CN114692500A - Deep learning-based complex amplitude type super-surface design method, system and device - Google Patents
Deep learning-based complex amplitude type super-surface design method, system and device Download PDFInfo
- Publication number
- CN114692500A CN114692500A CN202210334747.5A CN202210334747A CN114692500A CN 114692500 A CN114692500 A CN 114692500A CN 202210334747 A CN202210334747 A CN 202210334747A CN 114692500 A CN114692500 A CN 114692500A
- Authority
- CN
- China
- Prior art keywords
- model
- complex amplitude
- training
- target
- reverse
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a complex amplitude type super surface design method, a system and a device based on deep learning, which comprises the following steps of firstly determining the structure of a super unit and regulating and controlling the degree of freedom of the structure; acquiring complex amplitude response data sets of superunits with different structural parameters; collecting target complex amplitude response data sets generated randomly, and dividing the two data sets into a training set, a verification set and a test set respectively; constructing a deep learning model; predicting the complex amplitude response of the superunit according to the structural parameters by using a forward prediction model, and predicting the structural parameters according to the target response by using a reverse design model; then training by utilizing the training set and the verification set to obtain a reverse design model; and designing the complex amplitude type super-surface device by using the trained reverse design model, generating a target amplitude and a target phase according to the target device, and generating a full-mode structure of the device in the reverse design model. The design method provided by the invention reduces the difficulty of designing the complex amplitude type super surface and simultaneously ensures the processing feasibility of the super surface.
Description
Technical Field
The invention relates to the technical field of optical holography, in particular to a complex amplitude type super surface design method, a system and a device based on deep learning.
Background
The super surface is composed of a group of sub-wavelength structure arrays, and the amplitude, the phase and the polarization of electromagnetic waves can be controlled at will by adjusting the geometric dimension and the orientation of each unit structure. In the traditional super-surface design process, a designer needs to make physical or empirical inference on the optical response of a super-surface unit structure, and usually a continuous trial and error optimization process is carried out, so that the designer needs to have rich professional knowledge, and a large amount of optimization and selection are included in the process of selecting a proper unit structure, so that the whole process is time-consuming and labor-consuming.
To address this problem, deep learning has been introduced in the past few years into the design of hypersurfaces because of the ability to mine mutual representations with different levels of data. At present, the super-surface design based on deep learning is usually pure amplitude response or pure phase response, and the complex amplitude type super-surface (which simultaneously regulates and controls the amplitude and the phase) can realize more flexible electromagnetic regulation and control, and needs to be combined with the deep learning to carry out deep research. At present, researches have been made on arbitrary modulation of amplitude and phase by using a variable for controlling a multi-layered superunit, which is a bifocal lens and a polarization multiplexing lens designed by using a generated antagonistic network, with a fixed height of a three-dimensional unit structure and the remaining two-dimensional plane as a design variable.
Although the superunit of the two-dimensional model can realize arbitrary regulation and control of amplitude and phase, when the design wave band extends from the mid-infrared wave band to visible light, the two-dimensional model can bring some extremely small defects, or for the multilayer structure superunit of the single-layer structure, the two can bring greater difficulty for processing, and the practical development of the super surface is not facilitated.
Disclosure of Invention
In view of this, the present invention aims to provide a method, a system and a device for designing a complex amplitude type super-surface based on deep learning, in which the method determines a full-mode structure of a complex amplitude device by using a deep learning model, thereby reducing the difficulty of the design process and improving the precision of the model design.
In order to achieve the purpose, the invention provides the following technical scheme:
the invention provides a complex amplitude type super-surface design method based on deep learning, which comprises the following steps:
determining the structure of the superunit and regulating and controlling the degree of freedom of the structure;
acquiring complex amplitude response data sets of superunits with different structural parameters; collecting target complex amplitude response data sets generated randomly, and dividing the two data sets into a training set, a verification set and a test set respectively;
constructing a deep learning model; the deep learning model comprises a forward prediction model and a reverse design model, the forward prediction model is used for predicting the complex amplitude response of the superunit according to the structural parameters, and the reverse design model is used for predicting the structural parameters according to the target response;
training according to the collected training set and the verification set, wherein the training is carried out in two steps, a forward prediction model is trained by using a complex amplitude response data set, the forward prediction model is loaded into the whole model after the training is finished, the training of the forward model is frozen, the whole network is trained by using a target complex amplitude data set at the moment, only a reverse design model is trained in the process, and the reverse design model is derived from the stored whole model after the training is finished;
and designing the complex amplitude type super-surface device by using the trained reverse design model, generating a target amplitude and a target phase according to the target device, and generating a full-mode structure of the device in the reverse design model.
