Optical diffraction neural network device with nonlinear structure and regulation and control method of optical diffraction neural network device

Abstract

The invention discloses a regulation and control method for an optical diffraction neural network device with a nonlinear structure, and the method comprises the steps: constructing a mapping relation between an input polarization angle and an output phase difference based on a Jones matrix theory, taking a target nonlinear function as a constraint, and carrying out the iterative optimization of a target phase pattern of a phase plate through an error back propagation algorithm; preparing a reflective quarter-wave plate phase plate with a nonlinear phase modulation function, and integrating the reflective quarter-wave plate phase plate to a nonlinear optical path system; binarized nonlinear characteristics are verified by measuring emergent light polarization and phase difference, and the thickness of the phase pattern or the phase plate is adjusted according to a measurement result until the nonlinear activation requirement of the neural network is met, so that efficient optimization of the phase pattern is realized, the precision and stability of nonlinear activation are remarkably improved, and the application prospect is wide. And a reliable technical support is provided for the optical neural network to process a complex nonlinear problem.

Description

Technical Field

[0001] The present invention relates to the technical field of optical devices, and in particular to a light diffraction neural network device with a nonlinear structure and a control method thereof. Background Art

[0002] Optical Neural Networks (ONNs), a novel architecture based on optical computing, have shown great potential in image recognition, real-time signal processing, and quantum communications due to their ultra-high speed, low energy consumption, and high parallelism. Among them, diffractive deep neural networks (D2NNs) achieve passive computing through the diffraction and interference of light, capable of completing complex tasks directly during the propagation of light, significantly improving computational efficiency. However, existing optical neural networks primarily focus on linear computation scenarios, limiting their performance in tasks requiring nonlinear mapping, such as image classification and pattern recognition. There is an urgent need to develop efficient nonlinear activation devices to overcome these application bottlenecks.

[0003] Existing optical diffraction neural networks often use linear phase modulation devices (such as liquid crystal spatial light modulators) or static holographic gratings to achieve light field control and simulate nonlinear effects through cascaded linear layers. However, such methods rely on complex optical path designs or multi-layer stacking structures, resulting in bulky systems and significant energy loss. More importantly, their nonlinear activation function is essentially an approximate superposition of linear operations, making it difficult to achieve nonlinear mapping at the physical level, resulting in insufficient adaptability to complex tasks such as dynamic scene recognition and high-dimensional data classification. This limitation directly limits the practical application of optical neural networks in scenarios such as edge computing and real-time decision-making.

[0004] In view of this, it is necessary to improve the existing optical diffraction neural network device technology to solve the technical problem of limited complex task processing capability due to insufficient nonlinear activation function. Summary of the Invention

[0005] The purpose of the present invention is to provide a nonlinear optical diffraction neural network device and a control method thereof to solve the above technical problems.

[0006] To achieve this object, the present invention adopts the following technical solutions: A method for controlling a nonlinear optical diffraction neural network device comprises the following steps: S1, based on the Jones matrix theory, constructs the mapping relationship between the input polarization angle and the output phase difference. With the target nonlinear function as the constraint, the target phase pattern of the phase plate is iteratively optimized through the error back propagation algorithm; S2, prepare a reflective quarter-wave phase plate with nonlinear phase modulation function and integrate it into a nonlinear optical system; S3 verifies the binary nonlinear characteristics by measuring the polarization and phase difference of the outgoing light, and adjusts the phase pattern or phase plate thickness according to the measurement results until the nonlinear activation requirements of the neural network are met.

