Dynamic adjustable mode exchanger based on phase change material diffraction neural network

Abstract

The invention discloses a dynamic adjustable mode exchanger based on a phase change material diffraction neural network. The mode exchanger adopts a reflective diffraction neural network architecture. The reflective diffraction neural network comprises a stepped phase plate made of a two-dimensional phase change material, the height of the stepped phase plate is designed to be different, the optical fiber input and output port and the optical reflection layer are arranged at the top of the reflective diffraction neural network, and the optical reflection layer, the pulse control circuit and the temperature detection circuit are arranged at the bottom of the reflective diffraction neural network. The electric pulse output of the pulse control circuit is controlled through the temperature detection circuit, and the phase change proportion of the phase change phase plate is regulated and controlled, so that the accurate regulation and control of the phase response of the phase change phase plate are realized. In order to optimize the height distribution of the stepped phase modulation layer, the mode exchanger provided by the invention adopts a gradient descent algorithm to optimize a height distribution function.

Description

Technical Field

[0001] The invention relates to the field of optical communication and optoelectronic technology, and in particular to a dynamically adjustable mode switch based on a phase change material diffraction neural network. Background Art

[0002] With the rapid development of big data and artificial intelligence technologies, data throughput has shown explosive growth from the Internet to data centers. To meet this challenge, modern society has continuously increased its requirements for information transmission and distribution efficiency, driving modern communication systems to gradually develop in the direction of ultra-large capacity and ultra-high speed. In the past few decades, in order to improve the transmission bandwidth of optical fiber communication systems, multiplexing technologies such as polarization multiplexing (PDM), wavelength division multiplexing (WDM) and time division multiplexing (TDM) have been proposed. These technologies have effectively alleviated the bandwidth anxiety in the early big data era. However, with the further development of artificial intelligence technology, especially the advent of the big model era, the current multiplexing technology is still difficult to meet the required bandwidth requirements, so mode multiplexing has gradually been put on the agenda and has become a research focus.

[0003] Since the orbital angular momentum (OAM) mode was proposed by Professor Allen in 1992, it has quickly become a research hotspot in the field of optics in the past decade due to its infinite number of orthogonal modes and excellent anti-interference and weak crosstalk characteristics in optical fiber communication systems. In recent years, it has also been innovatively applied to optical fiber communication systems to break through the capacity limitations of existing communication systems.

[0004] As an important component of the optical fiber communication system based on OAM mode multiplexing, optical mode switching can realize information exchange between various mode channels, and provides key technical support for the flexible networking and traffic distribution of optical fiber communication systems, which has great practical significance. However, the current traditional optical switching solutions are mainly oriented to conventional dimensions such as wavelength switching or space switching, relying on optical multiplexers / demultiplexers and optoelectronic conversion systems. This approach is difficult to extend to the spatial mode dimension, and the demultiplexing and multiplexing systems using spiral phase plates (SPPs) have many shortcomings, including large system size, complex structure, difficulty in reconstruction, and low degree of integration. In addition, such solutions are usually only applicable to a small and limited number of modes, and cannot meet the actual needs of large-scale mode switching.

[0005] To address this problem, fiber-based mode switching systems have emerged. However, such systems also face many challenges and problems, such as large size and complex system. They can only be applied to the exchange of a small number of modes. At present, it seems that solutions based on fiber systems are not enough to solve the problem of flexible information exchange between large-scale mode channels. In recent years, the rapid development of optical neural networks (ONNs) has opened up a new research direction for optical mode switching. In on-chip solutions, by designing the topological structure of diffractive neural networks and regulating the weights of each node in the network, research has shown that mode switching can be achieved efficiently and flexibly. However, on-chip systems still face some technical bottlenecks on the road to practical application. For example, the system still needs to introduce mode multiplexing, demultiplexing and mode couplers. These modules not only increase the complexity of the system in practical applications, but also put forward higher requirements for its reliability and efficiency. In addition, with the increase in the number of multiplexed modes, the complexity of the system control module will increase exponentially, making it difficult to apply. In the future, how to optimize the structure of the optical mode switching system, simplify the complexity of the mode switching system, and improve the reconfigurability and scalability of the system will become the core research direction for achieving large-scale and efficient mode switching. Summary of the invention

