Method and system for preparing optical neural network by using femtosecond laser direct writing
Femtosecond laser direct writing addresses the limitations of 3D printing by enabling high-precision alignment and robust material choices for optical neural networks, enhancing their performance in visible and near-infrared applications.
Patent Information
- Application Number
- CN202410051420.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-12
- Publication Date
- 2025-07-15
AI Technical Summary
In the existing optical neural network implementation methods, the multi-layer terahertz phase plate has low alignment accuracy, limited material selectivity, poor robustness, and low recognition accuracy, especially in visible light and near infrared bands.
The femtosecond laser direct writing technology is used to form a multi-layer phase plate in transparent materials at one time. The high precision and flexibility of femtosecond laser are used to realize the preparation of optical neural networks. By controlling the focus of the femtosecond laser, it is modified on each unit phase element, and stable transparent materials such as silica, lithium niobate, yttrium aluminum garnet, and borosilicate glass are selected.
The alignment accuracy and material selectivity of the multi-layer phase plate are improved, the robustness of the optical neural network is enhanced, and the efficient preparation of optical neural networks in visible light and near-infrared bands is realized.
Smart Images
Figure CN120306861A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and system for preparing an optical neural network, and particularly to a method and system for preparing an optical neural network using femtosecond laser direct writing. Background Art
[0002] At present, in many fields such as driverless and image recognition, machine learning systems such as deep learning and neural networks play an increasingly important role. However, the implementation of existing machine learning systems relies on high-performance electronic hardware such as CPUs or GPUs. Especially for deep learning and neural networks, due to the need to perform a large number of parallel computations, it is inevitable to generate huge energy consumption, and the computing speed is also an important bottleneck restricting the performance of deep learning and neural networks.
[0003] Photon computing modulates optical field information based on different dimensions such as amplitude, phase, polarization, and angular momentum in classical optics, and can thus achieve high-density and high-throughput parallel information processing, with great advantages in terms of time delay and energy consumption, and is expected to play an important role in dedicated large models.
[0004] Currently, there are various ways to implement an optical neural network. For example, the first way is the cascaded modulator based on integrated optics. This way can utilize the CMOS industrial foundation to realize chip preparation, but it is limited by the two-dimensional architecture, and its integration scale lacks scalability. Currently, the largest scale in the industry is the optical computing chips realized by Lightmatter in the United States and Xizhi Technology in China, with a matrix scale of 64×64.
[0005] In addition, the second way is the diffractive optical computing architecture. Compared with the integrated optical matrix computing scheme, the diffractive computing architecture based on free-space optics makes full use of the parallelism of the two-dimensional space plane, has good scalability, and can realize large-scale matrix computing. And this computing architecture requires high-precision three-dimensional micro-nano junction manufacturing technology. The Ozcan research group at UCLA first proposed a spatial optical computing architecture in 2018 and realized multi-layer terahertz phase plates based on 3D printing technology to perform machine learning tasks. Summary of the Invention
[0006] However, each layer of the multi-layer terahertz phase plate needs to be individually manufactured by 3D printing technology. The alignment accuracy of the multi-layer terahertz phase plate is low, and the material selectivity of 3D printing technology is low and the robustness is poor. In addition, due to the relatively long terahertz wavelength, its recognition accuracy is lower than that of visible light and near-infrared bands. Therefore, it is more meaningful to realize diffractive optical neural networks in visible light and near-infrared bands.
[0007] The present invention is proposed in view of the above situation, and its purpose is to provide a method and system for preparing an optical neural network using femtosecond laser direct writing. By using femtosecond laser direct writing to form a multi-layer phase plate of the optical neural network in a transparent material at one time, the preparation of the optical neural network in visible light and near-infrared can be realized, and the alignment accuracy of the multi-layer phase plate is relatively high; in addition, since the material of each phase plate has a large selectivity, a transparent material with relatively stable material properties can be selected, so the robustness of the optical neural network is relatively good.