Further, after the training of the reverse design model is finished, the method further comprises the following steps:
and judging whether the difference between the predicted response and the real complex amplitude response of the structural parameters designed by the reverse design model meets a preset threshold value, if so, ending the training process of the reverse design model, and if not, continuing the training until the condition is met.
Further, the method also comprises a model test process, wherein the model test process comprises the following steps:
verifying whether the precision meets the requirement by using a test data set, generating a corresponding predicted complex amplitude response in a forward prediction model by using the structural parameters of a complex amplitude response data test set, subtracting the predicted value from the actual response, averaging the differences, and judging whether the error meets a preset condition; if the condition is met, carrying out the next reverse model training, if the condition is not met, adjusting the hyper-parameters in the forward model, and re-training the forward model until the precision meets the requirement;
further, the model test process further comprises the steps of:
generating prediction of super-unit structure parameters in a reverse design model by using a target complex amplitude data test set, calculating an actual response according to a predicted value, carrying out difference on the target complex amplitude and the actual response, averaging the difference, and judging whether the error meets a preset condition; if the condition is met, the model can be used for carrying out the next device design, if the condition is not met, the hyper-parameters in the reverse model are adjusted, and the reverse model is trained again until the precision meets the requirement.
Further, the full-mold structure is carried out according to the following steps:
the method comprises the steps of setting a target device as a complex amplitude multiplexing device for displaying a nano printing image in a near field and a holographic image in a Fraunhofer area, setting a target nano printing image and a target holographic image, determining the target amplitude according to the nano printing image, calculating a target phase by using a GS algorithm, and finally inputting the determined target complex amplitude into a reverse design model to generate a predicted full-mode structure.
Further, the method also comprises the following steps:
and performing full-mode simulation on the full-mode structure by adopting a finite difference time domain method, using boundary conditions of perfect matching layers in the x direction, the y direction and the z direction, adopting levorotatory circular polarized light incidence, obtaining a simulation result, and extracting and displaying right-rotatory circular polarized light.
The invention provides a complex amplitude type super surface design system based on deep learning, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor comprises:
the parameter setting unit is used for determining the structure of the superunit and regulating and controlling the degree of freedom of the structure;
the data acquisition unit is used for acquiring complex amplitude response data sets of the superunits with different structural parameters; collecting target complex amplitude response data sets generated randomly, and dividing the two data sets into a training set, a verification set and a test set respectively;
the model building unit is used for building a deep learning model; the deep learning model comprises a forward prediction model and a reverse design model, the forward prediction model is used for predicting the complex amplitude response of the superunit according to the structural parameters, and the reverse design model is used for predicting the structural parameters according to the target response;
the model training unit is used for training according to the collected training set and the collected verification set, the training is carried out in two steps, a forward prediction model is firstly trained by using a complex amplitude response data set, the forward prediction model is loaded into the whole model after the training is finished, the training of the forward model is frozen, the whole network is trained by using a target complex amplitude data set at the moment, only a reverse design model is trained in the process, and the reverse design model is derived from the stored whole model after the training is finished;
and the device full-mode generating unit is used for designing the complex amplitude type super-surface device by using the trained reverse design model, generating target amplitude and phase according to the target device and generating a full-mode structure of the device in the reverse design model.
Further, the model training unit is provided with a training control unit and a model testing unit, and the training control unit performs the following steps:
and judging whether the difference between the predicted response and the real complex amplitude response of the structural parameters designed by the reverse design model meets a preset threshold value, if so, ending the training process of the reverse design model, and if not, continuing the training until the condition is met.