[0006] The purpose of the present invention is to provide a dynamically adjustable mode switcher based on a phase change material diffraction neural network to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A dynamically adjustable mode exchanger based on a phase change material diffraction neural network, the dynamically adjustable mode exchanger comprises, from top to bottom, a first input / output port and a second input / output port embedded in a substrate layer and an optical reflective layer, an upper substrate layer, an upper optical reflective layer, a protective filling layer, a phase modulation layer based on a non-volatile phase change material, a lower optical reflective layer and a lower substrate layer, and also comprises a temperature detection circuit and a pulse control circuit connected in series, the temperature detection circuit and the pulse control circuit are respectively connected to the two ends of the lower optical reflective layer, the relative refractive index of the phase change material is regulated by the temperature detection circuit and the pulse control circuit, and the phase modulation amount is changed to realize different mode exchange functions.

[0009] The technical solution is further optimized, wherein the first input / output port and the second input / output port are both optical fibers, fixed to the upper surface of the state-adjustable mode exchanger by UV-curing adhesive, wherein the input / output mode is an optical orbital angular momentum mode.

[0010] The technical solution is further optimized in that the upper substrate layer and the lower substrate layer are silicon dioxide.

[0011] The technical solution is further optimized in that the upper optical reflective layer and the lower optical reflective layer are both gold films, and their reflectivity is greater than 95%.

[0012] The technical solution is further optimized, and the protective filling layer is SU-83000 photoresist.

[0013] The technical solution is further optimized in that the non-volatile phase change material used in the phase modulation layer is tellurium selenide, and its crystallization state can be changed by applying external stimulation to the phase change material.

[0014] The technical solution is further optimized in that the phase modulation layer adopts a stepped structure and the reflection phase is regulated by regulating the height of each unit.

[0015] The technical solution is further optimized. The dynamically adjustable mode switch uses a bidirectional diffraction neural network optimization algorithm based on gradient descent to optimize the height distribution function of the phase modulation layer. Based on the bidirectional diffraction neural network optimization model of gradient descent, after the input and output modes, working wavelength, and device thickness are known, the output light field distribution in the forward and backward directions of the system under different phase change states is numerically calculated by combining the angular spectrum diffraction forward and backward propagation algorithms based on the non-paraxial approximation with the initialized height distribution function; the crosstalk suppression loss function is used to evaluate the matching degree and crosstalk level of the output light field of the switch under different phase change states with the target mode; the gradient of the loss function to the height distribution function is calculated, and the gradient descent algorithm is used to update and optimize the height distribution function until the loss function converges.

[0016] The technical solution is further optimized, and the specific steps of the bidirectional diffraction neural network optimization algorithm based on gradient descent are as follows:

[0017] First, the surface height function is initialized, and the loss function is also initialized; next, the phase change state loop is entered, the output light field distribution is calculated by the angular spectrum diffraction algorithm, and the loss function at this time is accumulated and calculated in combination with the target light field distribution. The loss function is related to the output mode of each port in N states. Taking the first input and output port to the second input and output port as an example, the loss function L 1to2 It is expressed as:

[0018]

[0019] Where k represents the sum of the loss function phase transition states, i represents the sum of the loss function of N modes at the kth phase transition state, O i represents the output light field of the ith mode in the kth phase transition state, T i represents the target light field of the i-th mode at the k-th phase transition state, T jRepresents the non-target light field of the i-th mode in the k-th phase transition state, and the loss function L from the second input and output port to the first input and output port 2to1 It is the same, except that the input and output ports are swapped. The current loss function is obtained by adding the two together. Then, the gradient of the loss function is calculated and the height distribution function is updated, and the next cycle is entered again until the final loss function converges, which means that the height distribution function at this time is the local optimal solution in the current parameter space.

[0020] The technical solution is further optimized, and the thickness of the device is the thickness from the upper surface gold film to the lower surface gold film.