[0008] Technical solution for solving the technical problem
[0009] In order to solve the above problems, in the method for preparing an optical neural network using femtosecond laser direct writing according to the first aspect of the present invention, the optical neural network includes N phase plates, where N is a positive integer greater than or equal to 2, and the method includes the following steps:
[0010] (1) Control to make the focus of the femtosecond laser located on the first-layer phase plate as the bottom layer, and sequentially perform direct writing on each unit phase element of the first-layer phase plate to achieve corresponding modification on each unit phase element of the first-layer phase plate;
[0011] (2) Control to make the focus of the femtosecond laser located on the phase plate of the upper layer, and sequentially perform direct writing on each unit phase element of the upper-layer phase plate to achieve corresponding modification on each unit phase element of the upper-layer phase plate; and
[0012] (3) Repeat the step (2) until all N phase plates are prepared.
[0013] Further, after the nth-layer phase plate is prepared, turn off the femtosecond laser, control to make the focus of the femtosecond laser located on the (n + 1)th-layer phase plate, and sequentially perform direct writing on each unit phase element of the (n + 1)th-layer phase plate to achieve corresponding modification on each unit phase element of the (n + 1)th-layer phase plate, where n is a positive integer greater than or equal to 1 and less than N.
[0014] Further, when performing direct writing on each unit phase element of each phase plate, after the direct writing of a certain unit phase element is completed, turn off the femtosecond laser, control to make the focus of the femtosecond laser move to an adjacent unit phase element until the direct writing of all unit phase elements is completed.
[0015] Further, the N phase plates use a transparent material, and the transparent material is any one of silicon dioxide, lithium niobate, yttrium aluminum garnet, and borosilicate glass.
[0016] Further, a femtosecond laser is focused using an objective lens with a specified numerical aperture.
[0017] Further, a switching operation is performed on the femtosecond laser using a high-speed shutter.
[0018] Further, the movement of the focus of the femtosecond laser in the stacking direction of the N-layer phase plate is achieved by controlling the position of the N-layer phase plate.
[0019] Further, the movement of the focus of the femtosecond laser in the stacking direction of the N-layer phase plate is achieved by controlling the position of the objective lens.
[0020] In the system for preparing an optical neural network using femtosecond laser direct writing according to the second aspect of the present invention, the optical neural network includes an N-layer phase plate, where N is a positive integer greater than or equal to 2. The system includes a femtosecond laser control unit and an objective lens control unit. The femtosecond laser control unit and the objective lens control unit respectively control a femtosecond laser and an objective lens that focuses the femtosecond laser emitted by the femtosecond laser to perform the following steps:
[0021] (1) Control is performed such that the focus of the femtosecond laser is located on the first-layer phase plate as the bottom layer, and each unit phase element of the first-layer phase plate is sequentially directly written to achieve corresponding modification in each unit phase element of the first-layer phase plate;
[0022] (2) Control is performed such that the focus of the femtosecond laser is located on the phase plate of the upper layer, and each unit phase element of the phase plate of the upper layer is sequentially directly written to achieve corresponding modification in each unit phase element of the phase plate of the upper layer; and
[0023] (3) Step (2) is repeated until all N-layer phase plates are prepared.
[0024] Further, the system further includes a phase plate control unit. The movement of the focus of the femtosecond laser in the stacking direction of the N-layer phase plate is achieved by using the phase plate control unit to control the position of the N-layer phase plate.
[0025] Further, the movement of the focus of the femtosecond laser in the stacking direction of the N-layer phase plate is achieved by using the objective lens control unit to control the position of the objective lens.
[0026] Advantages of the Invention
[0027] According to the method and system for preparing an optical neural network using femtosecond laser direct writing of the present invention, by using femtosecond laser direct writing to form multiple-layer phase plates of an optical neural network in a transparent material at one time, the preparation of an optical neural network in visible light and near-infrared can be achieved, and the alignment accuracy of the multiple-layer phase plates is relatively high; in addition, since the materials of each phase plate have a large selectivity, transparent materials with relatively stable material properties can be selected, so the optical neural network has good robustness. Description of the Drawings
[0028] Figure 1 It is a schematic diagram of the principle showing an example of a method for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention.
[0029] Figure 2 It is a flowchart of a method for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention.
[0030] Figure 3 It is a schematic diagram showing an example of a system for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention.