[0007] Optionally, step S1 specifically includes the following steps: S11, establishing a Jones matrix model for a reflective quarter-wave plate: based on the fast axis direction of the phase plate, the properties of the reflector, and the liquid crystal molecular arrangement parameters, derive its Jones matrix expression and verify its modulation characteristics for linearly polarized light; S12, substituting the incident light polarization angle θ into the Jones matrix model, calculating the polarization direction and phase difference Δφ of the output light, generating a mathematical correlation formula of θ-Δφ, and defining the mapping relationship between the input polarization angle and the output phase difference; S13, performing discretization calculation on the correlation equation using a numerical simulation tool, plotting Δφ distribution curves under different θ values, and verifying the nonlinear characteristics of the binary phase jump; S14, defining a target phase difference function according to the activation requirements of the optical diffraction neural network, and establishing an error evaluation criterion between the target phase difference function and the measured phase difference to set the target nonlinear function constraint; S15, setting the liquid crystal azimuth angle φ of the phase plate as a trainable variable, establishing a neural network model based on the light field input-output mapping relationship, defining a loss function L, and constructing an error backpropagation framework; S16, adjusting the φ value through a gradient descent algorithm to minimize the loss function, and generating a phase distribution pattern that meets the target nonlinear response after a preset number of iterations to iteratively optimize the phase pattern.

[0008] Optionally, the reflective quarter-wave phase plate is prepared as follows: S11, substrate pretreatment: ultrasonically clean the reflective substrate in deionized water for 2 hours, blow dry with nitrogen, and then dry it. Then, activate the reflective substrate using plasma surface treatment technology to form a micro-nano rough surface with high adhesion; S12, dissolving the photosensitive alignment material in an organic solvent, forming a uniform thin film on the surface of the activated reflective substrate by a gradient spin coating process, pre-spreading at a low speed, switching to a high speed to complete the uniformization of the thin film, and then pre-baking and curing at a low temperature in a vacuum environment to prepare a photo-controlled alignment layer; S13, using a dynamic mask exposure technique to pattern the microstructure of the photo-controlled alignment layer, wherein the photosensitive material undergoes directional photocrosslinking, isomerization, or photolysis reaction under ultraviolet polarized light irradiation, and the exposure pattern is a preset neural network phase distribution pattern; S14, evenly spin-coating a UV-curable nematic liquid crystal onto the upper surface of the light-controlled alignment layer. Due to intermolecular interactions, the photosensitive material in the alignment layer induces the liquid crystal molecules to align along their arrangement to form a liquid crystal layer. S15, Phase plate packaging and calibration: Cover with a transparent protective layer and seal the edges, use a laser interferometer to detect the birefringence characteristics of the phase plate, adjust the packaging pressure so that the phase difference meets the quarter-wave condition, and complete the phase plate packaging.

[0009] Optionally, the optical path system specifically includes: A laser for outputting linearly polarized light with a wavelength of 633 nm; a beam expander, configured to expand the linearly polarized light into a parallel beam; Linear polarizer, used to adjust the polarization direction of light incident on the reflective quarter-wave phase plate; The reflective quarter-wave phase plate is used to control the polarization and phase of the incident light field based on nonlinear phase modulation; Beam splitter prism, used to separate the reflected light path to the CCD detection end; and CCD, which is used to capture and analyze the amplitude and phase distribution of the output light field in real time.

[0010] Optionally, the process of integrating the reflective quarter-wave phase plate into the optical path system is specifically as follows: The calibrated reflective quarter-wave phase plate is fixed to the optical path bracket and aligned with the laser, the beam expander and the linear polarizer. The phase plate tilt angle is adjusted in real time through the feedback of the beam splitter prism to ensure that the incident light is vertical and the reflected light path coincides with the CCD receiving end.

[0011] Optionally, the periodic microstructure of the mask pattern is specifically a rectangular waveguide array designed based on target phase jumps of 0° and 180°, and its period and width match a quarter-wave condition according to the operating wavelength.

[0012] Optionally, step S3 specifically includes the following steps: S31, selecting an input polarization angle range, configuring the laser to output linearly polarized light of the corresponding angle, and adjusting the light field parameters incident on the phase plate through a beam expander and a linear polarizer; S32, using the CCD to capture the amplitude distribution of the reflected light, and combining it with the polarization analysis module to measure the polarization direction and phase difference of the outgoing light to generate a measured data set of θ-Δφ; S33, comparing the measured data set with the target phase difference function, calculating the phase jump error (δ=|Δφ_{target}-Δφ_{measured}|), and identifying the angle interval that does not meet the binarization characteristic; S34, if the error δ exceeds a first threshold, re-optimizing the phase pattern or adjusting the thickness of the phase plate based on the error distribution, and updating the surface structure of the phase plate by photolithography or reactive ion etching; S35, reintegrate the corrected phase plate into the optical path system, and repeat steps S41-S43 until the phase jump error δ of all test angles is ≤ 5°, meeting the nonlinear activation requirements of the neural network.