[0021] Different from the prior art, the above technical solution has the following beneficial effects:

[0022] 1) The present invention adopts a spatial diffraction neural network structure, which has the advantages of large parameter space, large throughput, and strong scalability compared with the on-chip system. Compared with the optical fiber system, it has the advantages of high integration, small size, and simple design. It can achieve high-purity, low crosstalk, and low-loss direct conversion of spatial modes without multiplexing and demultiplexing the modes.

[0023] 2) The present invention uses phase change material as the phase modulation layer. The combination of excellent performance and non-volatile phase change material can realize dynamic optical mode switching, which is particularly effective for the current OAM mode multiplexing technology. At the same time, this solution only needs to control the temperature of the phase change material to achieve functional reconstruction, and its detection and control methods are simple.

[0024] 3) The present invention has a simple structure, mature processing technology, and a wide range of practical applications. The phase change material used exhibits low loss in the visible light to near-infrared bands and can work in any band only by being optimized according to the working wavelength. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic front view of the structure of a dynamically adjustable mode exchanger based on a phase change material diffraction neural network;

[0026] Figure 2 A schematic top view of the structure of a dynamically adjustable mode exchanger based on a phase change material diffractive neural network;

[0027] Figure 3 The complex amplitude distribution diagram of the output light field of the two ports of the dynamically adjustable mode switcher based on the phase change material diffraction neural network under different phase change states;

[0028] Figure 4 It is the transmission matrix diagram of the output light field of the two ports of the dynamically adjustable mode switcher based on the phase change material diffraction neural network under different phase change states;

[0029] Figure 5 This is the height distribution function diagram obtained by the bidirectional diffraction neural network optimization algorithm based on gradient descent;

[0030] Figure 6 Flowchart of the gradient descent-based bidirectional diffractive neural network optimization algorithm for optimizing the height distribution of a dynamically adjustable mode exchanger based on a phase change material diffractive neural network.

[0031] In the figure: mode exchanger 1, base layer 2, optical reflective layer 3, protective filling layer 4, phase modulation layer 5, first input and output port 6, second input and output port 7, light-curing UV glue 8, pulse control circuit 9, temperature detection circuit 10. DETAILED DESCRIPTION

[0032] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0033] See also Figure 1 As shown, it is a schematic front view of the structure of a dynamically adjustable mode switcher based on a phase change material diffraction neural network. A preferred embodiment of the present invention is a dynamically adjustable mode switcher based on a phase change material diffraction neural network. The structure of the dynamically adjustable mode switcher includes, from top to bottom, a first input / output port 6 and a second input / output port 7 embedded in a substrate layer 2 and an optical reflective layer 3, an upper substrate layer 2, an upper optical reflective layer 3, a protective filling layer 4, a phase modulation layer 5 based on a non-volatile phase change material, a lower optical reflective layer 3, a lower substrate layer 2, and a temperature detection circuit 10 and a pulse control circuit 9. The protective filling layer 4 is used as a filling part to support the upper and lower surfaces for fixing the device. The temperature detection circuit 10 is connected to the pulse control circuit 9, and a loop is realized through the gold film of the optical reflective layer 3 on the lower surface, thereby realizing the detection of temperature and the control of pulses. The relative refractive index of the phase change material is regulated by the temperature detection circuit 10 and the pulse control circuit 9, and the phase modulation amount is changed to realize different mode exchange functions.

[0034] The first input / output port 6 and the second input / output port 7 are optical fibers fixed to the upper surface of the mode exchanger by UV curing adhesive, wherein the input / output mode is an optical orbital angular momentum mode.

[0035] The base layer 2 is silicon dioxide, the optical reflective layer 3 is a gold film with a reflectivity greater than 95%. The protective filling layer 4 is SU-83000 photoresist, and the non-volatile phase change material used in the phase modulation layer 5 is tellurium selenide, and its crystallization state can be changed by applying external stimulation to the phase change material.

[0036] The phase modulation layer 5 adopts a stepped structure, and the reflection phase is regulated by regulating the height of each unit.

[0037] The dynamically adjustable mode exchanger adopts a reflective diffraction neural network architecture. A two-dimensional phase change material stepped phase plate with different heights is constructed in the reflective diffraction neural network as a phase modulation layer 5. The mode exchanger 1 also includes a first optical fiber input and output port 6, a second input and output port 7, an optical reflection layer 3 and an optical reflection layer 3 at the bottom, a pulse control circuit 9 and a temperature detection circuit 10 arranged at the top of the reflective diffraction neural network. By controlling the pulse output of the pulse control circuit 9, the phase change ratio of the phase change material in the phase modulation layer 5 is regulated to regulate its phase response. Since the phase modulation amount is related to the height distribution of the device surface, in order to realize the mode exchange function under different states, the mode exchanger proposed in the present invention uses a gradient descent algorithm to optimize the height distribution function of the stepped phase modulation layer, and uses the thermal phase change effect of the phase material to change the function of the diffraction neural network by means of electric pulses. The input optical OAM mode is modulated and reflected multiple times inside the mode exchanger 1 and converted into the target mode at another port to realize the mode exchange.

[0038] The height distribution function of the phase modulation layer is obtained by a bidirectional diffraction neural network optimization model based on gradient descent. The bidirectional diffraction neural network optimization model based on gradient descent, after the input and output modes, working wavelength, and device thickness are known, uses the angular spectrum diffraction forward and backward propagation algorithm based on the non-paraxial approximation combined with the initialization height distribution function to numerically calculate the output light field distribution of the system in different phase change states in the forward and backward directions; uses the crosstalk suppression loss function to evaluate the matching degree and crosstalk level of the output light field of the switcher in different phase change states with the target mode; calculates the gradient of the loss function to the height distribution function and uses the gradient descent algorithm to update and optimize the height distribution function until the loss function converges. The device thickness is the thickness from the upper surface gold film to the lower surface gold film. The temperature detection circuit 10 monitors the temperature of the gold film surface and the phase change material in real time, and the pulse control circuit 9 feeds back the system temperature to achieve the regulation of the phase change ratio of the phase change material, and realizes the dynamic exchange of different modes under different phase change ratios.

[0039] The key component of the mode switch is the phase modulation layer 5, whose stepped height distribution function determines the function of the diffraction neural network. Therefore, the core technology of the present invention is the calculation model of the height distribution function. The height distribution function of the phase modulation layer is obtained by a bidirectional diffraction neural network optimization model based on gradient descent. To achieve specific functions under specific phase change states, it is necessary to effectively describe its effective refractive index. As tellurium selenide transitions from an amorphous phase to a fully crystallized phase, its effective dielectric constant depends on the crystallization fraction m and follows the Lorentz relationship:

[0040]

[0041] in, and Represent amorphous a-Sb 2 Se 3 and crystalline c-Sb 2 Se 3 The wavelength λ is related to the dielectric constant. ε ef f (λ) is the effective dielectric parameter. When the crystal fraction m increases from 0 to 1, it means that the intermediate state changes from amorphous a-Sb 2 Se 3 Transformed into crystalline c-Sb 2 Se 3 , thereby causing multi-level changes in optical constants. In the present invention, the crystallization fraction is controlled by a pulse control circuit and a temperature detection circuit. At a given wavelength and crystallization fraction, the wavefront control amount of the reflective surface can be predicted and controlled. For the mutual exchange of N modes, at least N phase change states are required to realize that any input mode can be transformed into any output mode through the network, that is, the crystallization fraction m = {0, 1 / (N-1), 2 / (N-1), ..., 1}.

