurface layer selects a rectangular silicon pillar structure with half-wave plate properties. Each rectangular silicon pillar structure can adjust the rotation angle according to the above optimization results to achieve different beam control requirements.

[0051] Step 4: Apply coding voltage to the active modulation layer through a multi-channel voltage source, and form different phase-coded incident light fields under different coding voltages. After passing through the metasurface layer, the diffracted light beam can be deflected to the set position and focused according to the set diffraction process. By switching the applied voltage coding mode, the focusing position or deflection direction of the light beam can be dynamically controlled, so that the device can achieve dynamic beam deflection and precise adjustment of focusing. The specific implementation method includes: the incident terahertz wave is a circularly polarized Gaussian beam or a plane wave, and a multi-channel power supply is used to apply voltage to the active modulation layer to dynamically encode and control the phase of the terahertz wave; the modulated terahertz wave will continue to pass through the double-layer metasurface, thereby further diffracting and modulating the coded terahertz beam to achieve a specific terahertz beam emission. Among them, the metasurface structure is optimized by substituting the output target light field into the neural network, which can accurately control the control effect of the terahertz wave, so that the beam modulation accuracy can be <0.3°, and continuous terahertz beam active scanning can be achieved, and multiple functions (such as focus point control, etc.) can also be achieved;

[0052] The present invention proposes a method based on D 2NNs' terahertz phased array device includes an active modulation layer and two metasurface layers. The active modulation layer is composed of patterned metal micro-nano electrodes and active functional materials (such as liquid crystal, VO2, GaN, etc.). The electromagnetic properties of the active functional materials in the terahertz band can be changed by applying an external voltage, thereby performing phase control on the incident terahertz wave.

[0053] In this embodiment, the diffraction neural network is introduced to simulate the diffraction process of the terahertz wave, and the phase distribution of the metasurface layer is optimized by combining the gradient descent and back propagation algorithms to ensure efficient dynamic control of the light beam under different voltage conditions; the metasurface layer is composed of a rectangular silicon pillar structure array, and the design is based on the geometric phase principle. A rectangular silicon pillar structure with a half-wave plate function is selected, and the deflection control of the terahertz wave is realized according to the optimized metasurface layer phase distribution arrangement; specifically, when a beam of light passes through the metasurface structure, each metasurface unit can independently control the incident light, which can be expressed as:

[0054] (1)

[0055] Represents the complex amplitude of the light field, describing the light wave at position ( x , y , z ) and the electric field intensity and phase at wavelength λ, is the amplitude of the light field, expressed in ( x , y , z ) is the light field intensity at position and wavelength λ, is the phase distribution of the light field, which contains the phase information of the light wave. describes the phase factor of the light field, where i is an imaginary unit;

[0056] According to Huygens' principle, the light emitted from the metasurface can be equivalent to a secondary wavelet source. When the propagation distance is r When , the electric field generated by the secondary sub-wave source at this position is:

[0057] (2)

[0058] is the propagation factor of the secondary wave source, which contains information about distance and wavelength, is the equivalent area of ​​the hypersurface unit, r Describes the spatial distribution of the secondary wave source and the target location, the distance from the emission point to the observation point, is the phase factor; therefore, when the spatial position of the double-layer metasurface plane in the diffraction process and the position distribution of the unit structure on the metasurface are determined, the modulation of the light by the metasurface can be regarded as the weight that needs to be trained. By optimizing the phase or amplitude modulation coefficient through back propagation, a specific light field distribution can be obtained; if the output light fields of different phased array units are controlled into different diffraction patterns, then the total light field distribution on a specific plane will be the superposition of the light fields generated by the output light fields of these units on different planes, that is E total = E A + E B + E C + E D ; At this time, if the phase or amplitude modulation is performed on the phased array unit α , the total outgoing field becomes E total = αE A + E B + E C + E D , and so on:

[0059] (3)