The model test unit is carried out according to the following steps:
verifying whether the precision meets the requirement by using a test data set, generating a corresponding predicted complex amplitude response in a forward prediction model by using the structural parameters of a complex amplitude response data test set, subtracting the predicted value from the actual response, averaging the differences, and judging whether the error meets a preset condition; if the condition is met, carrying out the next reverse model training, if the condition is not met, adjusting the hyper-parameters in the forward model, and re-training the forward model until the precision meets the requirement;
generating prediction of super-unit structure parameters in a reverse design model by using a target complex amplitude data test set, calculating an actual response according to a predicted value, carrying out difference on the target complex amplitude and the actual response, averaging the difference, and judging whether the error meets a preset condition; if the condition is met, the model can be used for carrying out the next device design, if the condition is not met, the hyper-parameters in the reverse model are adjusted, and the reverse model is trained again until the precision meets the requirement.
The complex amplitude type super surface provided by the invention is a super surface device manufactured according to the complex amplitude type super surface design method based on deep learning.
The invention has the beneficial effects that:
the invention provides a complex amplitude type super surface design method, a system and a device based on deep learning. The random regulation and control of the amplitude and phase response of the superunit are realized, and designers do not need to physically deduce the response of the superunit with different structures according to the learned principle, so that a large number of optimization and selection processes are omitted, and the threshold of designing the complex amplitude type supersurface is greatly reduced; in addition, because the used deep learning model is of a full-connection layer structure, the model is simple to build and the training is easy to converge, so that in the process of building the model and the training model, a designer can realize high-precision training without deeply knowing the relevant principle of deep learning; the trained model can realize the prediction of the structural parameters very quickly according to the target device, so the method can be expanded to the design of more different complex amplitude type devices; finally, since the cell structure used is a relatively smooth single-layer structure, this ensures sufficient processing feasibility.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objectives and other advantages of the invention may be realized and attained by the means of the instrumentalities and combinations particularly pointed out hereinafter.
Drawings
In order to make the purpose, technical scheme and beneficial effect of the invention more clear, the invention provides the following drawings for explanation:
FIG. 1 is a super cell structure.
FIG. 2 is a deep learning model architecture.
FIG. 3 is a forward predictive model training process.
FIG. 4 is a process of inverse design model training.
FIG. 5 is a design target diagram and a full-mold simulation effect diagram.
Detailed Description
The present invention is further described with reference to the following drawings and specific examples so that those skilled in the art can better understand the present invention and can practice the present invention, but the examples are not intended to limit the present invention.
Example 1
The complex amplitude type super surface design method based on deep learning provided by the embodiment has the following basic principle: the super-surface is composed of a group of sub-wavelength structure arrays, and the amplitude, the phase and the polarization of electromagnetic waves can be arbitrarily controlled by adjusting the geometric dimension of each unit structure. The deep learning is formed by a neural network simulating human brain to carry out analysis learning, the mutual expression forms of data with different levels can be mined, the structural parameters of the super unit forming the super surface and the corresponding optical response of the super unit are two data sets with complex nonlinear relation, the implicit relation of the two data sets can be mined by introducing the deep learning, and the design of related super surface devices can be carried out according to the relation.
The complex amplitude type super surface design method based on deep learning provided by the embodiment specifically comprises the following steps:
1. determining the structure of the superunit and the degree of freedom of the structure which can be regulated and controlled, then acquiring complex amplitude response data sets of the superunit with different structural parameters by using CST Microwave Studio software, collecting randomly generated target complex amplitude response data sets, and dividing the two data sets into a training set, a verification set and a test set respectively.
2. The method is characterized in that a preset deep learning model is built by using Tensorflow and Keras frames, the whole model consists of two parts, a forward prediction model for predicting the complex amplitude response of the superunit according to the structural parameters and a reverse design model for predicting the structural parameters according to the target response can be used, and compared with a single network, the model can solve the problem that multiple solutions possibly occur in the reverse design process and the prediction precision is not high.
3. Training according to a collected training set and a collected verification set, wherein the training is carried out in two steps, a forward prediction model is trained by using a complex amplitude response data set, the forward prediction model is loaded into the whole model after the training is finished, the training of the forward model is frozen, the whole network is trained by using a target complex amplitude data set at the moment, only a reverse design model is trained in the process, and after the training is finished, if the prediction response of the structural parameters designed by the reverse design model is close to the real complex amplitude response of the structural parameters, namely the target complex amplitude is almost equal to the actual complex amplitude response of a superunit, the prediction of the superunit structural parameters can be realized by using the reverse design model at the moment according to the target complex amplitude. And finally, deriving the reverse design model from the stored whole model.