[0013] The present invention also provides a light diffraction neural network device with a nonlinear structure, which is manufactured using the control method of the light diffraction neural network device with a nonlinear structure as described above. The light diffraction neural network device includes an integrated bracket and an optical path system integrated on the integrated bracket.

[0014] Compared with the existing technology, the present invention has the following beneficial effects: first, based on the Jones matrix theory, a mapping relationship between the input polarization angle and the output phase difference is established, and with the target nonlinear function as a constraint, the target phase pattern is generated by iterative optimization through the error back propagation algorithm; a reflective quarter-wave plate phase plate with nonlinear phase modulation function is prepared and integrated into the optical path system; the binary nonlinear characteristics are verified by measuring the polarization and phase difference of the output light, and the phase pattern or phase plate thickness is dynamically adjusted according to the measured results until the performance requirements of the nonlinear activation of the neural network are met; this scheme combines Jones matrix modeling with the error back propagation algorithm to achieve efficient optimization of the phase pattern and significantly improve the accuracy and stability of the nonlinear activation; the dynamic verification and adjustment mechanism can quickly adapt to the needs of different neural network tasks, enhance the versatility and practicality of the device, and provide reliable technical support for optical neural networks to deal with complex nonlinear problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] The structures, proportions, sizes, etc. depicted in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not intended to limit the conditions under which the present invention can be implemented and therefore have no substantive technical significance. Any structural modifications, changes in proportional relationships, or adjustments in size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and objectives that can be achieved by the present invention.

[0017] Figure 1This is a flow chart of a method for controlling a light diffraction neural network device with a nonlinear structure according to the first embodiment; Figure 2 This is a second flow chart of the control method of the nonlinear structure light diffraction neural network device of the first embodiment; Figure 3 Schematic diagram of the structure of the reflective quarter-wave phase plate of the first embodiment; Figure 4 Schematic diagram of the structure layout of the optical path system of the first embodiment; Figure 5 is a line diagram showing the relationship between the incident polarization angle and the outgoing polarization angle of the first embodiment; Figure 6 1 is a diagram showing the corresponding relationship between the incident polarization angle and the output phase difference of the first embodiment. DETAILED DESCRIPTION

[0018] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0019] In the description of the present invention, it should be understood that the terms "upper," "lower," "top," "bottom," "inner," "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of the present invention and simplify the description. They are not intended to indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. It should be noted that when a component is considered to be "connected" to another component, it may be directly connected to the other component or there may be a centrally located component.

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and through specific implementation methods.

[0021] Example 1: Combine Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a method for controlling a light diffraction neural network device with a nonlinear structure, comprising: S1, constructing a mapping relationship between input polarization angle and output phase difference based on Jones matrix theory, and iteratively optimizing the target phase pattern of phase plate 405 through an error back propagation algorithm with a target nonlinear function as a constraint; Based on Jones matrix theory, a mathematical relationship between the input polarization angle and the output phase difference is established, and the phase pattern is optimized using an error back-propagation algorithm. By deriving a Jones matrix model for phase plate 405, the phase difference response corresponding to different incident polarization angles is quantified, and an error evaluation criterion is defined in conjunction with a target nonlinear function (such as a step function). A gradient descent algorithm is used to iteratively adjust the liquid crystal azimuth parameters, gradually approaching the target phase distribution. This process combines physical models with deep learning algorithms, improving the accuracy and adaptability of nonlinear activation and providing high-fidelity phase pattern data for photolithography.