[0042] The flowchart of the bidirectional diffractive neural network optimization algorithm based on gradient descent for optimizing the height distribution of the dynamically adjustable mode exchanger based on the phase change material diffractive neural network proposed in the present invention is as follows: Figure 6 As shown. Specifically, the surface height function is first initialized, and it is usually set to a constant. Then the loss function is also initialized. At this time, the loss function is converged, so the next step is to calculate a new loss function. Then, in order to realize the corresponding functions in different phase change states, the new loss function is related to the output light field in each phase change state, so it is necessary to enter the phase change state cycle. In each phase change state, the output light field distribution is calculated by the angular spectrum diffraction algorithm, and the loss function at this time is cumulatively calculated in combination with the target light field distribution, where the loss function is related to the output mode of each port in N states. Taking the first input and output port 6 to the second input and output port 7 of the top optical fiber of the device as an example, the loss function L LtoR It is expressed as:

[0043]

[0044] Where k represents the sum of the loss function phase change state, and i represents the sum of the loss function of N modes at the kth phase change state. i represents the output light field of the i-th mode in the k-th phase transition state, T i represents the target light field of the i-th mode at the k-th phase transition state, T j represents the non-target light field of the i-th mode in the k-th phase transition state. The loss function L from the second input-output port 7 to the first input-output port 6 RtoLThe same is true, except that the target modes of the input and output ports are swapped. The loss function is calculated in different phase change states and accumulated to obtain the final new loss function. At this point, the phase change cycle ends. In the next step, in order to update the current height distribution, the gradient of the loss function is calculated and the height distribution function is updated, and the next loss function solution cycle is entered again until the final loss function converges, which means that the height distribution function at this time is the local optimal solution in the current parameter space.

[0045] The present invention proposes a dynamically adjustable mode exchanger for optimizing the diffraction neural network based on phase change materials. Due to the physical architecture of spatial diffraction, this scheme can only realize the exchange between the cyclic permutation groups between modes. Specifically, assuming that the mode exchanger is used for the mutual exchange of N modes, it is expressed as (O 1 ,O 2 ,…,O N-1 ,O N ), the switch proposed by the present invention can only achieve (O 1 ,O 2 ,…,O N-1 ,O N ) to any element in its cyclic commutative group, such as (O 2 ,O 3 ,…,O N ,O 1 ),(O 3 ,O 4 ,…,O 1 ,O 2 )wait.

[0046] When the device operates at a wavelength of 1550nm, the amorphous a-Sb 2 Se 3 and crystalline c-Sb 2 Se 3 The refractive indexes of the protective filling layer 4 are 3.2850 and 4.0500 respectively, the refractive index of the SU-83000 photoresist is 1.5552, and the device thickness is 16 mm. The pixel size of the phase modulation layer 5 is 8 μm × 8 μm, the number of pixels per layer is 640 × 640, and the depth of the diffraction neural network is 6 layers. The number of input and output OAM modes is 3, and the topological charges are l 1 =1,l 2 =2 and l 3 =3. Therefore, the mode switch has three states, and the corresponding crystallization fractions are m = 0.0, 0.5 and 1.0. The height distribution function obtained by the bidirectional diffraction neural network optimization algorithm based on gradient descent is as follows Figure 5As shown, its maximum value is 339nm. The height distribution of the dynamically adjustable mode exchanger based on phase change material diffraction neural network at a working wavelength of 1550nm has a pixel size of 8μm×8μm and a pixel number of 640×3840.

[0047] When in use, when m=0.0, the topological charge mapping relationship of the first and second input and output ports 6 and 7 is (1,2,3) mapped to (1,2,3); when m=0.5, the mapping relationship of the first and second input and output ports 6 and 7 is (1,2,3) mapped to (2,3,1); when m=1.0, the mapping relationship of the first and second input and output ports 6 and 7 is (1,2,3) mapped to (3,1,2); under different phase change ratios, the corresponding output light field complex amplitude is as follows Figure 3 As shown in , the exchange between modes can be achieved in different phase change states. The confusion matrix between the output light field and the target light field is as follows: Figure 4 As shown in the figure, thanks to the design of the loss function, high-purity, low-crosstalk spatial mode conversion can be directly achieved in different phase change states, and its loss is less than 3dB. At the same time, the device performance shows good consistency in different states. The physical size of the device at this time is 5.12mm×30.72mm×16mm. Compared with the current mode switching device, it is small in size, compact in structure, and highly integrated.

[0048] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "include..." or "comprise..." do not exclude the existence of other elements in the process, method, article or terminal device including the elements. In addition, in this article, "greater than", "less than", "exceed" and the like are understood to exclude the number itself; "above", "below", "within" and the like are understood to include the number itself.

[0049] Although the above embodiments have been described, once those skilled in the art know the basic creative concepts, they can make additional changes and modifications to these embodiments. Therefore, the above description is only an embodiment of the present invention and does not limit the patent protection scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the specification and drawings of the present invention, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A dynamically adjustable mode switcher based on a phase change material diffraction neural network, characterized in that: The dynamically adjustable mode exchanger includes, from top to bottom, a first input / output port and a second input / output port embedded in a substrate layer and an optical reflective layer, an upper substrate layer, an upper optical reflective layer, a protective filling layer, a phase modulation layer based on a non-volatile phase change material, a lower optical reflective layer, and a lower substrate layer, and also includes a temperature detection circuit and a pulse control circuit connected in series, wherein the temperature detection circuit and the pulse control circuit are respectively connected to the two ends of the lower optical reflective layer, and the relative refractive index of the phase change material is regulated by the temperature detection circuit and the pulse control circuit, and the phase modulation amount is changed to realize different mode exchange functions.

2. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The first input / output port and the second input / output port are both optical fibers, fixed to the upper surface of the state-adjustable mode exchanger by UV-curing adhesive, wherein the input / output mode is an optical orbital angular momentum mode.

3. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The upper substrate layer and the lower substrate layer are silicon dioxide.

4. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The upper optical reflective layer and the lower optical reflective layer are both gold films, and the reflectivity thereof is greater than 95%.

5. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The protective filling layer is SU-83000 photoresist.

6. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The non-volatile phase-change material used in the phase modulation layer is tellurium selenide, and its crystallization state can be changed by applying external excitation to the phase-change material.

7. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The phase modulation layer adopts a stepped structure, and the reflection phase is regulated by regulating the height of each unit.

8. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 1, characterized in that: The dynamically adjustable mode switch adopts a bidirectional diffraction neural network optimization algorithm based on gradient descent to optimize the height distribution function of the phase modulation layer. Based on the bidirectional diffraction neural network optimization model based on gradient descent, after the input and output modes, working wavelength, and device thickness are known, the output light field distribution in the forward and backward directions of the system under different phase change states is numerically calculated by combining the initialization height distribution function with the angular spectrum diffraction forward and backward propagation algorithms based on the non-paraxial approximation; the matching degree and crosstalk level of the output light field of the switch under different phase change states with the target mode are evaluated by using the crosstalk suppression loss function; the gradient of the loss function to the height distribution function is calculated, and the gradient descent algorithm is used to update and optimize the height distribution function until the loss function converges.

9. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 8, characterized in that: The specific steps of the bidirectional diffractive neural network optimization algorithm based on gradient descent are as follows: First, the surface height function is initialized, and the loss function is also initialized; next, the phase change state loop is entered, the output light field distribution is calculated by the angular spectrum diffraction algorithm, and the loss function at this time is accumulated and calculated in combination with the target light field distribution. The loss function is related to the output mode of each port in N states. Taking the first input and output port to the second input and output port as an example, the loss function L 1to2 It is expressed as: Where k represents the sum of the loss function phase transition states, i represents the sum of the loss function of N modes at the kth phase transition state, O i represents the output light field of the i-th mode in the k-th phase transition state, T i represents the target light field of the i-th mode at the k-th phase transition state, T j Represents the non-target light field of the i-th mode in the k-th phase transition state, and the loss function L from the second input and output port to the first input and output port 2to1 It is the same, except that the input and output ports are swapped. The current loss function is obtained by adding the two together. Then, the gradient of the loss function is calculated and the height distribution function is updated, and the next cycle is entered again until the final loss function converges, which means that the height distribution function at this time is the local optimal solution in the current parameter space.

10. The dynamically adjustable mode switcher based on phase change material diffraction neural network according to claim 8, characterized in that: The thickness of the device is the thickness from the upper surface gold film to the lower surface gold film.

Images