4. Verifying whether the precision meets the requirement by using a test data set, generating corresponding predicted complex amplitude response in a forward prediction model by 1000 structural parameters of a complex amplitude response data test set, carrying out difference between the predicted value and the actual response, averaging the difference, judging whether the error meets preset conditions (the average amplitude deviation is less than 0.01, and the average phase deviation is less than 5 degrees), if so, carrying out the next reverse model training, if not, adjusting the hyper-parameter in the forward model, and re-training the forward model until the precision meets the requirement; the prediction of the super-unit structure parameters is generated in a reverse design model by utilizing a target complex amplitude data test set, the actual response is calculated by a predicted value in CST Microwave Studio software, the target complex amplitude and the actual response are subjected to difference and the average value of the difference is calculated, whether the error meets the preset conditions (the average amplitude deviation is less than 0.05, and the average phase deviation is less than 5 degrees) or not is judged, if the condition is met, the model can be used for carrying out the next device design, if the condition is not met, the super-unit structure parameters in the reverse model are adjusted, and the reverse model is trained again until the precision meets the requirement.
5. Designing a complex amplitude type super-surface device by using a trained reverse design model, generating a target amplitude and a target phase according to a target device, and generating a full-mode structure of the device in the reverse design model;
6. and carrying out full-mode simulation on the full-mode structure, and extracting and displaying the final effect from the simulation result.
Example 2
In this embodiment, a method for constructing a complex amplitude type super-surface design based on deep learning includes the following specific steps:
the structure of the superunit and its controllable structural freedom are determined, as shown in FIG. 1, the superunit is composed of lower silicon dioxide (SiO)2) Substrate and upper layer (TiO)2) The cross-shaped structure is formed by two cuboids which are vertical to each other and have the length and the width of Lx, W, Ly and W respectively, and the whole cross-shaped structure is rotated by theta degrees. Then, a CST Microwave Studio software is used to collect 21000 sets of complex amplitude response data sets (left-handed circular polarized light with the wavelength of 532nm is incident from the substrate of the super-unit and collected) of the super-unit with different structural parameters { Lx, Ly, theta }Response results of emergent right-handed circularly polarized light), and 21000 groups of randomly generated target complex amplitude response data sets are collected, and the two data sets are respectively divided into 16000 groups as training sets, 4000 groups as verification sets and 1000 groups as test sets;
a preset deep learning model is built by using a Tensorflow and Keras framework, as shown in FIG. 2, the whole model consists of two parts, namely a forward prediction model capable of predicting the complex amplitude response of a superunit according to structural parameters and a reverse design model capable of predicting the structural parameters according to target responses, as shown in FIG. 2, Dense (i) represents a fully-connected layer, i represents the number of nerve units of the layer, and Dense (400) represents a fully-connected layer with 400 nerve units;
the BatcNormal layer can adjust the distribution of the activation values of each layer to enable the activation values to have proper extent, so that the training speed of the model is increased, and overfitting (the phenomenon that the training loss is reduced but the verification loss is not reduced or even increased) can be inhibited to a certain extent;
the Dropout layer can randomly delete neurons in the training process, so that overfitting can be effectively inhibited, and Dropout (0.25) means that in the training process, the nerve units in the previous layer are randomly deleted according to the proportion of 0.25 and then are transmitted to the next layer;
input (2) represents an Input layer of a reverse design model, Input two dimensions are amplitude and phase response, Input (3) represents an Input layer of a forward prediction model, and 3-dimensional structure parameters are Input;
the predicted structure parameters comprise { Lx, Ly, theta };
wherein the input target complex amplitude response represents a target amplitude and a target phase of the input set light field;
the output prediction complex amplitude response represents that after a reverse design model predicts a structural parameter according to a certain target complex amplitude response, a predicted value of the structural parameter outputs the predicted complex amplitude response after being input into a forward prediction model;
the two networks except the input layer and the output layer are composed of 5 layers of fully connected layers (Dense) with the number of nerve units of 400, 1000, 2500, 1000 and 400 respectively, and in order to inhibit overfitting, a reverse design model and a forward design model are respectively adoptedThe prediction model uses BatchNormal and Dropout strategies, the forward model uses relu function, the reverse design model uses sigmoid function, the whole training process uses average absolute error as loss function, and the used learning rate and the number of batches are respectively 10-6And 100;