[0022] S2, preparing a reflective quarter-wave phase plate 405 with nonlinear phase modulation function and integrating it into a nonlinear optical system; A reflective quarter-wave phase plate 405 with nonlinear phase modulation capability is fabricated through precision processes and integrated into the optical path system. This involves cleaning and surface activation of the substrate, coating and patterning of the photo-controlled alignment layer 102, and directional filling of the liquid crystal material. Plasma surface treatment enhances substrate adhesion, and nanoimprinting technology is used to form a periodic microstructure, ensuring that the liquid crystal molecules are aligned in a predetermined direction, thereby endowing the phase plate 405 with nonlinear modulation capabilities. Finally, the phase plate 405 is aligned and integrated with components such as the laser 401, beam expander 402, and linear polarizer 403 to construct the initial optical path system, laying the hardware foundation for subsequent optimization and verification.

[0023] S3, verifying the binary nonlinear characteristics by measuring the polarization and phase difference of the output light, and adjusting the phase pattern or the thickness of the phase plate 405 according to the measurement results until the nonlinear activation requirements of the neural network are met.

[0024] The nonlinear performance of the device is verified and optimized through systematic testing and dynamic adjustment. The phase difference of the outgoing light is measured at different input polarization angles and compared with the target response curve to identify the angular range where the error exceeds the limit. If the error exceeds the threshold, the phase pattern is re-optimized or the thickness of the phase plate 405 is adjusted, and the surface structure is updated through the photolithography process. The closed-loop verification mechanism combines measured data with parameter corrections to gradually converge to the performance standards that meet the requirements of nonlinear activation. This method solves the problem of insufficient nonlinear response in traditional optical diffraction neural networks through iterative optimization, significantly improving the applicability of the device for complex tasks.

[0025] The working principle of the present invention is as follows: first, based on the Jones matrix theory, a mapping relationship between the input polarization angle and the output phase difference is established, and with the target nonlinear function as a constraint, the target phase pattern is generated through iterative optimization of the error back propagation algorithm; a reflective quarter-wave plate phase plate with nonlinear phase modulation function is prepared and integrated into the optical path system; the binary nonlinear characteristics are verified by measuring the polarization and phase difference of the output light, and the phase pattern or phase plate thickness is dynamically adjusted according to the measured results until the performance requirements of the nonlinear activation of the neural network are met; this scheme combines Jones matrix modeling with the error back propagation algorithm to achieve efficient optimization of the phase pattern and significantly improve the accuracy and stability of the nonlinear activation; the dynamic verification and adjustment mechanism can quickly adapt to the needs of different neural network tasks, enhance the versatility and practicality of the device, and provide reliable technical support for optical neural networks to deal with complex nonlinear problems.

[0026] In this embodiment, it is specifically explained that step S2 specifically includes the following steps: S11, establish a Jones matrix model for a reflective quarter-wave plate: based on the fast axis direction of the phase plate 405, the properties of the reflector, and the liquid crystal molecular arrangement parameters, derive its Jones matrix expression and verify its modulation characteristics for linearly polarized light; Based on the fast axis direction of the phase plate 405, the reflective characteristics of the reflector, and the arrangement parameters of the liquid crystal molecules, a Jones matrix model of the reflective quarter-wave plate is established. Its modulation capability of linearly polarized light is quantified through mathematical derivation, providing a theoretical basis for subsequent nonlinear mapping.

[0027] S12, substituting the incident light polarization angle θ into the Jones matrix model, calculating the polarization direction and phase difference Δφ of the output light, generating a mathematical correlation formula of θ-Δφ, and defining the mapping relationship between the input polarization angle and the output phase difference; Substitute the incident polarization angle θ into the Jones matrix model, calculate the polarization direction and phase difference Δφ of the output light, generate the mathematical correlation formula of θ-Δφ, clarify the nonlinear mapping relationship between input and output, and lay the data foundation for simulation and optimization.

[0028] Specifically, the amplitude of the input light field can be controlled by controlling the azimuth angle φ in the liquid crystal plane. Using the Jones matrix, when linearly polarized light transmitted along the z-axis and polarized in the x-direction passes through the liquid crystal molecules, the output light field can be obtained as follows: The reflective half-wave plate can be regarded as incident light passing through a quarter-wave plate, then reflected by a reflector, and then passing through the quarter-wave plate again. The wave plate is operated on. Assuming that the fast axis of the quarter-wave plate is along the x-axis, its Jones matrix is: The Jones matrix of the reflector is: Then the Jones matrix of the reflective quarter-wave plate is: For linearly polarized incident light that forms an angle θ with the x-axis, the Jones vector of the light exiting the reflective quarter-wave plate is: The phase difference between the outgoing light and the incident light is: Through the above formulas, the corresponding relationships between the incident polarization angle and the outgoing polarization angle, as well as the incident polarization angle and the outgoing phase difference can be plotted respectively.