[0060] E total is the sum of the light fields generated by the phased array units in all coding modes on the same plane, E A , E B , E C , E D is the independent output light field of each phased array unit, α is the modulation coefficient of the phased array unit, represents the phase adjustment parameter, and represents the different coding modes formed. E 1, E 2, E 3, E 4 is the total light field intensity under different modulation coefficient conditions, that is, under different input coding modes. The beam control principle is as shown in the attached figure. Figure 2 As shown;

[0061] By using a metasurface to construct a special diffraction light field, the control mode of the phased array can be turned into an eigensolution of the matrix. Using this characteristic to control the light field only requires designing a metasurface to satisfy the corresponding output light field, thereby greatly improving the performance of the phased array device in terms of modulation accuracy and continuity. When the required control mode is large enough, it becomes relatively difficult to solve the above eigenvalue problem using analytical methods. The present invention will introduce a deep learning optimization algorithm to optimize the structure, which can better optimize and reverse design.

[0062] In this implementation case, the super surface unit structure is as shown in the attached Figure 3 As shown, it consists of a dielectric substrate and a rectangular columnar structure, and its structural parameters are: width W = 65 μm, length L = 145 μm, height H = 450 μm. After design and optimization, the metasurface unit structure is a half-wave plate with excellent performance at 0.6 THz, with a transmittance greater than 80% and a polarization conversion efficiency (the efficiency of converting circularly polarized waves into orthogonal circularly polarized waves) greater than 95%. Figure 3 According to the geometric phase principle, the incident circularly polarized terahertz wave will be converted into another circularly polarized light orthogonal to the incident circularly polarized terahertz wave (such as left-handed circularly polarized light converted into right-handed circularly polarized light, and vice versa), and the outgoing light will obtain a rotation angle with the rectangular columnar structure. θ Related Phase 2 θ . Therefore, the desired phase modulation can be obtained by simply rotating each rectangular columnar structure in the metasurface. Note: Terahertz waves refer to light waves with a frequency in the terahertz band, while circularly polarized light is a special form of polarized light in which the electric field vector of the light wave rotates along a spiral trajectory.

[0063] The working mechanism and principle of the present invention are further described below in combination with specific functions. For example, the target plane is set to be 15000 μm away from the active modulation layer, the interval between the two metasurfaces is 7500 μm, and the length and width of the active modulation layer and the metasurface layer are both 12800 μm. At this time, the active modulation layer is set to produce a coded modulation effect on the terahertz wave, and the spatial phase distribution of the emitted terahertz wave is as shown in the attached figure. Figure 5 As shown. The target emitted terahertz light field is set as the focused spot. According to formula (3), the terahertz propagation process is modeled using the Python programming language and brought into the neural network optimization to obtain the spatial rotation angle distribution required for the rectangular silicon pillar structure in the two-layer metasurface, as shown in the attached figure. Figure 4 The simulation results of the terahertz light spot emitted by the overall structure under different coded phase modulation are shown in the attached figure. Figure 5As shown, the light beam can be well focused and the focus can be dynamically adjusted; similarly, this method can be used to design far-field control. At this time, the light emitted from the second layer propagates infinitely far, and the emitted light field can be projected into the wave vector space using Fourier transform, so that the final emitted terahertz waves under different active modulation states are at different positions in the far field.

[0064] Figure 6 This is a far-field angle scan diagram. Under the control of 19 continuous coding modes, the scanning angle deflects from -22.5° to -30°, the minimum deflection angle is less than 0.3°, and the average angle is less than 0.5°, achieving continuous and high-precision rapid beam control.

[0065] In this embodiment, the neural network control module can optimize and calculate the phase distribution of the metasurface layer according to external inputs and settings, such as the position of the diffraction spot focus on the target plane, the beam deflection direction, and the voltage distribution. Through repeated training and adjustment, the system can achieve high-precision control of the terahertz beam in complex environments, with fast response speed and high control efficiency, and is suitable for various real-time application scenarios.

[0066] The present invention proposes a method based on D 2 The terahertz phased array device of NNs and its implementation method utilize the voltage coding characteristics of the active control layer combined with the phase control of the metasurface layer to achieve dynamic focusing and control of the terahertz beam. By introducing the neural network optimization algorithm, the efficiency and accuracy of beam control are greatly improved, which can be widely used in wireless communications, terahertz imaging and detection, and other fields, providing a new solution for the application and development of terahertz technology.

[0067] The above-mentioned embodiment only expresses one implementation mode of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A D-based 2 A terahertz phased array device of NNs, characterized in that The terahertz phased array device includes an active modulation layer and two metasurface layers; the two metasurface layers are located behind the active modulation layer, the first metasurface layer is close to the active modulation layer, and the two metasurface layers are separated by a certain distance; the two metasurfaces constitute an optical diffraction neural network architecture, and the phase distribution of the metasurface layer is optimized by gradient descent and back propagation algorithms to achieve dynamic control of terahertz waves; The active modulation layer includes patterned micro-nano metal electrodes and active functional materials. An external multi-channel voltage source is connected to the patterned micro-nano metal electrodes and voltage is applied to cause the active functional material to undergo a phase change. Different voltages are applied to each channel to form a coded voltage pattern, and the corresponding phase change pattern of the active functional material forms a code, thereby realizing active modulation of the terahertz wave phase; the metasurface layer is composed of a metasurface unit structure, wherein each metasurface unit is composed of a rectangular silicon column structure located on the surface of the substrate, and the rectangular silicon column structure in each unit is rotated at different angles with the center of the unit as the origin. This angle is obtained through optical diffraction neural network calculation optimization.

2. The method according to claim 1 based on D 2 A terahertz phased array device of NNs, characterized in that The metasurface uses the geometric phase principle to control the phase of the terahertz wave. The rotation angle of each rectangular silicon column structure has a specific spatial distribution. The optimization process is as follows: Determine the light field of the target terahertz output. When a multi-channel voltage source applies different coding voltages to the active modulation layer, consider the two-layer metasurface diffraction, and use the Rayleigh-Sommerfeld diffraction algorithm to construct an optical diffraction neural network. Take multiple samples as a batch and input them into the neural network. Combined with the gradient descent and back propagation algorithms, the phase modulation parameters of the metasurface are obtained, and the corresponding rotation angle distribution is obtained using the geometric phase principle.

3. The method according to claim 1 based on D 2 A terahertz phased array device of NNs, characterized in that The incident light uses circularly polarized light. After passing through the active modulation layer, the phase distribution changes, and then passes through two layers of metasurface layers, thereby realizing the emission of a specific light beam.

4. The method according to claim 1 2 A terahertz phased array device of NNs, characterized in that For the metasurface unit, each rectangular silicon pillar structure is 145 microns long, 65 microns wide, and 450 microns high; the silicon substrate is 600 microns thick; the metasurface unit can be equivalent to a half-wave plate and can operate at a specified frequency.

5. The method according to claim 4 based on D 2 A terahertz phased array device of NNs, characterized in that The metasurface unit structure has been designed and optimized. At 0.6 THz, the metasurface unit structure is a half-wave plate with excellent performance, with a transmittance greater than 80% and a polarization conversion rate greater than 95%. According to the principle of geometric phase, the incident circularly polarized terahertz wave will be converted into another circularly polarized light orthogonal to the incident circularly polarized terahertz wave, and the outgoing light will obtain a rotation angle with the rectangular silicon column structure. θ Related Phase 2 θ ; Therefore, the desired phase modulation can be achieved by simply rotating each rectangular silicon pillar structure in the metasurface.

6. The method according to claim 1 based on D 2 A terahertz phased array device of NNs, characterized in that The active functional material in the active modulation layer is a material having different electromagnetic properties to the terahertz wave under different voltages.

7. A D-based 2 A method for realizing a terahertz phased array device of NNs, characterized in that: Based on any one of claims 1-6 based on D 2 The terahertz phased array device of NNs is realized by the following specific steps: Step 1: Construction process of optical diffraction neural network: Assume that the incident circularly polarized terahertz wave passes through the active modulation layer, two metasurface layers and is output to the detection plane. Use the Rayleigh-Sommerfeld diffraction algorithm to simulate the diffraction process of the circularly polarized terahertz wave passing through the active modulation layer, the metasurface layer and the spatial propagation. Model the above process using programming language, and construct the optical diffraction neural network using the deep learning neural network technology framework. Step 2, optimization method: divide the active modulation layer into columns, apply different phase conditions to different columns to form different coding modes, form an incident light field with a coding mode, and set and determine the form of the target light field to be emitted; calculate the diffraction process of the incident light passing through the double-layer metasurface layer by using the Rayleigh-Sommerfeld diffraction algorithm, solve the diffraction light field to be emitted, calculate the error between the emitted diffraction light field and the set target light field, use the gradient descent and back propagation algorithms to form the effect of beam deflection in different directions according to the diffraction spot position of the target light field, and obtain the optimized phase distribution of the metasurface in the above process; Step 3: Reverse design of the metasurface layer: By applying the geometric phase principle, the metasurface layer selects a rectangular silicon pillar structure with half-wave plate properties; each rectangular silicon pillar structure can adjust the rotation angle according to the phase distribution result of the optimized metasurface obtained in step 2 to achieve different beam control requirements; Step 4: Apply coding voltage to the active modulation layer through a multi-channel voltage source to form incident light fields with different phase coding under different coding voltages. After passing through the metasurface layer, the diffracted light beam can be deflected to the set position and focused according to the set diffraction process.

8. The method according to claim 7 based on D 2 A method for realizing a terahertz phased array device of NNs, characterized in that: The metasurface layer is composed of a rectangular silicon pillar structure array. The design is based on the principle of geometric phase. A rectangular silicon pillar structure with a half-wave plate function is selected, and the deflection control of the terahertz wave is realized according to the optimized metasurface layer phase distribution arrangement. When a beam of light passes through the metasurface structure, each metasurface unit can independently control the incident light, which can be expressed as: Represents the complex amplitude of the light field, describing the light wave at position ( x , y , z ) and the electric field intensity and phase at wavelength λ, is the amplitude of the light field, expressed in ( x , y , z ) is the light field intensity at position and wavelength λ, is the phase distribution of the light field, which contains the phase information of the light wave. describes the phase factor of the light field, where i is an imaginary unit; According to Huygens' principle, the light emitted from the metasurface can be equivalent to a secondary wavelet source. When the propagation distance is r When , the electric field generated by the secondary sub-wave source at this position is: is the propagation factor of the secondary wave source, which contains information about distance and wavelength, is the equivalent area of ​​the hypersurface unit, r Describes the spatial distribution of the secondary wave source and the target location, the distance from the emission point to the observation point, is the phase factor; therefore, when the spatial position of the double-layer metasurface plane in the diffraction process and the position distribution of the unit structure on the metasurface are determined, the modulation of the light by the metasurface can be regarded as the weight that needs to be trained. By optimizing the phase or amplitude modulation coefficient through back propagation, a specific light field distribution can be obtained; if the output light fields of different phased array units are controlled into different diffraction patterns, then the total light field distribution on a specific plane will be the superposition of the light fields generated by the output light fields of these units on different planes, that is E total = E A + E B + E C + E D ; At this time, if the phase or amplitude modulation is performed on the phased array unit α , the total outgoing field becomes E total = αE A + E B + E C + E D , and so on: E total is the sum of the light fields generated by the phased array units in all coding modes on the same plane, E A , E B , E C , E D is the independent output light field of each phased array unit, α is the modulation coefficient of the phased array unit, represents the phase adjustment parameter, and represents the different coding modes formed. E 1 , E 2 , E 3 , E 4 is the total light field intensity under different modulation coefficient conditions, that is, different input coding modes; By using metasurfaces to construct special diffraction light fields, the control mode of the phased array can be converted into the eigensolution of the matrix. Using this characteristic to control the light field only requires designing the metasurface to satisfy the corresponding output light field.

Images