training according to the collected training set and the verification set, wherein the training is carried out in two steps, a forward prediction model is trained by utilizing a complex amplitude response data set, and the training process is shown in FIG. 3; after the training is finished, loading the model into the whole model, freezing the training of the forward model, training the whole network by using the target complex amplitude data set, wherein only the reverse design model is trained in the process, and the training process is shown in fig. 4. After the training is finished, the predicted response of the structural parameters designed by the reverse design model is close to the real complex amplitude response of the structural parameters, namely the target complex amplitude is almost equal to the actual complex amplitude response of the superunit, which means that the reverse design model can realize the prediction of the superunit structural parameters according to the target complex amplitude. Finally, deriving a reverse design model from the stored whole model;
verifying whether the precision meets the requirement by using a test data set, generating corresponding predicted complex amplitude response in a forward prediction model by using 1000 structural parameters of a complex amplitude response data test set, subtracting the predicted value from the actual response, and calculating the average value of the difference, wherein the average amplitude prediction deviation is calculated to be 0.002, the average phase prediction deviation is calculated to be 3.11 degrees, and the precision meets the requirement; generating prediction of super-unit structure parameters in a reverse design model by utilizing a target complex amplitude data test set, calculating actual response of a predicted value in CST software, carrying out difference between the target complex amplitude and the actual response, and calculating the average value of the difference, wherein the calculated average amplitude prediction deviation is 0.01, the average phase prediction deviation is 3.28 degrees, and the precision meets the requirement;
the method comprises the steps of designing a complex amplitude type super-surface device by using a trained reverse design model, setting a target device as a complex amplitude multiplexing device capable of displaying a nano printing image in a near field and a holographic image in a Fraunhofer area, setting a target nano printing image and a target holographic image as (a) and (b) in fig. 5 respectively, determining target amplitude according to the nano printing image, calculating a target phase by using a GS algorithm (the GS algorithm provided in the embodiment is the prior art, and the specific calculation process is not repeated here), and finally inputting the determined target complex amplitude into the reverse design model to generate a predicted full-mode structure.
And (d) performing full-mode simulation on the full-mode structure by adopting a finite difference time domain method (FDTD), using boundary conditions of perfect matching layers in the x direction, the y direction and the z direction, adopting levorotatory circular polarized light incidence, extracting dextrorotatory circular polarized light from a simulation result and displaying a final effect, wherein the displayed nano printing gray scale image shows a good visual effect and the holographic image shows high resolution as shown in (c) and (d) of fig. 5.
The above-mentioned embodiments are merely preferred embodiments for fully illustrating the present invention, and the scope of the present invention is not limited thereto. The equivalent substitution or change made by the technical personnel in the technical field on the basis of the invention is all within the protection scope of the invention. The protection scope of the invention is subject to the claims.
Claims (9)
1. The method for designing the complex amplitude type super surface based on deep learning is characterized by comprising the following steps of: the method comprises the following steps:
determining the structure of the superunit and regulating and controlling the degree of freedom of the structure;
acquiring complex amplitude response data sets of superunits with different structural parameters; collecting target complex amplitude response data sets generated randomly, and dividing the two data sets into a training set, a verification set and a test set respectively;
constructing a deep learning model; the deep learning model comprises a forward prediction model and a reverse design model, the forward prediction model is used for predicting the complex amplitude response of the superunit according to the structural parameters, and the reverse design model is used for predicting the structural parameters according to the target response;
training according to the collected training set and the verification set, wherein the training is carried out in two steps, a forward prediction model is trained by using a complex amplitude response data set, the forward prediction model is loaded into the whole model after the training is finished, the training of the forward model is frozen, the whole network is trained by using a target complex amplitude data set at the moment, only a reverse design model is trained in the process, and the reverse design model is derived from the stored whole model after the training is finished;
and designing the complex amplitude type super-surface device by using the trained reverse design model, generating a target amplitude and a target phase according to the target device, and generating a full-mode structure of the device in the reverse design model.