[0029] S13, discretize the correlation equation using numerical simulation tools, plot the Δφ distribution curve under different θ values, and verify the nonlinear characteristics of the binary phase jump; The distribution of Δφ at different θ values was discretized and calculated using numerical simulation tools to verify the binary jump characteristics of the phase difference between 0° and 180°, confirming the physical feasibility of the device's nonlinear response.

[0030] S14, defining a target phase difference function (such as a step function) according to the activation requirements of the optical diffraction neural network, and establishing an error evaluation criterion between the target phase difference function and the measured phase difference to set the target nonlinear function constraint; According to the nonlinear activation requirements of the optical neural network, the target phase difference function (such as a step function) is defined, and an error evaluation criterion between it and the measured value is established to provide constraints for parameter optimization.

[0031] S15, setting the liquid crystal azimuth angle φ of the phase plate 405 as a trainable variable, establishing a neural network model based on the light field input-output mapping relationship, defining a loss function L, and constructing an error back-propagation framework; In step S25 , the liquid crystal azimuth angle φ of the phase plate 405 is set as a trainable variable, a neural network model is constructed based on the input-output relationship of the light field, and a loss function (L=Σ(Δφ_{target}-Δφ_{actual})²) is defined to form an error backpropagation framework.

[0032] S16, adjusting the φ value through a gradient descent algorithm to minimize the loss function, and generating a phase distribution pattern that meets the target nonlinear response after a preset number of iterations to iteratively optimize the phase pattern.

[0033] Combine Figure 5 and Figure 6 As shown, Figure 5 is the relationship between the incident polarization angle and the outgoing polarization angle, Figure 6It is the correspondence between the incident polarization angle and the outgoing phase difference, which overall reflects the polarization relationship between the incident light and the outgoing light of the reflective quarter-wave plate.

[0034] Depend on Figure 6 It can be seen that for incident linearly polarized light with different polarization angles, the reflective quarter-wave plate has a binary nonlinear effect at 0° and 180°. By utilizing this nonlinear effect, the accuracy of the neural network algorithm can be improved.

[0035] The gradient descent algorithm is used to adjust the φ value, and the loss function is minimized through iteration (≥1000 times) to generate a phase distribution pattern that matches the target nonlinear response, completing the algorithm-driven high-precision phase optimization.

[0036] In this embodiment, the manufacturing process of the reflective quarter-wave phase plate 405 is specifically described as follows: S21, substrate pretreatment: ultrasonically cleaning the reflective substrate 101 in deionized water for 2 hours, drying with nitrogen, and then drying the reflective substrate 101 using a plasma surface treatment technique to form a micro-nano rough surface with high adhesion; Ultrasonic cleaning is used to remove contaminants from the substrate surface, and nitrogen drying is used to avoid secondary contamination. After drying, plasma treatment is used to activate the substrate surface to form a micro-nano rough structure, which significantly improves the adhesion of the subsequent light-controlled alignment layer 102 and lays the foundation for the directional arrangement of liquid crystal molecules.

[0037] S22, dissolving the photosensitive alignment material in an organic solvent and forming a uniform thin film on the surface of the activated reflective substrate 101 by a gradient spin coating process. After pre-spreading at a low speed (500 rpm, 10 seconds), the spin coating is switched to a high speed (4000 rpm, 30 seconds) to complete the thin film homogenization. The thin film is then pre-baked at a low temperature (80°C, 15 minutes) in a vacuum environment for curing to prepare the photo-controlled alignment layer 102; A gradient spin coating process (low-speed pre-spreading + high-speed homogenization) is used to form a light-controlled alignment layer 102 thin film on the activated substrate. Combined with vacuum low-temperature pre-baking (80°C, 15 minutes) to solidify the material, it avoids high-temperature damage to the substrate and ensures the uniformity and stability of the alignment layer.

[0038] S23, using dynamic mask exposure technology to perform microstructure patterning on the photo-controlled orientation layer, the photosensitive material undergoes directional photocrosslinking, isomerization or photolysis reaction under ultraviolet polarized light irradiation, and the exposure pattern is a preset neural network phase distribution pattern.

[0039] S24, evenly spin-coating the UV-curable nematic liquid crystal onto the upper surface of the light-controlled alignment layer. Due to the interaction between molecules, the photosensitive material in the alignment layer induces the liquid crystal molecules to align along their arrangement to form the liquid crystal layer 103; Liquid crystal materials mixed with nonlinear response dopants are injected into microgrooves, and bubbles are eliminated by vacuum adsorption. The groove structure is used to guide the directional arrangement of liquid crystal molecules. Dopants enhance the phase jump sensitivity to ensure the physical realization of the nonlinear modulation function.

[0040] S25, packaging and calibration of the phase plate 405: Cover with a transparent protective layer and seal the edges, use a laser interferometer to detect the birefringence characteristics of the phase plate 405, adjust the packaging pressure so that the phase difference meets the quarter-wave condition, and complete the packaging of the phase plate 405.

[0041] The phase plate 405 is encapsulated in a transparent protective layer, and the birefringence characteristics are detected by a laser interferometer. The packaging pressure is dynamically adjusted to ensure that the phase difference accurately meets the λ / 4 requirement, completing the device performance calibration and ensuring the compatibility of the nonlinear modulation function with the optical path system.

[0042] Combine Figure 3 and Figure 4 As shown, in this embodiment, it is specifically described that the optical path system specifically includes: Laser 401, configured to output linearly polarized light with a wavelength of 633 nm; A beam expander 402 is used to expand the linearly polarized light into a parallel beam; The linear polarizer 403 is used to adjust the polarization direction of the light incident on the reflective quarter-wave phase plate 405; A reflective quarter-wave phase plate 405 is used to control the polarization and phase of the incident light field based on nonlinear phase modulation; The beam splitter prism 404 is used to separate the reflected light path to the CCD detection end; The collimating lens 406 and the CCD 407 are used to capture and analyze the amplitude and phase distribution of the output light field in real time.

[0043] The optical system operates as follows: Laser 401 outputs linearly polarized light with a wavelength of 633nm. After being expanded into a parallel beam by beam expander 402, the polarization direction of the incident light is adjusted by linear polarizer 403 before entering reflective quarter-wave phase plate 405. The incident light undergoes polarization and phase changes based on nonlinear phase modulation on the surface of phase plate 405. The reflected light is separated by beam splitter prism 404 and transmitted to the CCD detection terminal. CCD 407 captures the amplitude and phase distribution of the output light field in real time. The phase difference jump characteristics are quantified in conjunction with the polarization analysis module to verify the nonlinear response performance of the device. This optical system integrates light field control, nonlinear modulation, and real-time detection through the coordinated efforts of its components, significantly improving the verification accuracy and dynamic adjustment efficiency of nonlinear phase jumps.

[0044] In this embodiment, it is further explained that the process of integrating the reflective quarter-wave phase plate 405 into the optical path system is specifically as follows: The calibrated reflective quarter-wave phase plate 405 is fixed to the optical path bracket and aligned with the laser 401, beam expander 402 and linear polarizer 403. The tilt angle of the phase plate 405 is adjusted in real time through feedback from the beam splitter 404 to ensure that the incident light is vertical and the reflected light path coincides with the CCD407 receiving end.

[0045] In this embodiment, it is specifically described that the periodic microstructure of the mask pattern is a rectangular waveguide array designed based on the target phase jumps of 0° and 180°, and its period and width match the quarter-wave condition according to the operating wavelength (633nm).

[0046] In this embodiment, it is specifically explained that step S3 specifically includes the following steps: S31 , selecting an input polarization angle range (0°-180°, 5° step), configuring the laser 401 to output linearly polarized light of the corresponding angle, and adjusting the light field parameters incident on the phase plate 405 through the beam expander 402 and the linear polarizer 403 ; By setting the input polarization angle range (0°-180°, 5° step) and configuring the laser 401 to output linearly polarized light at the corresponding angle, combined with the beam expander 402 and the linear polarizer 403 to adjust the incident light parameters, standardized test conditions are established to provide an input benchmark for systematic verification of nonlinear responses.

[0047] S32, using CCD407 to capture the amplitude distribution of the reflected light, and combining it with the polarization analysis module to measure the polarization direction and phase difference of the outgoing light to generate a measured data set of θ-Δφ; The CCD407 is used to capture the amplitude distribution of the reflected light in real time, and the polarization analysis module is used to accurately measure the polarization direction and phase difference of the outgoing light, generating a measured data set of θ-Δφ, which provides a quantitative basis for subsequent error analysis.

[0048] S33 compares the measured data set with the target phase difference function, calculates the phase jump error (δ = |Δφ_{target} - Δφ_{measured}|), identifies the angle interval that does not meet the binary characteristics, and locates the nonlinear response defect area.

[0049] S34, if the error δ exceeds a first threshold (δ ≥ 10°), re-optimizing the phase pattern or adjusting the thickness of the phase plate 405 based on the error distribution, and updating the surface structure of the phase plate 405 by photolithography or reactive ion etching; Based on the error distribution, the phase pattern is re-optimized (adjusting the photolithography mask design) or the thickness of the phase plate 405 is corrected (±10nm), and the surface structure is updated through photolithography or reactive ion etching to dynamically correct the nonlinear modulation performance of the device.

[0050] S35, reintegrate the corrected phase plate 405 into the optical path system, and repeat steps S41-S43 until the phase jump error δ of all test angles is ≤5°, meeting the nonlinear activation requirements of the neural network.

[0051] Example 2: The present invention also provides a light diffraction neural network device with a nonlinear structure, which is manufactured using the control method of the light diffraction neural network device with a nonlinear structure as in Example 1. The light diffraction neural network device includes an integrated bracket and an optical path system integrated on the integrated bracket.

[0052] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. 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 method for controlling a nonlinear optical diffraction neural network device, characterized in that: The following steps are involved: S1, based on the Jones matrix theory, constructs the mapping relationship between the input polarization angle and the output phase difference. With the target nonlinear function as a constraint, the target phase pattern of the phase plate is iteratively optimized through the error back propagation algorithm; S2, prepare a reflective quarter-wave phase plate with nonlinear phase modulation function and integrate it into a nonlinear optical system; S3 verifies the binary nonlinear characteristics by measuring the polarization and phase difference of the outgoing light, and adjusts the phase pattern or phase plate thickness according to the measurement results until the nonlinear activation requirements of the neural network are met.

2. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11, establishing a Jones matrix model for a reflective quarter-wave plate: based on the fast axis direction of the phase plate, the properties of the reflector, and the liquid crystal molecular arrangement parameters, derive its Jones matrix expression and verify its modulation characteristics for linearly polarized light; S12, substituting the incident light polarization angle θ into the Jones matrix model, calculating the polarization direction and phase difference Δφ of the output light, generating a mathematical correlation formula of θ-Δφ, and defining the mapping relationship between the input polarization angle and the output phase difference; S13, performing discretization calculation on the correlation equation using a numerical simulation tool, plotting Δφ distribution curves under different θ values, and verifying the nonlinear characteristics of the binary phase jump; S14, defining a target phase difference function according to the activation requirements of the optical diffraction neural network, and establishing an error evaluation criterion between the target phase difference function and the measured phase difference to set the target nonlinear function constraint; S15, setting the liquid crystal azimuth angle φ of the phase plate as a trainable variable, establishing a neural network model based on the light field input-output mapping relationship, defining a loss function L, and constructing an error backpropagation framework; S16, adjusting the φ value through a gradient descent algorithm to minimize the loss function, and generating a phase distribution pattern that meets the target nonlinear response after a preset number of iterations to iteratively optimize the phase pattern.

3. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 1, characterized in that: The preparation process of the reflective quarter-wave phase plate is as follows: S11, substrate pretreatment: ultrasonically clean the reflective substrate in deionized water for 2 hours, blow dry with nitrogen, and then dry it. Then, activate the reflective substrate using plasma surface treatment technology to form a micro-nano rough surface with high adhesion; S12, dissolving the photosensitive alignment material in an organic solvent, forming a uniform thin film on the surface of the activated reflective substrate by a gradient spin coating process, pre-spreading at a low speed, switching to a high speed to complete the uniformization of the thin film, and then pre-baking and curing at a low temperature in a vacuum environment to prepare a photo-controlled alignment layer; S13, using a dynamic mask exposure technique to pattern the microstructure of the photo-controlled alignment layer, wherein the photosensitive material undergoes directional photocrosslinking, isomerization, or photolysis reaction under ultraviolet polarized light irradiation, and the exposure pattern is a preset neural network phase distribution pattern; S14, evenly spin-coating a UV-curable nematic liquid crystal onto the upper surface of the light-controlled alignment layer, wherein the photosensitive material in the alignment layer induces the liquid crystal molecules to align along their arrangement due to intermolecular interactions, thereby forming a liquid crystal layer; S15, Phase plate packaging and calibration: Cover with a transparent protective layer and seal the edges, use a laser interferometer to detect the birefringence characteristics of the phase plate, adjust the packaging pressure so that the phase difference meets the quarter-wave condition, and complete the phase plate packaging.

4. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 3, characterized in that: The optical path system specifically includes: A laser for outputting linearly polarized light with a wavelength of 633 nm; a beam expander, configured to expand the linearly polarized light into a parallel beam; Linear polarizer, used to adjust the polarization direction of light incident on the reflective quarter-wave phase plate; The reflective quarter-wave phase plate is used to control the polarization and phase of the incident light field based on nonlinear phase modulation; Beam splitter prism, used to separate the reflected light path to the CCD detection end; and CCD, which is used to capture and analyze the amplitude and phase distribution of the output light field in real time.

5. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 4, characterized in that: The process of integrating the reflective quarter-wave phase plate into the optical path system is specifically as follows: The calibrated reflective quarter-wave phase plate is fixed to the optical path bracket and aligned with the laser, the beam expander and the linear polarizer. The phase plate tilt angle is adjusted in real time through the feedback of the beam splitter prism to ensure that the incident light is vertical and the reflected light path coincides with the CCD receiving end.

6. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 1, characterized in that: The periodic microstructure of the mask pattern is specifically a rectangular waveguide array designed based on the target phase jumps of 0° and 180°, and its period and width match the quarter-wave condition according to the operating wavelength.

7. The control method of the optical diffraction neural network device with a nonlinear structure according to claim 1, characterized in that: The step S3 specifically includes the following steps: S31, selecting an input polarization angle range, configuring the laser to output linearly polarized light of the corresponding angle, and adjusting the light field parameters incident on the phase plate through a beam expander and a linear polarizer; S32, using the CCD to capture the amplitude distribution of the reflected light, and combining it with the polarization analysis module to measure the polarization direction and phase difference of the outgoing light to generate a measured data set of θ-Δφ; S33, comparing the measured data set with the target phase difference function, calculating the phase jump error (δ=|Δφ_{target}-Δφ_{measured}|), and identifying the angle interval that does not meet the binarization characteristic; S34, if the error δ exceeds a first threshold, re-optimizing the phase pattern or adjusting the thickness of the phase plate based on the error distribution, and updating the surface structure of the phase plate by photolithography or reactive ion etching; S35, reintegrate the corrected phase plate into the optical path system, and repeat steps S41-S43 until the phase jump error δ of all test angles is ≤ 5°, meeting the nonlinear activation requirements of the neural network.

8. A nonlinear optical diffraction neural network device, characterized in that: It is manufactured using the control method of the optical diffraction neural network device with a nonlinear structure as described in any one of claims 1 to 7, and the optical diffraction neural network device includes an integrated bracket and an optical path system integrated on the integrated bracket.

Images