Quantum neural network-based periodic target electromagnetic scattering characteristic rapid simulation method

Abstract

The invention discloses a periodic target electromagnetic scattering characteristic rapid simulation method based on a quantum neural network, and the method comprises the following steps: selecting finite periodic arrays with different structure gaps and array scales in a certain range, and generating a data set through employing a sub-global basis function method; selecting a classic neural network technology, designing a quantum neural network unit, and building a hybrid quantum neural network model; training and testing the hybrid quantum neural network model, and storing parameters in the model in an optimal state; and generating a sub-global basis function by using a sub-global basis function method, predicting an expansion coefficient of each unit in the periodic structure by using the trained mixed quantum neural network model, and finally calculating to obtain surface current distribution of the whole target periodic structure. According to the method, the hybrid neural network technology is utilized to avoid consuming a lot of time to generate the reduced impedance matrix, the consumed time can be effectively reduced while the accuracy is ensured, and the efficiency advantage brought by quantum equipment in the future is expected to be obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic simulation, and particularly relates to a fast simulation method for electromagnetic scattering characteristics of periodic targets based on a quantum neural network. Background Art

[0002] Methods for accurately and effectively solving the scattering problem of large-scale finite periodic structures have been widely studied for many years. Different from infinite periodic structures, edge effects need to be considered in large-scale finite periodic structures. At present, commercial software such as HFSS, CST, and FEKO based on full-wave simulation has become the main tool for analyzing scattering problems of targets. However, for large-scale finite periodic structures, it is very difficult for a computer to simulate scattering characteristics using the full-wave method. The sub-domain basis function method is an effective method to alleviate this problem, which greatly reduces the number of unknowns. However, in analyzing large-scale finite periodic structures containing a large number of elements, due to considering the mutual coupling between all elements, it is still time-consuming to establish a simplified matrix equation based on these methods. Especially for periodic structures with different parameters, the entire process of the global domain basis function method must be recalculated, resulting in low efficiency in using the traditional sub-domain basis function method to solve practical engineering applications. Therefore, it is necessary to combine other new technologies to improve the efficiency of the sub-domain basis function method in analyzing large-scale finite periodic structures. Summary of the Invention

[0003] The purpose of the present invention is to provide a fast simulation method for electromagnetic scattering characteristics of periodic targets based on a quantum neural network, so as to achieve accurate and efficient analysis of the electromagnetic scattering characteristics of the target periodic array and solve the technical problems mentioned in the background art.

[0004] To solve the above technical problems, the specific technical solution of the present invention is as follows:

[0005] A fast simulation method for electromagnetic scattering characteristics of periodic targets based on a quantum neural network, which consists of the following steps:

[0006] Step 1, generate a data set: Select finite periodic arrays with different structural gaps and array scales within a certain range, and use the sub-domain basis function method to generate a data set.

[0007] Step 2, build a hybrid quantum neural network model: Select classical neural network technology, design quantum neural network units, and build a hybrid quantum neural network model;

[0008] Step 3, train a prediction model: Train and test the hybrid quantum neural network model, and save the structure and parameters in the model that reaches the optimal state;

[0009] Step 4, Analyze the scattering characteristics of the target periodic structure: Use the sub-domain basis function method to generate sub-domain basis functions, use the trained hybrid quantum neural network model to predict the expansion coefficients of each unit in the periodic structure, and finally calculate the surface current distribution of the entire target periodic structure.

[0010] Further, generating the dataset based on the sub-domain basis function method in Step 1 includes the following steps:

[0011] Step 1, Generate a dataset: Select finite periodic arrays with different array scales and structural gaps within a certain range, use the sub-domain basis function method to generate data, divide the data to form a dataset, and preprocess the dataset.

[0012] Step 1.1, Select finite periodic arrays with different structures: The selection range of the array scale of the finite periodic structure is from 6 to 50, and the selection range of the periodic gap is from 0.6 to 1.2 wavelengths. The unit of the selected finite periodic structure array is a pure metal square with a side length of 0.5 wavelengths, and the incident wave used is a plane wave with a frequency of 300 MHz;

[0013] Step 1.2, Generate the features and labels of the dataset: Use the sub-domain basis function method to calculate the selected finite periodic structure array, and obtain the expansion coefficients of the sub-domain basis functions of each cell. This coefficient includes two parts: the real part and the imaginary part. The array scale, periodic gap, and cell coordinates of the finite periodic structure are used as the feature part of the dataset, and the real and imaginary part coefficients of the calculated sub-domain basis functions are used as the label part of the dataset;

[0014] Step 1.3, Preprocess the dataset: Split the data obtained in Step 1.2 into corner, edge, and internal cell data according to the position of the cells in the finite periodic structure, and then three groups of datasets can be obtained. Each group of datasets contains the real part and the imaginary part. Perform Z-Score normalization on the feature part of the dataset so that the values of the feature data are normally distributed between (-1, 1).

[0015] Further, in Step 2, build three groups of hybrid quantum neural network models corresponding to corner, edge, and internal cells. Each hybrid quantum neural network model consists of neurons and quantum modules, and includes three parts: feature dimension augmentation, quantum neural network, and output. Specifically, it includes the following steps:

[0016] Step 2.1, Feature dimension augmentation part: It consists of the input layer and hidden layer of a classical neural network. The input layer transmits the preprocessed 6-dimensional feature information to the subsequent hidden layer. The first hidden layer contains 32 neurons with ReLU activation functions, and the second hidden layer contains 28 neurons with Tanh activation functions. The feature dimension augmentation part can non-linearly map low-dimensional information to high-dimensional space, thus realizing the expansion of the input data;

[0017] Step 2.2, Quantum Neural Network Part: It consists of three parts: quantum encoding, quantum circuit, and quantum measurement. The quantum encoding method is amplitude encoding, which encodes the classical information input by neurons onto five qubits; the quantum circuit is composed of parametric quantum gates RY gates and non-parametric quantum gates CNOTs to adjust the quantum state of the quantum unit to the optimal; the quantum measurement measures the quantum state of the entire quantum unit, thereby representing the probability of collapsing the quantum state to the ground state;

[0018] Step 2.3, Output Part: It consists of only one neuron without an activation function, which linearly superimposes the measurement values of the quantum neural network unit as expansion coefficients;

[0019] Step 2.4, Forming a Prediction Model System: Steps 2.1, 2.2, and 2.3 together form a hybrid quantum neural network model. Since the periodic structure consists of corner, edge, and internal units, three groups of hybrid quantum neural network models are built, and each group consists of a real-part and an imaginary-part hybrid quantum neural network model.

[0020] Step 2.5, Parameter Setting: According to Steps 2.1 - 2.4, set the initial learning rate of the hybrid quantum to 0.01, and use the Adam optimizer to dynamically adjust the learning rate; the batch size of the model is set to 64; the maximum number of iterations of the model is set to 5000 rounds, and then make appropriate adjustments according to the training effect.

[0021] Furthermore, the hybrid quantum neural network model described in Step 3 includes the following steps:

[0022] Step 3.1, Generating Training Set, Test Set, and Validation Set: Randomly sample 80%, 10%, and 10% of the dataset samples as the training set, validation set, and test set. The training set is used to train the neuron parameters and quantum gate parameters in the hybrid quantum neural network model, the validation set is used to tune the hyperparameter configuration and conduct a preliminary evaluation of the model, and the test set is used to verify the effectiveness and generalization of the model;

[0023] Step 3.2, model training: Initialize the parameters in the three sets of hybrid quantum neural network models, use the training set and validation set to train and tune the three sets of hybrid quantum neural network models; use the iterable data loader to read 64 samples in the training set in random order each time into the neural network for training, calculate the training loss value, and update the neuron weights and bias terms and quantum gate parameters through the backpropagation algorithm. The training of all samples is considered one round; the training parameters of the current round are fixed to calculate the validation set loss value. After each round of training, the neural network model corresponding to the minimum loss on the validation set is saved; after several rounds of training, observe the trend and value of the loss values of the training set and validation set, and modify the model or hyperparameters until the loss values of the training set and validation set simultaneously reach the set threshold range and the loss value of the validation set satisfies the requirement of first decreasing and then increasing. At this time, the saved minimum validation set loss model is the trained neural network model.

[0024] Step 3.3, save the model and parameters: Save the architecture and optimal parameters of the three sets of hybrid quantum neural network models, and use the test set to verify and predict the effects of the three sets of hybrid quantum neural network models.

[0025] Furthermore, step 4 uses the sub-global basis function method to generate 9 sub-global basis functions, uses the trained hybrid quantum neural network model to predict the expansion coefficients of the sub-global basis functions of the target periodic array, and calculates the induced surface current distribution of the entire target periodic array, which specifically includes the following steps:

[0026] Step 4.1, generate sub-global basis functions: Given a finite periodic structure array to be predicted, extract a 3×3 sub-global basis function subarray. The sub-global basis function can be formulated as:

[0027]

[0028] Where M is the number of triangular mesh edges on a cell in the target periodic array, and f nm (r) is the RWG basis function of the mth edge on the nth cell, I nm is f nm (r) expansion coefficient, r is the position vector of the observation point on the periodic array;

[0029] Step 4.2, predict the expansion coefficient of the sub-global basis function: the array size, structural gap and cell coordinates of the target periodic array are standardized and input into the trained hybrid quantum neural network model to predict the expansion coefficient of each cell in the target periodic array;

[0030] Step 4.3: The electromagnetic scattering induced current distribution of the periodic array to be predicted is as follows:

[0031]

[0032] Where r is the position vector of the observation point on the periodic array, J(r) is the surface current distribution of the array; a n is the expansion coefficient of the sub-global basis function; is the subglobal basis function at the nth element in the array.

[0033] The present invention provides a method for rapidly simulating the electromagnetic scattering characteristics of a periodic target based on a quantum neural network, which has the following advantages: the present invention mainly uses a hybrid quantum neural network to learn the distribution of sub-global basis functions on periodic structures of different array scales and unit spacings, thereby obtaining the current distribution of the periodic structure and rapidly predicting the scattering characteristics of the periodic structure. By training and calculating different periodic structures within a selected range and training a neural network model, the scattering characteristics of the periodic structure under different feed distributions can be predicted. The neural network can also be used to extremely quickly predict the expansion coefficient of the sub-global product function, greatly reducing computational costs and improving solution efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a method for rapid simulation of electromagnetic scattering characteristics of periodic targets based on quantum neural networks of the present invention;

[0035] Figure 2 Generate a schematic diagram for the dataset of the present invention;

[0036] Figure 3 The hybrid quantum neural network model group of the present invention;

[0037] Figure 4 It is a hybrid quantum neural network model architecture;

[0038] Figure 5 This is the 20×20 square patch periodic structure of the present invention;

[0039] Figure 6 This is the 70×70 square patch periodic structure of the present invention;

[0040] Figure 7 This is the 100×100 square ring patch periodic structure of the present invention;

[0041] Figure 8 This is a diagram showing the radar cross section prediction results of a dual-base station with a 20×20 square patch periodic structure according to the present invention;

[0042] Figure 9 This is a diagram showing the radar cross section prediction results of a dual-base station with a 70×70 square patch periodic structure according to the present invention;

[0043] Figure 10 This is a diagram showing the radar cross section prediction results of the dual-base station with a 100×100 square ring patch periodic structure according to the present invention. DETAILED DESCRIPTION

[0044] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of a method for rapid simulation of electromagnetic scattering characteristics of periodic targets based on quantum neural networks in conjunction with the accompanying drawings.

[0045] The present invention proposes a method for rapid simulation of electromagnetic scattering characteristics of periodic targets based on quantum neural networks, such as Figure 1 As shown, the following steps are included:

[0046] Step 1: Generate a data set: Select finite period arrays with different array sizes and structural gaps within a certain range, generate data using the sub-global basis function method, divide the data into a data set, and preprocess the data set.

[0047] Step 1.1. Select finite period arrays of different structures: the array size of the finite period structure can be selected from 6 to 50, and the period interval can be selected from 0.6 to 1.2 wavelengths. The units of the selected finite period array are uniformly 0.5 wavelength pure metal squares, and the incident wave used is a 300 MHz plane wave;

[0048] Step 1.2, generate the features and labels of the data set: use the sub-global basis function method to calculate the selected finite periodic structure array, and obtain the expansion coefficient of the sub-global basis function of each unit cell, which includes two parts: real part and imaginary part. Figure 2 As shown, the array size, periodic gap and cell coordinates of the finite periodic structure are used as the feature part of the data set, and the real and imaginary coefficients of the calculated sub-global basis function are used as the label part of the data set;

[0049] Step 1.3. Preprocess the dataset: Split the data obtained in Step 1.2 into corner, edge, and interior cell data based on the cell locations within the finite periodic structure. This yields three datasets, each containing real and imaginary components. Perform Z-score normalization on the characteristic components of the dataset, ensuring that the characteristic data values are normally distributed between (-1, 1).

[0050] Step 2: Build three sets of hybrid quantum neural network models corresponding to corners, edges, and internal cells. Each hybrid quantum neural network model consists of neurons and quantum modules, and includes three parts: feature dimension increase, quantum neural network, and output. Figure 3 As shown, the architecture of a single hybrid quantum neural network model is as follows Figure 4 As shown:

[0051] Step 2.1, Feature Dimension Expansion Part: It consists of the input layer and the hidden layer of a classical neural network. The input layer transmits the preprocessed 6-dimensional feature information to the subsequent hidden layer. The first hidden layer contains 32 neurons with ReLU activation functions, and the second hidden layer contains 28 neurons with Tanh activation functions. The feature dimension expansion part can non-linearly map low-dimensional information to a high-dimensional space, thereby realizing the expansion of the input data;

[0052] Step 2.2, Quantum Neural Network Part: It consists of three parts: quantum encoding, quantum circuit, and quantum measurement. The quantum encoding method is amplitude encoding, which encodes the classical information input by neurons onto five qubits; the quantum circuit is composed of parameterized quantum gates RY gates and non-parameterized quantum gates CNOTs to adjust the quantum state of the quantum unit to the optimal; the quantum measurement measures the quantum state of the entire quantum unit, thereby characterizing the probability of collapsing the quantum state to the ground state;

[0053] Step 2.3, Output Part: It consists of only one neuron without an activation function, which linearly superimposes the measurement values of the quantum neural network unit as the expansion coefficient;

[0054] Step 2.4, Forming a Prediction Model System: Steps 2.1, 2.2, and 2.3 together form a hybrid quantum neural network model. Since the periodic structure consists of corner, edge, and internal units, three groups of hybrid quantum neural network models are built, and each group consists of a real part and an imaginary part hybrid quantum neural network model.

[0055] Step 2.5, Parameter Setting: According to Steps 2.1 - 2.4, set the initial learning rate of the hybrid quantum to 0.01, and use the Adam optimizer to dynamically adjust the learning rate; the batch size of the model is set to 64; the maximum number of iterations of the model is set to 5000 rounds, and then make appropriate adjustments according to the training effect.

[0056] Step 3, Training the Prediction Model, includes the following steps:

[0057] Step 3.1, Generating a Training Set, a Test Set, and a Validation Set: Randomly sample 80%, 10%, and 10% of the dataset samples as the training set, validation set, and test set. The training set is used to train the neuron parameters and quantum gate parameters in the hybrid quantum neural network model, the validation set is used to optimize the hyperparameter configuration and conduct a preliminary evaluation of the model, and the test set is used to verify the effectiveness and generalization of the model;

[0058] Step 3.2, Model Training: Initialize the parameters in the three groups of hybrid quantum neural network models, and use the training set and validation set to train and optimize the three groups of hybrid quantum neural network models; through the iterable data loader, randomly read 64 samples from the training set into the neural network for training each time, calculate the training loss value, and update the weights and bias terms of the neurons and the quantum gate parameters through the backpropagation algorithm. After all samples are trained, it is considered as one round; fix the training parameters of the current round and calculate the validation set loss value. After each round of training, save the neural network model corresponding to the minimum loss on the validation set; after several rounds of training, observe the trends and values of the training set and validation set loss values, modify the model or hyperparameters until the loss values of the training set and validation set simultaneously reach the set threshold range and the loss value of the validation set satisfies the condition of first decreasing and then increasing. At this time, the neural network model with the minimum validation set loss saved is the trained neural network model.

[0059] Step 3.3, Save the Model and Parameters: Save the architectures and optimal parameters of the three groups of hybrid quantum neural network models, and use the test set to verify and predict the effects of the three groups of hybrid quantum neural network models.

[0060] Step 4: Use the sub-domain basis function method to generate 9 sub-domain basis functions, and use the trained hybrid quantum neural network model to predict the expansion coefficients of the sub-domain basis functions of the target periodic array, and calculate the induced surface current distribution of the entire target periodic array, which specifically includes the following steps:

[0061] Step 4.1, Generate Sub-domain Basis Functions: Given the finite periodic structure array to be predicted, extract the 3×3 sub-domain basis function sub-array. The sub-domain basis function can be formulated as:

[0062]

[0063] where M is the number of edges of the triangular mesh divided on a cell in the target periodic array, f nm (r) is the RWG basis function of the m-th edge on the n-th cell, I nm is the expansion coefficient of f nm (r), and r is the position vector of the observation point on the periodic array;

[0064] Step 4.2, Predict the Expansion Coefficients of Sub-domain Basis Functions: Normalize the array scale, structure gap, and cell coordinates of the target periodic array and input them into the trained hybrid quantum neural network model to predict the expansion coefficients of each cell in the target periodic array respectively;

[0065] Step 4.3, The electromagnetic scattering induced current distribution of the periodic array to be predicted is as follows:

[0066]

[0067] Where r is the position vector of the observation point on the periodic array, J(r) is the surface current distribution of the array; a n is the expansion coefficient of the sub-global basis function; is the subglobal basis function at the nth element in the array.

[0068] In order to verify the accuracy and efficiency of the present invention, three pure metal patch periodic arrays are taken as examples below. The calculation examples are calculated on the computing platform of Intel(R) Core(TM) i9-14900KF 3.20GHz Nvidia 4090GPU.

[0069] like Figure 5 As shown in the figure, a periodic array of pure metal square pieces with an array size of 20×20 and a side length of 0.5λ, the incident plane wave frequency is 300MHz, and the center spacing of the unit cells on the periodic array is d x =d y =0.55λ.

[0070] like Figure 6 As shown in the figure, a periodic array of pure metal square pieces with an array size of 70×70 and a side length of 0.5λ, the incident plane wave frequency is 300MHz, and the center spacing of the unit cells on the periodic array is d x =d y =0.8λ.

[0071] like Figure 7 As shown, a pure metal square ring periodic array with an array size of 100×100, a side length of 0.5λ and an internal side length of 0.2λ, the incident plane wave frequency is 300MHz, and the center spacing of the unit cells on the periodic array is d x =d y =0.85λ.

[0072] Figure 8 、 Figure 9 and Figure 10 The dual-base radar cross sections (RSCs) for the three target arrays calculated using the sub-global basis function method and the method of the present invention are presented, demonstrating consistent results between the two methods. This is because the trained hybrid quantum neural network model exhibits excellent generalization performance, and changes in the array size, structural gap, or cell type of the periodic structure within a certain range do not significantly affect the model's prediction accuracy.

[0073] Table 1 presents the time consumed by the novel sub-universal basis function method and the method of the present invention in the calculation example. It can be seen that, compared with the novel sub-universal basis function method, the time consumption of the method of the present invention is significantly reduced, thus proving the high efficiency of the method of the present invention. In addition, the hybrid quantum neural network model proposed by the present invention is expected to be deployed on noisy intermediate-scale quantum devices in the near future to further obtain quantum-level acceleration advantages.

[0074] Table 1 Time consumption of the sub-universal basis function method and the method of the present invention for the calculation example

[0075]

[0076] It can be understood that the present invention is described by some embodiments. Those skilled in the art know that, without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. Additionally, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A fast simulation method for the electromagnetic scattering characteristics of periodic targets based on a quantum neural network, characterized in that, The following steps are involved: Step 1: Generate a data set: Select finite periodic arrays with different structural gaps and array sizes within a certain range, and use the sub-global basis function method to generate a data set; Step 2: Build a hybrid quantum neural network model: Select classical neural network technology, design quantum neural network units, and build a hybrid quantum neural network model; Step 3: Train the prediction model: Train and test the hybrid quantum neural network model, and save the structure and parameters of the optimal model. Step 4. Analyze the scattering characteristics of the target periodic structure: Use the sub-global basis function method to generate the sub-global basis function, use the trained hybrid quantum neural network model to predict the expansion coefficient of each unit in the periodic structure, and finally calculate the surface current distribution of the entire target periodic structure.

2. The rapid simulation method for the electromagnetic scattering characteristics of periodic targets based on a quantum neural network according to claim 1, wherein In step 1, the dataset is generated based on the sub-global basis function method, which includes the following steps: Step 1.

1. Select finite period arrays of different structures: the array size of the finite period structure can be selected from 6 to 50, and the period interval can be selected from 0.6 to 1.2 wavelengths. The units of the selected finite period arrays are uniformly pure metal squares with a wavelength of 0.5 wavelength, and the incident wave used is a 300 MHz plane wave. Step 1.

2. Generate dataset features and labels: Use the subglobal basis function method to calculate the selected finite periodic structure array and obtain the expansion coefficient of the subglobal basis function for each cell. This coefficient consists of two parts: the real part and the imaginary part. The array size, period gap, and cell coordinates of the finite periodic structure are used as the dataset features, and the calculated real and imaginary coefficients of the subglobal basis function are used as the dataset labels. Step 1.

3. Preprocess the data set: Split the data obtained in step 1.2 into corner, edge, and interior cell data according to the location of the cells in the finite periodic structure. This will yield three data sets, each containing real and imaginary parts. Perform Z-Score normalization on the characteristic parts of the data sets so that the characteristic data values are normally distributed between (-1, 1).

3. A fast simulation method for the electromagnetic scattering characteristics of periodic targets based on a quantum neural network according to claim 1, characterized in that The construction of the hybrid quantum neural network model described in step 2 includes the following steps: Step 2.

1. Feature Dimension Enhancement: This consists of the input layer and hidden layers of a classic neural network. The input layer transmits the preprocessed 6-dimensional feature information to the subsequent hidden layers. The first hidden layer contains 32 neurons with ReLU activation function, and the second hidden layer contains 28 neurons with Tanh activation function. The feature dimension enhancement part can nonlinearly map low-dimensional information to high-dimensional space, thereby expanding the input data. Step 2.2, quantum neural network: It consists of three parts: quantum coding, quantum circuit, and quantum measurement. The quantum coding method is amplitude coding, which encodes the classical information transmitted by the neuron into five quantum bits. The quantum circuit is composed of a parameterized quantum gate RY gate and a parameterless quantum gate CNOT, which adjusts the quantum state of the quantum unit to the optimal state. Quantum measurement measures the quantum state of the entire quantum unit, thereby representing the probability of the quantum state collapsing to the ground state. Step 2.3, Output part: It consists of only one neuron without an activation function, which linearly superimposes the measurement values of the quantum neural network unit into expansion coefficients; Step 2.4, Forming a prediction model system: Steps 2.1, 2.2, and 2.3 together form a hybrid quantum neural network model; The periodic structure consists of corner, edge, and internal units, and three groups of hybrid quantum neural network models are built, with each group consisting of a real-part and an imaginary-part hybrid quantum neural network model; Step 2.5, Set the learning rate, batch size of the model, and maximum number of iterations of the model for the hybrid quantum neural network according to Steps 2.1 - 2.4, and then make appropriate adjustments according to the training effect.

4. A fast simulation method for the electromagnetic scattering characteristics of periodic targets based on a quantum neural network according to claim 1, characterized in that, The hybrid quantum neural network model described in Step 3 includes the following steps: Step 3.1, Generate a training set, a test set, and a validation set: Randomly sample 80%, 10%, and 10% of the dataset samples as the training set, validation set, and test set respectively; The training set is used to train the neuron parameters and quantum gate parameters in the hybrid quantum neural network model, the validation set is used to tune the hyperparameter configuration and conduct a preliminary evaluation of the model, and the test set is used to verify the effectiveness and generalization of the model; Step 3.2, Model training: Initialize the parameters in the three groups of hybrid quantum neural network models, and use the training set and validation set to train and tune the three groups of hybrid quantum neural network models; Through an iterable data loader, randomly read 64 samples from the training set into the neural network for training each time, calculate the training loss value, and update the weights and biases of the neurons as well as the quantum gate parameters through the backpropagation algorithm. One round is completed after all samples are trained; Fix the training parameters of the current round and calculate the validation set loss value. After each round of training, save the neural network model corresponding to the minimum loss on the validation set; After several rounds of training, observe the trends and values of the training set and validation set loss values, and modify the model or hyperparameters until the loss values of both the training set and validation set reach the set threshold range and the loss value of the validation set satisfies the condition of first decreasing and then increasing. At this time, the neural network model with the minimum validation set loss saved is the trained neural network model; Step 3.3, Save the model and parameters: Save the architectures and optimal parameters of the three groups of hybrid quantum neural network models, and use the test set to verify and predict the effects of the three groups of hybrid quantum neural network models.

5. A fast simulation method for the electromagnetic scattering characteristics of periodic targets based on a quantum neural network according to claim 1, characterized in that The specific steps of Step 4 include the following: Step 4.1, Generate sub-universal basis functions: Given a finite periodic structure array to be predicted, extract a 3×3 sub-universal basis function sub-array. The sub-universal basis function can be formulated as: where M is the number of edges of the triangular mesh dissected on a cell in the target periodic array, f nm (r) is the RWG basis function of the m-th edge on the n-th cell, I nm is the expansion coefficient of f nm (r), and r is the position vector of the observation point on the periodic array; Step 4.2, Predict the expansion coefficients of the sub-universal basis functions: Normalize the array scale, structure gap, and cell coordinates of the target periodic array and input them into the trained hybrid quantum neural network model to predict the expansion coefficients of each cell in the target periodic array respectively; The electromagnetic scattering induced current distribution of the periodic array to be predicted is as follows: Where r is the position vector of the observation point on the periodic array, J(r) is the surface current distribution of the array; a n is the expansion coefficient of the sub-global basis function; is the subglobal basis function at the nth element in the array.

Images