2. The method of claim 1, wherein: after the training of the reverse design model is finished, the method further comprises the following steps:
and judging whether the difference between the predicted response and the real complex amplitude response of the structural parameters designed by the reverse design model meets a preset threshold value, if so, ending the training process of the reverse design model, and if not, continuing the training until the condition is met.
3. The method of claim 1, wherein: the method also comprises a model test process, wherein the model test process comprises the following steps:
verifying whether the precision meets the requirement by using a test data set, generating a corresponding predicted complex amplitude response in a forward prediction model by using the structural parameters of a complex amplitude response data test set, subtracting the predicted value from the actual response, averaging the differences, and judging whether the error meets a preset condition; and if the condition is met, carrying out next reverse model training, if the condition is not met, adjusting the hyper-parameters in the forward model, and re-training the forward model until the precision meets the requirement.
4. The method of claim 3, wherein: the model test procedure further comprises the steps of:
generating prediction of super-unit structure parameters in a reverse design model by using a target complex amplitude data test set, calculating an actual response according to a predicted value, carrying out difference on the target complex amplitude and the actual response, averaging the difference, and judging whether the error meets a preset condition; and if the condition is met, using the model to carry out the next device design, if the condition is not met, adjusting the hyper-parameters in the reverse model, and re-training the reverse model until the precision meets the requirement.
5. The method of claim 1, wherein: the full-mold structure is carried out according to the following steps:
the method comprises the steps of setting a target device as a complex amplitude multiplexing device for displaying a nano printing image in a near field and a holographic image in a Fraunhofer area, setting a target nano printing image and a target holographic image, determining the target amplitude according to the nano printing image, calculating a target phase by using a GS algorithm, and finally inputting the determined target complex amplitude into a reverse design model to generate a predicted full-mode structure.
6. The method of claim 1, wherein: further comprising the steps of:
and performing full-mode simulation on the full-mode structure by adopting a finite difference time domain method, using boundary conditions of perfect matching layers in the x direction, the y direction and the z direction, adopting levorotatory circular polarized light incidence, obtaining a simulation result, and extracting and displaying right-rotatory circular polarized light.
7. A complex amplitude-type super-surface design system based on deep learning, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor comprises:
the parameter setting unit is used for determining the structure of the superunit and regulating and controlling the degree of freedom of the structure;
the data acquisition unit is used for acquiring complex amplitude response data sets of the superunits with different structural parameters; collecting target complex amplitude response data sets generated randomly, and dividing the two data sets into a training set, a verification set and a test set respectively;
the model building unit is used for building a deep learning model; the deep learning model comprises a forward prediction model and a reverse design model, the forward prediction model is used for predicting the complex amplitude response of the superunit according to the structural parameters, and the reverse design model is used for predicting the structural parameters according to the target response;
the model training unit is used for training according to the collected training set and the collected verification set, the training is carried out in two steps, a forward prediction model is firstly trained by using a complex amplitude response data set, the forward prediction model is loaded into the whole model after the training is finished, the training of the forward model is frozen, the whole network is trained by using a target complex amplitude data set at the moment, only a reverse design model is trained in the process, and the reverse design model is derived from the stored whole model after the training is finished;
and the device full-mode generating unit is used for designing the complex amplitude type super-surface device by using the trained reverse design model, generating target amplitude and phase according to the target device and generating a full-mode structure of the device in the reverse design model.
8. The method of claim 7, wherein: the model training unit is provided with a training control unit and a model testing unit, and the training control unit performs the following steps:
and judging whether the difference between the predicted response and the real complex amplitude response of the structural parameters designed by the reverse design model meets a preset threshold value, if so, ending the training process of the reverse design model, and if not, continuing the training until the condition is met.
