Method for On-Demand Design of Superatom Structures Based on Diffusion Model and Contrastive Learning

Abstract

The present invention discloses a method for on-demand design of superatom structures based on diffusion models and contrastive learning. According to the target spectral response given by the user, by introducing "multiple generation condition constraints" (connected domain constraint, spectral response transmittance coefficient constraint) in the latent space encoded by the diffusion model, a superatom structure library with high degrees of freedom that meets the condition constraints is reversely designed. Through contrastive learning, the generated superatom structure library is sorted according to similarity, and the optimal superatom structure information parameters that meet the target spectral response are designed on demand. The invention patent of the present invention can not only reversely design superatom structures with high degrees of freedom and solve the "one-to-many" problem in reverse design, but also design the optimal superatom structure that meets the target spectral response and is easy to process on demand, and solve the trade-off problem between optimal design and design result diversity. Compared with the traditional method of using numerical simulation for reverse design, it saves a large amount of on-demand design time and the design results are accurate.

Description

Technical Field:

[0001] The present invention belongs to the technical field of metamaterial surface structure design, and particularly relates to a method for designing metaatom structures on demand based on a diffusion model and contrastive learning. Background Art:

[0002] Metamaterials refer to non-natural, artificially synthesized composite materials, which are composed of specific spatial arrangement structures and have extraordinary physical properties that natural materials do not have. A metasurface is a two-dimensional (2D) version of a metamaterial, which is a planar device composed of subwavelength structures and is also called a metaatom. Currently, due to its unique properties, the metaatom structure has played a crucial role in application fields such as vibration and noise suppression, energy harvesting, stealth materials, and beam steering.

[0003] In order to prepare a metasurface structure with desired optical properties, it is necessary to accurately predict and design the spectral response transmission coefficient and geometry of the metaatom structure. The inverse design from the spectral response (SR) to the metaatom structure is a mapping problem from low dimension to high dimension, and the complex relationship between the two cannot be solved by establishing a simple mapping relationship through a generalized theory. Moreover, during the inverse design process, there is also a problem that different metaatom structures may produce very similar spectral responses ("one-to-many" mapping problem).

[0004] However, traditional design methods can only rely on empirical reasoning or experimental trial and error. Since this method involves a large number of full-wave numerical simulations (for example, the finite-difference time-domain method (FDTD) and the finite element method (FEM)), the design efficiency is low and the design results are usually not satisfactory. When deep neural networks (CNN, GAN, VAE) emerged, they were also applied to the design process of nanophotonic devices, but were only used to establish a one-to-one mapping relationship between structural parameters and spectral responses. Therefore, the "one-to-many" mapping problem in inverse design engineering cannot be solved, and at the same time, it is impossible to design on demand an optimal metaatom structure that meets multiple generation condition constraints (spectral response transmission coefficient constraint, connected domain constraint) and is easy to process and prepare.

[0005] Patent No. CN202310455940.9 discloses "a method for designing a multifunctional coded metasurface based on deep learning". This design method proposes a jointly trained conditional generative adversarial model, adding a neural network prediction model to the metasurface inverse design network, increasing the depth of the network model, and making the network have stronger generalization performance for the design of complex metasurfaces. However, this metasurface design method generates multiple metasurface designs and cannot select the optimal metasurface design suitable for the target electromagnetic response. In addition, no constraint conditions are added during the inverse design process. Therefore, it is impossible to design on demand a metasurface design that meets the generation condition constraints.

[0006] Patent No. CN202210839542.2 discloses "An Inverse Design Method for Metasurface Unit Structures Based on Improved Generative Adversarial Networks". This design method groups the dataset and inputs it into the generator and discriminator for training. By inputting the electromagnetic response, the corresponding metasurface unit structure can be obtained, thus reducing the trial-and-error time for designing metasurfaces. However, this design method utilizes a generative adversarial network, and the training and debugging process of this model is usually unstable and prone to mode collapse. In addition, this design process directly designs the meta-atom structure from the latent space through the GAN model, unable to control the design direction and easily falling into local optima.

[0007] Therefore, there is an urgent need to design an inverse design method that can design the optimal meta-atom structure that meets multiple generation condition constraints and is easy to process and prepare according to the target spectral response, save time costs, solve the "one-to-many" mapping problem, and control the design direction.

[0008] The information disclosed in this background art section is only intended to enhance the overall understanding of the present invention and should not be regarded as an admission or any form of implication that this information constitutes prior art already known to those of ordinary skill in the art. Summary of the Invention:

[0009] The purpose of the present invention is to provide an on-demand design method for meta-atom structures based on diffusion models and contrastive learning, which combines the excellent generation ability of the latent diffusion model and the similarity matching of the contrastive learning model, thereby overcoming the defects in the above-mentioned prior art.

[0010] To achieve the above purpose, the present invention provides an on-demand design method for meta-atom structures based on diffusion models and contrastive learning, and its steps are as follows:

[0011] S01. Determine the information parameters of the meta-atom structure, and use the meta-atom structure images in the existing dataset to train the encoder-decoder in the diffusion model. The encoder-decoder includes an encoder and a decoder, and it provides a low-dimensional latent space to train the diffusion model; after the encoder-decoder is trained once, it can be used for the diffusion model training without repeated training;

[0012] After the encoder-decoder of the diffusion model is trained, train the diffusion model; encode and compress the superatom structure image from the high-dimensional space to the low-dimensional space, i.e., the latent space, through the encoder-decoder of the diffusion model; use the encoded superatom structure image and the superatom structure parameters together as the input of the forward diffusion process of the diffusion model, and output a noise image after a process of continuously adding noise; then input the noise image and the spectral response transmittance coefficient into the backward diffusion process of the diffusion model, continuously denoise the noise image in the latent space, and at the same time introduce multiple generation condition constraints, and the multiple generation condition constraints include connected domain constraints and spectral response transmittance coefficient constraints; after continuously denoising the noise image through the backward diffusion process, generate superatom structure information parameters that meet the spectral response conditions and have the smallest connected domain, and the superatom structure information parameters include the superatom structure image and the superatom structure parameters; the decoder then decodes and restores the generated superatom structure image to the original high-dimensional space;

[0013] S03. Train the contrast learning model with the absorber dataset fabricated by FDTD software simulation. After the contrast learning model is trained, input the superatom structure information parameters and the spectral response transmittance coefficient in the existing dataset into the contrast learning model, calculate the similarity values between the superatom structure information parameters and the spectral response transmittance coefficient and sort them, and find the optimal superatom structure information parameters with the highest similarity to the spectral response transmittance coefficient;

[0014] S04. Given the spectral response transmittance coefficient, reverse design a superatom structure library that meets the spectral response transmittance coefficient conditions in the latent space through the diffusion model. The superatom structure library consists of multiple superatom structure information parameters, and then input the spectral response transmittance coefficient and the superatom structure library into the contrast learning model together; calculate the similarity values between the multiple superatom structure information parameters in the superatom structure library and the spectral response transmittance coefficient through the contrast learning model; sort the calculated similarity values from large to small, and select the superatom structure information parameters with the most matching similarity, so as to be able to design the optimal superatom structure information parameters that meet the spectral response transmittance coefficient conditions as required.

[0015] Preferably, in the technical solution, in step S01, the existing dataset is composed of the real part and the imaginary part of the superatom structure information parameters and their corresponding spectral response transmittance coefficients; the superatom structure information parameters are: 2D superatom structure image and three 1D superatom structure parameters.

[0016] Preferably, in the technical solution, in step S02, the encoder-decoder is also divided into an encoder and a decoder. The encoder of the diffusion model encodes and compresses the superatom structure image into a low-dimensional image, while the decoder restores the low-dimensional image into a superatom structure image. The loss function for training the encoder-decoder is the reconstruction term loss function loss re :

[0017]

[0018] Among them, Z is the input superatom structure image, is the reconstructed superatom structure image.

[0019] Preferably, in the technical solution, in step S02, the diffusion process of the diffusion model is divided into a forward diffusion process and a backward diffusion process; the forward diffusion process is to gradually add noise to the encoded superatom structure image z0 until it becomes a noise image z T at the T-th step; the noise addition formula in the forward diffusion process is as follows:

[0020]

[0021] where z0 represents the original superatom structure image, and z t represents the structure image obtained after adding noise at the t-th step, and α t is an intermediate variable, N is a normal distribution, and ∈ t is a Gaussian noise that follows N(0,1) at the time step t; t ∈ [1, T], and [1, T] is a series of fixed values; is the cumulative of the α t function from time step 1 to t;

[0022] The backward diffusion process is a process of gradually denoising the noise image z T until it is finally restored to the original superatom structure image z0; in addition, during the continuous denoising process of the noise image, a spectral response transmittance coefficient constraint and a connected component constraint are added; when inputting in the backward diffusion process, the spectral response transmittance coefficient is input as a generation condition to constrain the output result; when outputting in the backward diffusion process, the connected component value corresponding to the generated superatom structure image is calculated, and the superatom structure image with the smallest connected component value is output; the denoising sampling formula for the reverse diffusion process is:

[0023]

[0024] where z is a Gaussian noise that follows N(0,1), and ∈ t (z t , t) is the predicted noise, and β t is the variance of the Gaussian distribution; ∈ t (z t , t) is updated through the loss function of the diffusion model, and the loss function of the diffusion model is:

[0025]

[0026] Among them, ε(x) is the latent feature vector encoded in the latent space, is the Gaussian noise to be predicted, and E is the mean of the reverse diffusion process.

[0027] Preferably, in the technical solution, in step S03, during the training of the contrast learning model, the gradient backpropagation method is adopted to minimize the cross-entropy loss of the diagonal elements and optimize the weight parameters of the network; the loss function of the entire contrast learning model framework is expressed as:

[0028]

[0029] where x n is the feature of the nth spectral response, and y n is the information parameter of the nth meta-atom structure. Therefore, they form a pair. y i is an arbitrary information parameter of the meta-atom structure, batch size is the number of a batch, and τ is a learnable temperature parameter used to increase the tolerance of the contrast learning model to the most matching samples;

[0030] The contrast learning model includes a metamaterial structure encoder and a spectral response encoder; the training process of the contrast learning model is as follows: the meta-atom structure image and the meta-atom structure parameters first extract 64 feature maps through the convolutional layer; subsequently, these feature maps are converted into one-dimensional vectors through the Flatten layer; the one-dimensional vectors are passed to two consecutive network architectures, which are composed of a linear layer, a normalization layer, and a ReLU layer, where the linear layer is the Linear layer and the normalization layer is the BN layer; the metamaterial structure encoder connects the outputs of the two linear layers through a residual connection and then forms a structural feature vector with the encoding form of all×1024 after passing through the linear layer; the spectral response encoder passes through 3 convolutional layers, and each convolutional layer is followed by a normalization layer and a ReLU layer, connects the outputs of the two linear layers through a residual connection, and then forms a spectral response feature vector with the encoding form of 2×1024 after passing through a Flatten layer and a linear layer; finally, a matrix of batch size×batch size is obtained through the scaled similarity calculation, which is the similarity relationship between the spectral response feature vector and the structural feature vector; when the training loss value of the contrast learning model is less than the threshold or reaches the set maximum number of training times, the optimal weight of the contrast learning model training is output.

[0031] Preferably, in the technical solution, the formula for the similarity value between the spectral response feature vector and the structural feature vector is:

[0032]

[0033] Among them, A represents the spectral response feature vector, B represents the structural feature vector, and k represents the number of feature vectors. The closer the angle between the two vectors is to 0 and the closer the cosine value is to 1, the more similar the spectral response feature vector and the structural feature vector are.

[0034] Preferably, in the technical solution, in step S04, the process of the contrast learning model to find the optimal hyperatom structure information parameters in the hyperatom structure library is as follows: Given the spectral response transmission coefficient, the diffusion model that has completed training is used to reverse design the hyperatom structure library that meets the spectral response transmission coefficient; then, the contrast learning model inputs the hyperatom structure information parameters and the spectral response transmission coefficient in the hyperatom structure library into the metamaterial structure encoder and the spectral response encoder respectively for feature extraction and encoding operations; the contrast learning model outputs the structural feature vector with the encoded form of all×1024 and the spectral response feature vector of 2×1024 respectively; then, normalization operations are performed on the encoded structural feature vector and spectral response feature vector to eliminate the scale difference problem in the multi-dimensional data processing process; the scaling matrix multiplication is used to calculate the similarity value between the structural feature vector and the spectral response feature vector; a matrix of batch size×batch size is established through the scaling similarity calculation to establish the similarity relationship between the structural feature vector and the spectral response feature vector; after the similarity value calculation is completed, the similarity values are sorted from large to small; the optimal hyperatom structure information parameters that meet the spectral response transmission coefficient correspond to the maximum similarity value.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] Compared with the large amount of training data and repeated iterative calculations of traditional full numerical simulation methods (FEM, FDDT), the method of the present invention is carried out in the encoded latent space, greatly reducing the computational complexity and computational resources, accelerating the reverse design speed, and saving a large amount of manpower time and cost. In addition, the diffusion model performs reverse design in the latent space (low-dimensional space), and can learn more potential feature information in the low-dimensional space, making the designed hyperatom structure more diverse. At the same time, compared with the traditional generative models (GAN, VAE) that cannot control the direction of reverse design, the diffusion model decomposes the complex generation task into multiple easy-to-handle steps, indirectly mastering the direction of hyperatom structure design and avoiding the design from falling into local optimality. This reverse method can not only realize the reverse design of high-degree-of-freedom hyperatom structures, solve the "one-to-many" problem in reverse design, but also design the optimal hyperatom structure information parameters that meet the spectral response transmission coefficient and generation condition constraints (connected domain constraints), with a simple structure and easy to process. Description of the Drawings:

[0037] Figure 1 Flow chart of the on-demand design method for superatom structures based on diffusion models and contrast learning of the present invention;

[0038] Figure 2 Information parameter diagram of superatom structures composed of 2D structure images and three 1D structure parameters in the existing dataset of the present invention;

[0039] Figure 3 Structural framework diagram of the diffusion model training of the present invention;

[0040] Figure 4 Schematic diagram of the training error of the diffusion model of the present invention;

[0041] Figure 5 Structural framework diagram of the contrast learning model of the present invention;

[0042] Figure 6 Structural framework diagram of the input of structure images in the contrast learning model of the present invention

[0043] Figure 7 Structural framework diagram of the input of spectral response in the contrast learning model of the present invention

[0044] Figure 8 Schematic diagram of the training error of the contrast learning model of the present invention;

[0045] Figure 9 Structural framework diagram of the on-demand design method for superatom structures based on diffusion models and contrast learning of the present invention;

[0046] Figure 10 Schematic diagram of 100 structure images randomly generated by the diffusion model of the present invention;

[0047] Figure 11 Schematic diagram of the prediction results of the present invention's FDTD through the validation set in the dataset;

[0048] Figure 12 Information parameter diagram of the superatom structure reversely designed according to the target spectral response transmittance coefficient of the present invention;

[0049] Figure 13 Optimal superatom structure information parameter diagram calculated by the present invention based on contrast learning in the global superatom structure library designed by the diffusion model. Specific implementation method:

[0050] The following describes the specific implementation method of the present invention in detail, but it should be understood that the protection scope of the present invention is not limited by the specific implementation method.

[0051] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or variations thereof such as "comprises" or "comprising" shall be understood to include the stated element or component without excluding other elements or other components.

[0052] As Figure 1 shown, the method for on-demand design of superatom structures based on diffusion models and contrast learning comprises the following steps:

[0053] S01. Determine the superatom structure information parameters. The existing dataset consists of the superatom structure information parameters and the real and imaginary parts of the corresponding spectral response transmission coefficients. As Figure 2 shown, the superatom structure information parameters are: 2D superatom structure images and three 1D superatom structure parameters;

[0054] Use the superatom structure images in the existing dataset to train the encoder-decoder in the diffusion model. The encoder-decoder includes an encoder and a decoder, which provides a low-dimensional latent space for training the diffusion model. In addition, after the encoder-decoder is trained once, it can be used for diffusion model training without repeated training.

[0055] S02. The diffusion model is divided into two parts: the encoder-decoder and the diffusion model. Among them, the encoder-decoder is also divided into an encoder and a decoder, as Figure 3 shown; the encoder of the diffusion model encodes and compresses the superatom structure image into a low-dimensional image, while the decoder restores the low-dimensional image into a superatom structure image. The loss function for training the encoder-decoder is the reconstruction term loss function loss re :

[0056]

[0057] where Z is the input superatom structure image, is the reconstructed superatom structure image;

[0058] After the encoder-decoder of the diffusion model is trained, train the diffusion model; encode and compress the superatom structure image from the high-dimensional space to the low-dimensional space, i.e., the latent space, through the encoder-decoder of the diffusion model; use the encoded superatom structure image and the superatom structure parameters together as the input of the forward diffusion process of the diffusion model, and output a noisy image after a process of continuously adding noise; then input the noisy image and the spectral response transmittance coefficient into the backward diffusion process of the diffusion model, and the spectral response transmittance coefficient is used as a generation condition constraint to constrain the output result; continuously denoise the noisy image in the latent space, and at the same time, multiple generation condition constraints are introduced, including connected domain constraints and spectral response transmittance coefficient constraints; by calculating the connected domain value corresponding to the generated superatom structure image, output the superatom structure image with the smallest connected domain value, and optimize the superatom structure information parameters designed reversely. Among them, the role of the U-Net network model is to predict the noise added to the noisy image; after the process of continuously denoising the noisy image and updating the superatom structure image in the backward diffusion process, generate superatom structure information parameters that meet the spectral response conditions and have the smallest connected domain. The superatom structure information parameters include the superatom structure image and the superatom structure parameters; the decoder then decodes and restores the generated superatom structure image to the original high-dimensional space;

[0059] The diffusion process of the diffusion model is divided into a forward diffusion process and a backward diffusion process; the forward diffusion process is to gradually add noise to the encoded superatom structure image z0 until it becomes a noisy image z at the T-th step T The process; the noise addition formula in the forward diffusion process is as follows:

[0060]

[0061] Among them, z0 represents the original superatom structure image, and z t represents the structure image obtained after adding noise at the t-th step, and α t is an intermediate variable, N is a normal distribution, and ∈ t is a Gaussian noise that follows N(0,1) at the time step t; t ∈ [1, T], and [1, T] is a series of fixed values; is the cumulative of the α t function from time step 1 to t;

[0062] The backward diffusion process is to denoise the noisy image z TThe process of gradually denoising until finally restoring to the original superatom structure image z0; in addition, during the continuous denoising process of the noisy image, spectral response transmission coefficient constraints and connected domain constraints are added; when inputting in the backward diffusion process, the spectral response transmission coefficient is used as the generation condition input to constrain the output result; when outputting in the backward diffusion process, the connected domain value corresponding to the generated superatom structure image is calculated, and the superatom structure image with the smallest connected domain value is output; optimize the information parameters of the reverse-designed superatom structure so that the generated superatom structure information parameters can meet the spectral response transmission coefficient and its connected domain is also the smallest; the denoising sampling formula for the reverse diffusion process is:

[0063]

[0064] where z is Gaussian noise obeying N(0,1), ∈ t (z t ,t) is the predicted noise, and β t is the variance of the Gaussian distribution; ∈ t (z t ,t) is updated through the loss function of the diffusion model, and the loss function of the diffusion model is:

[0065]

[0066] where ε(x) is the latent feature vector encoded in the latent space, is the Gaussian noise to be predicted, and E is the mean of the backward diffusion process; as Figure 4 shown is the schematic diagram of the training error of the latent diffusion model.

[0067] S03. Train the contrastive learning model with the absorber dataset fabricated by FDTD software simulation, as Figure 5 、 6As shown in FIGS. 6 and 7, the contrast learning model includes a metamaterial structure encoder and a spectral response encoder. The training process of the contrast learning model is as follows: The metamaterial structure image and the metamaterial structure parameters first extract 64 feature maps through a convolutional layer; subsequently, these feature maps are converted into one-dimensional vectors through a Flatten layer; the one-dimensional vectors are passed into two consecutive network architectures, which are composed of a linear layer, a normalization layer, and a ReLU layer, where the linear layer is a Linear layer and the normalization layer is a BN layer; the metamaterial structure encoder connects the outputs of the two linear layers through a residual connection, and then forms a structural feature vector with an encoding form of all×1024 after passing through a linear layer; the spectral response encoder passes through 3 convolutional layers, each followed by a normalization layer and a ReLU layer, connects the outputs of the two linear layers through a residual connection, and then forms a spectral response feature vector with an encoding form of 2×1024 after passing through a Flatten layer and a linear layer; finally, a matrix of batch size×batch size is obtained through scaled similarity calculation, which is the similarity relationship between the spectral response feature vector and the structural feature vector; when the training loss value of the contrast learning model is less than the threshold or reaches the set maximum number of training times, the optimal weights of the contrast learning model training are output;

[0068] During the training process of the contrast learning model, the gradient backpropagation method is adopted to minimize the cross-entropy loss of the diagonal elements and optimize the weight parameters of the network; the loss function of the entire contrast learning model framework is expressed as:

[0069]

[0070] where x n is the feature of the nth spectral response transmittance coefficient, y n is the nth metamaterial structure information parameter, so they form a pair, y i is any metamaterial structure information parameter, batch size is the number of a batch, and τ is a learnable temperature parameter used to increase the tolerance of the contrast learning model to the most matching samples; Figure 8 is the schematic diagram of the training error of the contrast learning model;

[0071] After the contrast learning model is trained, the metamaterial structure information parameters and the spectral response transmittance coefficients in the existing dataset are input into the contrast learning model, the similarity values between the metamaterial structure information parameters and the spectral response transmittance coefficients are calculated and sorted, and the optimal metamaterial structure information parameters with the highest similarity to the spectral response transmittance coefficients are found;

[0072] The formula for the similarity value between the spectral response feature vector and the structural feature vector is:

[0073]

[0074] Among them, A represents the spectral response feature vector, B represents the structural feature vector, k represents the number of feature vectors. The closer the angle between the two vectors is to 0, the closer the cosine value is to 1, indicating that the spectral response feature vector and the structural feature vector are more similar.

[0075] S04. The process of the contrast learning model to find the optimal hyperatom structure information parameters in the hyperatom structure library is as follows: As Figure 9 shown, given the spectral response transmission coefficient, the diffusion model completed by training is used to inversely design the hyperatom structure library that meets the spectral response transmission coefficient; among them, in the process of continuously denoising the noisy image during the backward diffusion process, the role of the U-Net network model is to predict the noise added to the noisy image; in addition, the generated hyperatom structure image is constrained by the connected domain condition restriction to design a hyperatom structure library with the smallest connected domain value, which is easy to be manufactured manually; then, the hyperatom structure information parameters and the spectral response transmission coefficient in the hyperatom structure library are respectively input into the metamaterial structure encoder and the spectral response encoder by the contrast learning model for feature extraction and encoding operations; the contrast learning model outputs a structural feature vector with the encoded form of all×1024 and a spectral response feature vector of 2×1024 respectively; then, the normalized operation is performed on the encoded structural feature vector and spectral response feature vector to eliminate the scale difference problem in the multi-dimensional data processing process; the scaling matrix multiplication is used to calculate the similarity value between the structural feature vector and the spectral response feature vector; a matrix of batch size×batch size is obtained through the scaling similarity calculation to establish the similarity relationship between the structural feature vector and the spectral response feature vector; after the similarity value calculation is completed, the similarity values are sorted from large to small; the optimal hyperatom structure information parameters that meet the spectral response correspond to the largest similarity value; finally, the generated hyperatom structure information parameters are reconstructed into a three-dimensional hyperatom structure by using the FDTD software, so as to simulate the corresponding electric field distribution of the structure, and the phase and amplitude corresponding to the hyperatom structure are calculated by inputting the electric field distribution.

[0076] As Figure 10 shown are 100 structure images randomly generated by the latent diffusion model, Figure 10 (a) is the structure image in the original dataset, Figure 10 (b) is the structure image restored by the latent diffusion model. It can be seen from the figure that the generated hyperatom structure images with high degrees of freedom are almost the same.

[0077] To verify the accuracy of the FDTD software in predicting the spectral response transmission coefficient, we randomly selected 25 hyperatom structure information parameters, such as Figure 11As shown, the spectral response transmittance coefficients predicted by the randomly selected structures are almost the same as the target spectral response transmittance coefficients. Therefore, the accuracy of FDTD in predicting spectral response transmittance coefficients can be demonstrated.

[0078] In Figure 12 , to verify the accuracy of the inverse design of this method, we reverse-designed the superatom structure image and superatom structure parameters given the target spectral response transmittance coefficient, and then predicted the spectral response transmittance coefficient through FDTD. As Figure 12 shown, the predicted spectral response transmittance coefficient is the same as the target spectral response transmittance coefficient, thus demonstrating the accuracy of the inverse design method for the superatom structure information parameters.

[0079] As Figure 13 shown is to verify whether this inverse design method can solve the "one-to-many" mapping problem in the inverse design process. Sort the calculated similarity values from largest to smallest, and select the one with the highest similarity value as the optimal superatom structure information parameter. As Figure 13 shown on the right are the spectral response transmittance coefficient and the superatom structure information parameter with the highest similarity value. As Figure 13 shown on the left is the similarity value ranking of each superatom structure information parameter.

[0080] The foregoing description of specific exemplary embodiments of the invention has been presented for purposes of illustration and exemplification. These descriptions are not intended to limit the invention to the precise forms disclosed, and it is obvious that many changes and variations are possible in light of the above teaching. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the invention and its practical applications, so that those skilled in the art can implement and utilize the various different exemplary embodiments of the invention, as well as various different selections and changes. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for designing super-atom structures on demand based on diffusion model and contrastive learning, the steps of which are as follows: S01. Determine super-atomic structure information parameters, and use super-atomic structure images in existing data sets to train the codec in the diffusion model. The codec includes an encoder and a decoder, which provides a low-dimensional latent space to train the diffusion model. S02. After the encoder and decoder of the diffusion model are trained, the diffusion model is trained; the super-atom structure image is encoded and compressed from a high-dimensional space to a low-dimensional space, i.e., a latent space, through the encoder and decoder of the diffusion model; The encoded super-atom structure image and super-atom structure parameters are used as the input of the forward diffusion process of the diffusion model, and the noise image is output after the process of continuously adding noise; Then, the noise image and the spectral response transmission coefficient are input together into the backward diffusion process of the diffusion model, and the noise image is continuously denoised in the latent space. At the same time, multiple generation condition constraints are introduced, and the multiple generation condition constraints include connected domain constraints and spectral response transmission coefficient constraints. After the noise image is continuously denoised in the backward diffusion process, super-atomic structure information parameters that meet the spectral response conditions and have the smallest connected domain are generated. The super-atomic structure information parameters include super-atomic structure images and super-atomic structure parameters. The decoder then decodes the generated super-atom structure image and restores it to the original high-dimensional space; S03, training the comparative learning model with the absorber data set simulated by FDTD software. After the comparative learning model is trained, the super-atom structure information parameters and the spectral response transmission coefficient in the existing data set are input into the comparative learning model, and the similarity values ​​between the super-atom structure information parameters and the spectral response transmission coefficient are calculated and sorted to find the optimal super-atom structure information parameters with the highest similarity to the spectral response transmission coefficient; S04. Given a spectral response transmission coefficient, a super-atom structure library satisfying the spectral response transmission coefficient condition is reversely designed in the latent space through a diffusion model. The super-atom structure library is composed of multiple super-atom structure information parameters. The spectral response transmission coefficient and the super-atom structure library are then input into the contrastive learning model together. The similarity values ​​between multiple superatomic structure information parameters and spectral response transmission coefficients in the superatomic structure library are calculated through a comparative learning model; the calculated similarity values ​​are sorted from large to small, and the superatomic structure information parameters with the most similarity are selected, so that the optimal superatomic structure information parameters that meet the spectral response transmission coefficient conditions can be designed as needed.

2. The method for designing super-atom structures on demand based on diffusion model and contrastive learning according to claim 1, characterized in that: In step S01, the existing data set is composed of super-atomic structure information parameters and the real and imaginary parts of the corresponding spectral response transmission coefficients; the super-atomic structure information parameters are: a 2D super-atomic structure image and three 1D super-atomic structure parameters.

3. The method for designing super-atom structures on demand based on diffusion model and contrastive learning according to claim 1, characterized in that: In step S02, the codec is also divided into an encoder and a decoder. The encoder of the diffusion model encodes and compresses the super-atomic structure image into a low-dimensional image, while the decoder restores the low-dimensional image into a super-atomic structure image. The loss function of the codec training is the reconstruction loss function loss re : Among them, Z is the input super-atomic structure image, This is the reconstructed superatomic structure image.

4. The method for designing super-atom structures on demand based on diffusion model and contrastive learning according to claim 1, characterized in that: In step S02, the diffusion process of the diffusion model is divided into a forward diffusion process and a backward diffusion process; the forward diffusion process is to gradually add noise to the encoded super-atom structure image z0 until the Tth step turns into a noise image z T The process of adding noise in the forward diffusion process is as follows: Among them, z0 represents the original super-atom structure image, z t represents the structural image obtained after adding noise in step t, α t is the intermediate variable, N is the normal distribution, ∈ t is Gaussian noise that obeys N(0,1) at time step t; t∈[1,T], [1,T] is a series of fixed values; is α t The accumulation of the function from time step 1 to t; The back diffusion process is to perform a backscattering of the noise image z T The process of gradually denoising until the original super-atomic structure image z0 is finally restored; in addition, the spectral response transmission coefficient constraint and the connected domain constraint are added to the noise image during the continuous denoising process; when the backward diffusion process is input, the spectral response transmission coefficient is input as the generation condition to constrain the output result; when the backward diffusion process is output, the connected domain value corresponding to the generated super-atomic structure image is calculated, and the super-atomic structure image with the smallest connected domain value is output; the super-atomic structure information parameters designed inversely are optimized so that the generated super-atomic structure information parameters can meet the spectral response transmission coefficient and its connected domain is also the smallest; the denoising sampling formula of the reverse diffusion process is: Where z is Gaussian noise subject to N(0,1), ∈ t (z t ,t) is the prediction noise, β t is the variance of the Gaussian distribution; ∈ t (z t ,t) is updated through the loss function of the diffusion model. The loss function of the diffusion model is: Among them, ε(x) is the potential feature vector encoded in the latent space, is the Gaussian noise to be predicted, and E is the mean of the back diffusion process.

5. The method for designing super-atom structures on demand based on diffusion model and contrastive learning according to claim 1, characterized in that: In step S03, during the training process, the contrastive learning model uses the gradient back propagation method to minimize the cross entropy loss of the diagonal elements and optimize the weight parameters of the network; the loss function of the entire contrastive learning model framework is expressed as: Among them, x n is the characteristic of the nth spectral response transmission coefficient, y n is the nth superatomic structure information parameter, so they form a pair, y i is any superatomic structure information parameter, batch size is the number of batches, and τ is a learnable temperature parameter used to increase the tolerance of the contrastive learning model to the best matching sample; The contrastive learning model includes a metamaterial structure encoder and a spectral response encoder. The training process of the contrastive learning model is as follows: the metaatomic structure image and the metaatomic structure parameters are first extracted into 64 feature maps through a convolutional layer. Subsequently, these feature maps are converted into one-dimensional vectors through a Flatten layer. The one-dimensional vectors are passed to two consecutive network architectures, which consist of a linear layer, a normalization layer, and a ReLU layer, where the linear layer is a Linear layer and the normalization layer is a BN layer. The metamaterial structure encoder connects the outputs of the two linear layers through a residual connection, and then forms a structural feature vector with an encoding form of all×1024 after passing through the linear layer. The spectral response encoder passes through three convolutional layers, each of which is followed by a normalization layer and a ReLU layer, and connects the outputs of the two linear layers through a residual connection, and then forms a spectral response feature vector with an encoding form of 2×1024 after passing through a Flatten layer and a linear layer. Finally, the batch size×batch is obtained by scaling the similarity calculation. size, which is the similarity relationship between the spectral response eigenvector and the structural eigenvector. When the contrastive learning model training loss value is less than the threshold or reaches the set maximum number of training times, the optimal weights for contrastive learning model training are output.

6. The method for designing super-atom structures on demand based on diffusion model and contrastive learning according to claim 1, characterized in that: In step S04, the process of the contrastive learning model searching for the optimal super-atomic structure information parameters in the super-atomic structure library is as follows: given the spectral response transmission coefficient, the super-atomic structure library that satisfies the spectral response transmission coefficient is reversely designed through the trained diffusion model; the super-atomic structure information parameters and the spectral response transmission coefficient in the super-atomic structure library are respectively input into the metamaterial structure encoder and the spectral response encoder through the contrastive learning model to perform feature extraction and encoding operations; the contrastive learning model outputs a structural feature vector encoded in the form of all×1024 and a spectral response feature vector of 2×1024 respectively; then, the encoded structural feature vector and spectral response feature vector are normalized to eliminate the scale difference problem in the multi-dimensional data processing process; scaling matrix multiplication is used to calculate the similarity value between the structural feature vector and the spectral response feature vector; a batch size×batch size matrix is ​​obtained by scaling the similarity calculation, and a similarity relationship between the structural feature vector and the spectral response feature vector is established; after the similarity value calculation is completed, the similarity values ​​are sorted from large to small; the optimal super-atomic structure information parameter that satisfies the spectral response transmission coefficient corresponds to the maximum similarity value.

Images