[0031] Figure 4 It is a schematic diagram of the principle showing the realization of a diffraction deep neural network using a multi-layer terahertz phase plate prepared by a 3D printing technology in the prior art.
[0032] Description of Reference Numerals
[0033] 1001 First-layer phase plate (the bottommost phase plate)
[0034] 1002 Second-layer phase plate
[0035] 1003 Third-layer phase plate
[0036] 1004 Fourth-layer phase plate
[0037] 1005 Fifth-layer phase plate
[0038] 1006 Sixth-layer phase plate
[0039] 2001 Objective lens
[0040] 3001 Objective lens control unit
[0041] 4001 Phase plate control unit
[0042] 5001 Femtosecond laser
[0043] 6001 Femtosecond laser control unit
[0044] 61, 62, 63 Unit phase elements
[0045] 10 System for preparing an optical neural network using femtosecond laser direct writing Detailed Embodiments
[0046] The present invention will be described in detail below in conjunction with the drawings and specific embodiments. These embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.
[0047] For convenience of description, spatial relative terms, such as "under", "below", "lower", "above", "upper", etc., may be used herein to describe the relationship of one element or feature to another element or feature as shown in the figures. It should be understood that the spatial relative terms are intended to encompass different orientations of the device in use or operation in addition to the orientation shown in the figures. For example, if the device in the figures is turned over, an element described as "under" or "below" another element or feature will be oriented "above" the other element or feature.
[0048] Unless otherwise defined, the terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terms should be understood to have a meaning consistent with the context of the related art and should not be understood in an idealized or overly formal sense unless clearly so defined herein.
[0049] <Method for Preparing an Optical Neural Network by Using Femtosecond Laser Direct Writing>
[0050] Before describing the method for preparing an optical neural network by using femtosecond laser direct writing involved in the embodiments of the present invention, the principle of the diffraction deep neural network in the prior art will be described first. Figure 4 FIG. shows the schematic principle diagram of implementing a diffraction deep neural network by using a multi-layer terahertz phase plate prepared based on 3D printing technology in the prior art.
[0051] As Figure 4 shown, the diffraction deep neural network (D 2 NN) in the prior art includes multiple layers of diffractive optical elements ( Figure 4 5 layers in the figure), and the micro-nano structure units on each layer of diffractive optical elements act as artificial neurons. During the learning and training process, when the light emitted by an object passes through the D 2 NN, a computer is used to learn the diffraction light pattern generated by each object, and the neural network is trained to recognize the object in front. Once the learning and training are completed, the system can solidify the neural network through physical processing means, and then control the optical diffraction propagation mode between layers to achieve the preset recognition task in the form of a specific output light field distribution. For example, when the handwritten "5" shown in Figure 4 is placed in front of the D 2 NN, through the optical diffraction of the D 2 NN, a specific output light field distribution will be formed behind the D 2 NN (i.e., a relatively bright light spot will be generated at the position of the detector 5), and the detector arranged behind the D 2 NN can recognize the handwritten "5" object.
[0052] However, inFigure 4 In the existing diffractive deep neural network shown, each layer of the multi-layer terahertz phase plate needs to be individually manufactured by 3D printing technology. The alignment accuracy of the multi-layer terahertz phase plate is low, and the material selectivity of 3D printing technology is low, and the robustness is poor. In addition, due to the long terahertz wavelength, its recognition accuracy is lower compared to the visible light and near-infrared bands.
[0053] To solve the above problems, the present invention proposes a method and system for preparing an optical neural network using femtosecond laser direct writing. The femtosecond laser direct writing technology has flexible three-dimensional micro-nano processing capabilities, and can achieve maskless, super-resolution, and high-quality three-dimensional processing. In the present invention, the femtosecond laser can directly write three-dimensional patterns arbitrarily in a transparent material, can flexibly construct any number of network layers and neurons of any size in a large range, and can cause a change in the localized refractive index. That is, in the present invention, by using femtosecond laser direct writing to form a multi-layer phase plate of an optical neural network in a transparent material at one time, the preparation of an optical neural network in visible light and near-infrared can be realized, and the alignment accuracy of the multi-layer phase plate is high; in addition, due to the large material selectivity of each phase plate, a transparent material with relatively stable material properties can be selected, so the optical neural network has good robustness.
[0054] Next, with reference to Figure 1 and Figure 2 , the method for preparing an optical neural network using femtosecond laser direct writing according to the embodiments of the present invention will be described in detail. Figure 1 FIG. is a schematic principle diagram showing an example of the method for preparing an optical neural network using femtosecond laser direct writing according to the embodiments of the present invention. Figure 2 FIG. is a flowchart showing the method for preparing an optical neural network using femtosecond laser direct writing according to the embodiments of the present invention.
[0055] As shown in Figure 1 , as an example, the optical neural network in the embodiments of the present invention includes 6 layers of phase plates, that is, the first layer of phase plate 1001, the second layer of phase plate 1002, the third layer of phase plate 1003, the fourth layer of phase plate 1004, the fifth layer of phase plate 1005, and the sixth layer of phase plate 1006 as the bottom layer, and each phase plate corresponds to a network layer of the neural network.
[0056] In the method for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention, first, in step ST101, femtosecond laser light from the outside is focused by an objective lens 2001 having a specified numerical aperture, and is controlled so that the focal point of the femtosecond laser is located on the first layer phase plate 1001 as the bottom layer, and direct writing is sequentially performed on each unit cell of the first layer phase plate 1001 to achieve corresponding modification on each unit cell of the first layer phase plate 1001.
[0057] Next, in step ST102, after direct writing has been performed on all unit cells on the first layer phase plate 1001, that is, after the preparation of the first layer phase plate 1001 is completed, the femtosecond laser is turned off, and then it is controlled so that the focal point of the femtosecond laser is located on the second layer phase plate 1002, and direct writing is sequentially performed on each unit cell of the second layer phase plate 1002 to achieve corresponding modification on each unit cell of the second layer phase plate 1002.
[0058] Next, in step ST103, after direct writing has been performed on all unit cells on the second layer phase plate 1002, that is, after the preparation of the second layer phase plate 1002 is completed, the femtosecond laser is turned off, and then it is controlled so that the focal point of the femtosecond laser is located on the third layer phase plate 1003, and direct writing is sequentially performed on each unit cell of the third layer phase plate 1003 to achieve corresponding modification on each unit cell of the third layer phase plate 1003.
[0059] Then, similarly, direct writing is performed on each unit cell of the fourth layer phase plate 1004, the fifth layer phase plate 1005, and the sixth layer phase plate 1006 to achieve corresponding modification on each unit cell of the fourth layer phase plate 1004, the fifth layer phase plate 1005, and the sixth layer phase plate 1006.
[0060] As Figure 1 shown in the upper right of, the sixth layer phase plate 1006 includes a plurality of unit cells 61, 62, 63 ···, and each unit cell corresponds to a neuron of the neural network. Different shades of each unit cell 61, 62, 63 ··· represent different modifications, which will cause the refractive index (diffraction angle) of light passing through each unit cell to change.
[0061] Since an optical neural network can be formed at one time in the stacked multi-layer phase plates by using femtosecond laser direct writing, the alignment accuracy of the multi-layer phase plates constituting the optical neural network is relatively high.
[0062] In addition, Figure 1 in the example of, an embodiment of a six-layer phase plate is shown, but the number of layers of the phase plate in the present invention is not particularly limited.
[0063] In addition, each phase plate can use a transparent material, which can be any one of silica, lithium niobate, yttrium aluminum garnet, and borosilicate glass.
[0064] In addition, the switch of the femtosecond laser can be operated by using a high-speed shutter (not shown in the figure).
[0065] In addition, after the preparation of a certain layer of the phase plate is completed, it is necessary to move the focus of the femtosecond laser to the upper layer of the phase plate. At this time, the movement of the focus of the femtosecond laser in the stacking direction of the multi-layer phase plate can be achieved by controlling the position of the multi-layer phase plate, or can be achieved by controlling the position of the objective lens.
[0066] Thus, according to the method for preparing an optical neural network using femtosecond laser direct writing of the present invention, by using femtosecond laser direct writing to form the multi-layer phase plates of the optical neural network in a transparent material at one time, the preparation of the optical neural network in visible light and near-infrared can be realized, and the alignment accuracy of the multi-layer phase plates is relatively high; in addition, since the materials of each phase plate have a large selectivity, a transparent material with relatively stable material properties can be selected, so the robustness of the optical neural network is relatively good.
[0067] <System for Preparing an Optical Neural Network Using Femtosecond Laser Direct Writing>
[0068] Next, with reference to Figure 3 , a system for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention will be described in detail. Figure 3 FIG. is a schematic diagram showing an example of a system for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention.
[0069] As Figure 3 shown, the system 10 for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention at least includes a femtosecond laser control unit 6001 and an objective lens control unit 3001. Among them, the femtosecond laser control unit 6001 can control the three-dimensional movement of the femtosecond laser 5001 and the intensity of the femtosecond laser; the objective lens control unit 3001 can control the three-dimensional movement of the objective lens 2001 that focuses the femtosecond laser emitted by the femtosecond laser 5001.
[0070] By using the femtosecond laser control unit 6001 to control the femtosecond laser 5001 and using the objective lens control unit 3001 to control the objective lens 2001, it is possible to execute Figure 2The method shown. That is, it is controlled so that the focus of the femtosecond laser is located on the first-layer phase plate 1001 which is the bottom layer, and direct writing is sequentially performed on each unit phase element of the first-layer phase plate 1001 to achieve corresponding modification on each unit phase element of the first-layer phase plate 1001. Then, after direct writing has been performed on all the unit phase elements on the first-layer phase plate 1001, that is, after the preparation of the first-layer phase plate 1001 is completed, the femtosecond laser is turned off, and then it is controlled so that the focus of the femtosecond laser is located on the second-layer phase plate 1002, and direct writing is sequentially performed on each unit phase element of the second-layer phase plate 1002 to achieve corresponding modification on each unit phase element of the second-layer phase plate 1002. Then, after direct writing has been performed on all the unit phase elements on the second-layer phase plate 1002, that is, after the preparation of the second-layer phase plate 1002 is completed, the femtosecond laser is turned off, and then it is controlled so that the focus of the femtosecond laser is located on the third-layer phase plate 1003, and direct writing is sequentially performed on each unit phase element of the third-layer phase plate 1003 to achieve corresponding modification on each unit phase element of the third-layer phase plate 1003. And, similarly, direct writing is performed on each unit phase element of the fourth-layer phase plate 1004, the fifth-layer phase plate 1005, and the sixth-layer phase plate 1006 to achieve corresponding modification on each unit phase element of the fourth-layer phase plate 1004, the fifth-layer phase plate 1005, and the sixth-layer phase plate 1006.
[0071] In addition, in the above example, the movement of the focus of the femtosecond laser in the stacking direction of the multi-layer phase plate is achieved by controlling the position of the objective lens, but the present invention is not limited thereto. For example, it can also be achieved by controlling the position of the multi-layer phase plate. In this case, the system 10 for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention further includes a phase plate control unit 4001 for controlling the position of the multi-layer phase plate ( Figure 3 shown by a dashed box in).
[0072] In addition, the system 10 for preparing an optical neural network using femtosecond laser direct writing according to an embodiment of the present invention may also include a femtosecond laser 5001 and an objective lens 2001 ( Figure 3 both shown by dashed boxes in). Of course, the system 10 of the present invention may also not include the femtosecond laser 5001 and the objective lens 2001, but only include a femtosecond laser control unit 6001 and an objective lens control unit 3001 to control the femtosecond laser 5001 and the objective lens 2001 outside the system.
[0073] Therefore, the system for preparing an optical neural network using femtosecond laser direct writing according to the present invention can form multi-layer phase plates of the optical neural network in a transparent material at one time by using femtosecond laser direct writing, so as to realize the preparation of the optical neural network in visible light and near-infrared, and the alignment accuracy of the multi-layer phase plates is relatively high; in addition, since the materials of the respective phase plates have a large selectivity, transparent materials with relatively stable material properties can be selected, so the optical neural network has good robustness.
[0074] It should be understood that the above description is illustrative rather than restrictive. For example, the above embodiments (and / or aspects thereof) can be used in combination with each other. In addition, many modifications can be made without departing from the scope of the present invention to adapt a specific situation or material to the teachings of various embodiments of the present invention. Although the sizes and types of the materials described herein are used to define the parameters of various embodiments of the present invention, the various embodiments are not meant to be restrictive but are exemplary embodiments. Many other embodiments will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of the various embodiments of the present invention should be determined with reference to the appended claims and the full scope of the equivalent forms claimed thereby.
[0075] Industrial applicability
[0076] The method and system for preparing an optical neural network using femtosecond laser direct writing of the present invention can be widely applied to the preparation of optical neural networks in many fields such as driverless and image recognition.
Claims
1. A method for preparing an optical neural network using femtosecond laser direct writing. The optical neural network includes N phase plates, where N is a positive integer greater than or equal to 2. The method includes the following steps: (1) Control to make the focus of the femtosecond laser locate at the first layer of phase plate as the bottom layer, and sequentially perform direct writing on each unit cell of the first layer of phase plate to achieve corresponding modification on each unit cell of the first layer of phase plate; (2) Control to make the focus of the femtosecond laser locate at the phase plate of the upper layer, and sequentially perform direct writing on each unit cell of the phase plate of the upper layer to achieve corresponding modification on each unit cell of the phase plate of the upper layer; And (3) Repeat the step (2) until all N layer phase plates are prepared.
2. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1, wherein after the nth layer of phase plate is prepared, turn off the femtosecond laser, control to make the focus of the femtosecond laser locate at the (n + 1)th layer of phase plate, and sequentially perform direct writing on each unit cell of the (n + 1)th layer of phase plate to achieve corresponding modification on each unit cell of the (n + 1)th layer of phase plate, where n is a positive integer greater than or equal to 1 and less than N.
3. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1 or 2, wherein when performing direct writing on each unit cell of each phase plate, after the direct writing of a certain unit cell is completed, turn off the femtosecond laser, control to make the focus of the femtosecond laser move to the adjacent unit cell until the direct writing of all unit cells is completed.
4. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1 or 2, wherein the N layer phase plates use a transparent material, the transparent material is any one of silicon dioxide, lithium niobate, yttrium aluminum garnet, and borosilicate glass.
5. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1 or 2, wherein focus the femtosecond laser using an objective lens with a specified numerical aperture.
6. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1 or 2, wherein perform a switching operation on the femtosecond laser using a high-speed shutter.
7. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 1 or 2, wherein the movement of the focus of the femtosecond laser along the stacking direction of the N layer phase plates is achieved by controlling the position of the N layer phase plates.
8. The method for preparing an optical neural network using femtosecond laser direct writing according to claim 5, wherein the movement of the focus of the femtosecond laser along the stacking direction of the N layer phase plates is achieved by controlling the position of the objective lens.
9. A system for fabricating an optical neural network using femtosecond laser direct writing, the optical neural network comprising N phase plates, where N is a positive integer greater than or equal to 2. The system includes a femtosecond laser control unit and an objective lens control unit, and the femtosecond laser control unit and the objective lens control unit respectively control a femtosecond laser and an objective lens that focuses the femtosecond laser emitted by the femtosecond laser to perform the following steps: (1) Control is such that the focus of the femtosecond laser is located on the first layer of phase plates, which is the bottommost layer, and each unit cell of the first layer of phase plates is sequentially directly written to achieve corresponding modification in each unit cell of the first layer of phase plates; (2) Control is such that the focus of the femtosecond laser is located on the phase plate of the upper layer, and each unit cell of the phase plate of the upper layer is sequentially directly written to achieve corresponding modification in each unit cell of the phase plate of the upper layer; and (3) Repeat step (2) until all N layer phase plates are fabricated.
10. The system for fabricating an optical neural network using femtosecond laser direct writing according to claim 9, wherein the system further includes a phase plate control unit, the movement of the focus of the femtosecond laser in the stacking direction of the N layer phase plates is achieved by controlling the position of the N layer phase plates using the phase plate control unit.
11. The system for fabricating an optical neural network using femtosecond laser direct writing according to claim 9 or 10, wherein the movement of the focus of the femtosecond laser in the stacking direction of the N layer phase plates is achieved by controlling the position of the objective lens using the objective lens control unit.