The model test unit is carried out according to the following steps:
verifying whether the precision meets the requirement by using a test data set, generating a corresponding predicted complex amplitude response in a forward prediction model by using the structural parameters of a complex amplitude response data test set, subtracting the predicted value from the actual response, averaging the differences, and judging whether the error meets a preset condition; if the condition is met, carrying out the next reverse model training, if the condition is not met, adjusting the hyper-parameters in the forward model, and re-training the forward model until the precision meets the requirement;
generating prediction of super-unit structure parameters in a reverse design model by using a target complex amplitude data test set, calculating an actual response according to a predicted value, carrying out difference on the target complex amplitude and the actual response, averaging the difference, and judging whether the error meets a preset condition; if the condition is met, the model can be used for carrying out the next device design, if the condition is not met, the hyper-parameters in the reverse model are adjusted, and the reverse model is trained again until the precision meets the requirement.
9. The super surface of complex amplitude type, its characterized in that: the complex amplitude type super surface is a super surface device manufactured according to the complex amplitude type super surface design method based on deep learning of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210334747.5A CN114692500A (en) | 2022-03-31 | 2022-03-31 | Deep learning-based complex amplitude type super-surface design method, system and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210334747.5A CN114692500A (en) | 2022-03-31 | 2022-03-31 | Deep learning-based complex amplitude type super-surface design method, system and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114692500A true CN114692500A (en) | 2022-07-01 |
Family
ID=82140192
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210334747.5A Pending CN114692500A (en) | 2022-03-31 | 2022-03-31 | Deep learning-based complex amplitude type super-surface design method, system and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114692500A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116205111A (en) * | 2023-03-09 | 2023-06-02 | 北京理工大学 | Multi-dimensional multi-channel multiplexing super-surface holographic optimization method based on reverse design |
-
2022
- 2022-03-31 CN CN202210334747.5A patent/CN114692500A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116205111A (en) * | 2023-03-09 | 2023-06-02 | 北京理工大学 | Multi-dimensional multi-channel multiplexing super-surface holographic optimization method based on reverse design |
| CN116205111B (en) * | 2023-03-09 | 2024-01-16 | 北京理工大学 | Multi-dimensional multi-channel multiplexing super-surface holographic optimization method based on reverse design |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tercan et al. | Transfer-learning: Bridging the gap between real and simulation data for machine learning in injection molding | |
| US20200342301A1 (en) | Convolutional neural network on-chip learning system based on non-volatile memory | |
| US20180114110A1 (en) | Method and apparatus to reduce neural network | |
| JPH02297117A (en) | Nerve system simulation system | |
| CN114692500A (en) | Deep learning-based complex amplitude type super-surface design method, system and device | |
| JP6851634B2 (en) | Feature conversion module, pattern identification device, pattern identification method, and program | |
| Sullivan et al. | Homeostatic synaptic scaling in self-organizing maps | |
| Anochi et al. | Self-configuring two types of neural networks by mpca | |
| Tarkhov et al. | Neural network modelling methods for creating digital twins of real objects | |
| CN113705031B (en) | Nano antenna array electromagnetic performance prediction method based on deep learning | |
| Ribeiro et al. | Artificial neural networks applied in mechanical structural design | |
| Tasdemır et al. | The effect on the wind power performance of different normalization methods by using multilayer feed-forward backpropagation neural network | |
| CN111542818B (en) | Network model data access method and device and electronic equipment | |
| CN112861332B (en) | Cluster dynamics prediction method based on graph network | |
| Altundogan et al. | Genetic algorithm based fuzzy cognitive map concept relationship determination and sigmoid configuration | |
| Zhu et al. | Integration of vergence, cyclovergence, and saccades through active efficient coding | |
| Baran et al. | Cellular automata with irregular structure: a compact representation | |
| CN117610630A (en) | Super-surface optical neural network based on random inactivation and parameter design and configuration method | |
| Cheng et al. | Research on Calibration Method of Camera Stabilizer Based on Binocular Vision | |
| Catrina et al. | Learning about Growing Neural Cellular Automata | |
| Duan et al. | Ship hull surface reconstruction from scattered points cloud using an RBF neural network mapping technology | |
| Pastorello | Quantum neural networks | |
| Wang et al. | Design Elements Extraction from Point-Cloud Dataset Using Deep Enhancement Learning | |
| Zhang et al. | Cellular GPU model for structured mesh generation and its application to the stereo-matching disparity map | |
| 장성제 | Tree-shaped support structure generation for 3D printing by deep reinforcement learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |