A Wide-Range Modeling Method and System for Microwave Devices Based on Adjoint Neural Networks

Abstract

The present invention provides a method and system for large-range modeling of microwave devices based on adjoint neural networks, which obtain the ranges of multiple device parameters of a microwave device, divide the ranges of the device parameters into multiple sub-ranges with overlapping intervals, construct sub-models based on the multiple sub-ranges of the multiple device parameters, and each sub-model is correspondingly provided with an adjoint neural network; pre-train each sub-model based on training data and perform auxiliary training through the adjoint neural network corresponding to each sub-model; construct sigmoid functions corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval; determine the correction function of each sub-model based on the sub-range of each device parameter corresponding to each sub-model; construct all the sub-models provided with the correction function into a large-range model, output the electromagnetic frequency-domain characteristic values corresponding to the device parameter group to be calculated based on the large-range model, and output the derivative information corresponding to the device parameter group through the adjoint neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of microwave device design, and in particular, to a method and system for large-range modeling of microwave devices based on adjoint neural networks. Background Art

[0002] With the continuous development of technologies such as the new generation of communication networks and the Internet of Things, various new requirements and new applications have emerged continuously. It is crucial to develop radio frequency (RF) microwave devices and circuits with high precision, low design cost, and stable performance to meet these new requirements and new applications. To achieve the above design requirements, RF microwave designers increasingly rely on computer-aided design (CAD) for circuit simulation and verification. RF microwave modeling and simulation have become two essential steps in the design of RF microwave devices and systems. However, when using existing 3D simulation software for electromagnetic (EM) design, since it is necessary to continuously repeat the adjustment of geometric parameter values for calculation to find appropriate values, its computational cost is very high and very time-consuming. Establishing an accurate and reliable RF microwave device model can effectively accelerate the design speed of high-frequency and high-performance RF microwave circuits, shorten the design cycle, and meet the development needs of current RF microwave technologies.

[0003] Among the existing technologies for quickly obtaining accurate microwave device surrogate models, artificial neural networks (ANNs) are recognized as a powerful technology for parametric modeling and design optimization of microwave devices. ANNs can effectively learn the input-output relationships of electromagnetic devices, and the trained neural network model can accurately and quickly predict the electromagnetic responses of microwave devices with geometric parameters as variables. Various technologies have now been introduced to develop and improve the parametric models of microwave devices. For example, the adjoint neural network model (SAANN) uses sensitivity analysis to reduce the amount of training data required for development and predict the derivatives of the output response with respect to each geometric parameter, and the generalization of this model is high. For the design of microwave devices with geometric variables varying in a large range, the existing technologies cannot solve the problem. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for large-range modeling of microwave devices based on adjoint neural networks to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for large-range modeling of microwave devices based on adjoint neural networks, and the steps of the method include:

[0006] Obtain the ranges of multiple device parameters of a microwave device, divide the ranges of the device parameters into multiple sub-intervals, there are overlapping intervals between adjacent sub-intervals, construct sub-models based on the multiple sub-intervals of the multiple device parameters, and each sub-model is correspondingly provided with an adjoint neural network;

[0007] Obtain training data based on the sub-intervals of the device parameters corresponding to each sub-model, pre-train each sub-model based on the training data, and perform auxiliary training through the adjoint neural network corresponding to each sub-model during the training process;

[0008] Construct sigmoid functions corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval;

[0009] Based on the sub-intervals of each device parameter corresponding to each sub-model, determine the correction sub-functions of each sub-model for each device parameter, and construct a correction function based on all the correction sub-functions of the sub-model;

[0010] Construct all sub-models with the correction function into a large-range model, and output the electromagnetic frequency-domain characteristic values corresponding to the device parameter group to be calculated based on the large-range model.

[0011] Adopting the above solution, in the initial training process of this solution, by dividing the intervals, the training duration of each sub-model is reduced, and the bridging degree between each model is initially ensured through the overlapping intervals. In the subsequent processing process, the correction function of each sub-model is constructed through the sigmoid function. The sigmoid function only works at the boundary part of the sub-model, while the internal part remains unchanged. This solution uses the sigmoid function to modify the trained sub-model. For each segmentation region at each geometric input parameter, the modification process is performed through a sigmoid function, so that the internal output value of the sub-model remains unchanged, the value at the boundary is modified, and the value outside the range of the sub-model is 0, which will not affect the output values of other models. While ensuring the output accuracy, it can be applied to the design of microwave devices with large-range changes in geometric variables.

[0012] In some embodiments of the present invention, in the step of constructing sigmoid functions corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval, calculate the parameters and parameter values of the parameters of the sigmoid function respectively based on the upper boundaries of the overlapping intervals, and construct sigmoid functions corresponding to the overlapping intervals.

[0013] In some embodiments of the present invention, in calculating the parameters and In the step of obtaining the parameter value of the parameter, the sigmoid function is calculated based on the following formula parameter and parameter:

[0014]

[0015]

[0016] where r is a preset calculation parameter, is the value of the upper boundary of the overlapping interval, is the value of the lower boundary of the overlapping interval.

[0017] In some embodiments of the present invention, the calculation parameter r is calculated according to the following formula:

[0018]

[0019] where is a preset adjustment width value.

[0020] In some embodiments of the present invention, in the step of determining the correction sub-function of each sub-model for each device parameter based on the sub-interval corresponding to each device parameter of each sub-model:

[0021] Determine whether there are adjacent sub-intervals in the shrinking direction and the increasing direction for the sub-interval of the device parameter corresponding to the sub-model;

[0022] If the sub-model does not have adjacent sub-intervals in the shrinking direction but has adjacent sub-intervals in the increasing direction, then construct the correction sub-function based on the first adjustment method;

[0023] If the sub-model has adjacent sub-intervals in the shrinking direction but does not have adjacent sub-intervals in the increasing direction, then construct the correction sub-function based on the second adjustment method;

[0024] If the sub-model has adjacent sub-intervals in the shrinking direction and has adjacent sub-intervals in the increasing direction, then construct the correction sub-function based on the third adjustment method.

[0025] In some embodiments of the present invention, the first adjustment method is to use the sigmoid function S corresponding to the overlapping interval of the sub-interval as the correction sub-function; the second adjustment method is to calculate 1 - S for the sigmoid function S corresponding to the overlapping interval of the sub-interval as the correction sub-function; the third adjustment method is to use the two sigmoid functions corresponding to the overlapping interval of the sub-interval. For the sigmoid function S1 corresponding to the left overlapping interval and the sigmoid function S2 corresponding to the right overlapping interval in the two overlapping intervals, calculate S1(1 - S2) as the correction sub-function.

[0026] In some embodiments of the present invention, in the step of constructing a correction function based on all the correction sub-functions of the sub-models, the correction sub-functions corresponding to each device parameter of the sub-models are multiplied together to obtain the correction function corresponding to each sub-model.

[0027] In some embodiments of the present invention, in the step of outputting the electromagnetic frequency-domain characteristic values corresponding to the device parameter group to be calculated based on the large-scale model:

[0028] Input the device parameter group to be calculated into the large-scale model;

[0029] Correct the output of each sub-model in the large-scale model through the corresponding correction function to obtain the value of the correction function;

[0030] Superimpose the values of the correction functions of all the sub-models to obtain the final electromagnetic frequency-domain characteristic values.

[0031] In some embodiments of the present invention, in the step of constructing sub-models based on multiple sub-intervals of multiple device parameters, for each device parameter, the sub-intervals are sequentially numbered based on the positions where the sub-intervals are located, and each sub-model corresponds to a combination of the numbers of the sub-intervals of each device parameter.

[0032] The second aspect of the present invention further provides a large-scale modeling system for microwave devices based on an adjoint neural network. The system includes a computer device, the computer device includes a processor and a memory, computer instructions are stored in the memory, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0033] The third aspect of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps implemented by the foregoing large-scale modeling method for microwave devices based on an adjoint neural network are implemented.

[0034] The additional advantages, objectives, and features of the present invention will be partially described below, and will become partially apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be pointed out and obtained specifically in the description and the drawings.

[0035] Those skilled in the art will understand that the objectives and advantages that can be achieved by the present invention are not limited to the above specifically described, and the above and other objectives that the present invention can achieve will be more clearly understood according to the following detailed description. Description of the Drawings

[0036] The accompanying drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention.

[0037] Figure 1 It is a schematic diagram of an embodiment of a method for large - scale modeling of microwave devices based on adjoint neural networks according to the present invention;

[0038] Figure 2 It is a schematic diagram of the architecture of a method for large - scale modeling of microwave devices based on adjoint neural networks according to the present invention;

[0039] Figure 3 It is a schematic diagram of the output of this solution when no correction function is set;

[0040] Figure 4 It is a schematic diagram of the output of the sigmoid function;

[0041] Figure 5 It is a schematic diagram of the processing principle of the correction function. Detailed implementation manners

[0042] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the implementation manners and the accompanying drawings. Here, the illustrative implementation manners of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0043] Here, it should also be noted that in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less related to the present invention are omitted.

[0044] As Figure 1 and 2 shown, the present invention proposes a method for large - scale modeling of microwave devices based on adjoint neural networks. The steps of this method include:

[0045] Step S100, obtaining the ranges of multiple device parameters of the microwave device, dividing the ranges of the device parameters into multiple sub - intervals, there is an overlapping interval between adjacent sub - intervals, constructing sub - models based on the multiple sub - intervals of the multiple device parameters, and each sub - model is correspondingly provided with an adjoint neural network;

[0046] In some embodiments of the present invention, the microwave device is a microwave radio - frequency device, and the device parameters include the height, length, width of the device, and the material properties of the device itself, such as dielectric constant and magnetic permeability.

[0047] Step S200: Obtain training data based on the sub-intervals of the device parameters corresponding to each sub-model, and pre-train each sub-model based on the training data. During the training process, perform auxiliary training through the adjoint neural network corresponding to each sub-model;

[0048] In the specific implementation process, the sub-model adopts the SAANN neural network structure. SAANN (Sparse Adversarial Adjoint Neural Network) is a neural network structure with strong generalization ability assisted by sensitivity information;

[0049] The SAANN model consists of two parts: one is the original neural network for training the electromagnetic response relationship of microwave devices, and the other is the adjoint neural network for training the derivative information of the electromagnetic response with respect to sensitivity variables. They share a set of internal neuron connection parameters, but their activation functions are different. The adjoint neural network has the same input as the original neural network, and the output of the entire SAANN model includes the output of the original network and the output of the adjoint network. During the training process, both the original network and the adjoint network exist, but the finally obtained model only includes the structure of the original network.

[0050] In some embodiments of the present invention, during the pre-training process, the sub-model can simultaneously learn the input-output behavior of the filter and the derivative of the output response with respect to the sensitivity parameters. The adjoint neural network model is derived from the original neural network model and shares the same weight coefficients with the original neural network. After the model training is completed, a simple final neural network model structure is directly obtained. The input variables of this model are usually the design variables of the device and the frequency parameters.

[0051] Adopting the above scheme, after adding the sensitivity derivative to the sub-model, the derivative can provide a guiding role for the change direction of the adjoint neural network fitting function, so that less data can be used for modeling, accelerating the modeling process in terms of data volume and simulation. And due to the addition of the derivative, the generalization of the model is better. To accelerate the modeling speed, the training of each sub-model is carried out in parallel. When training the sub-model, to avoid the situation of over-training or under-training of the neural network model, the number of neurons is adjusted according to the training error and the test error during the training process until a neural network surrogate model that meets the user-defined accuracy requirements is established. The established surrogate model can provide a fast and accurate prediction of the electromagnetic characteristics and their derivatives of the microwave device within a certain range within its training range.

[0052] Step S300: Construct a sigmoid function corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval;

[0053] Step S400: Based on the sub-intervals of each device parameter corresponding to each sub-model, determine the correction sub-functions of each sub-model for each device parameter, and construct a correction function based on all the correction sub-functions of the sub-models.

[0054] Step S500: Construct all the sub-models with the correction function into a large-range model, and output the electromagnetic frequency-domain characteristic values corresponding to the device parameter group to be calculated based on the large-range model.

[0055] In the specific implementation process, derivative information corresponding to the device parameter group is output through an adjoint neural network.

[0056] In the specific implementation process, the electromagnetic frequency-domain characteristic values can be S-parameter values, gain, quality factor, etc.

[0057] Adopting the above solution, in the initial training process of this solution, by dividing the intervals, the training duration of each sub-model is reduced, and the degree of bridging between each model is initially ensured through overlapping intervals. In the subsequent processing process, the correction function of each sub-model is constructed through the sigmoid function. The sigmoid function only works at the boundary part of the sub-model, while the internal part remains unchanged. This solution uses the sigmoid function to modify the trained sub-model. For each segmentation region at each geometric input parameter, the modification process is performed through a sigmoid function, so that the internal output value of the sub-model remains unchanged, the values at the boundary are modified, and the values outside the sub-model range are 0, which will not affect the output values of other models. This solution corrects the boundaries of each sub-model through the correction function, which can be applicable to the design of microwave devices with large-range changes in geometric variables while ensuring the output accuracy.

[0058] In some embodiments of the present invention, in the step of constructing the sigmoid function corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval, the parameters and parameter values of the

[0059] parameters of the sigmoid function are calculated respectively based on the upper boundary of the overlapping interval, and the sigmoid function corresponding to the overlapping interval is constructed. parameters and In some embodiments of the present invention, in the step of calculating the parameters and parameter values of the

[0060]

[0061]

[0062] where r is a preset calculation parameter, is the value of the upper boundary of the overlapping interval, is the value of the lower boundary of the overlapping interval.

[0063] In some embodiments of the present invention, the calculation parameter r is calculated according to the following formula:

[0064]

[0065] where, is a preset adjustment width value, which can be specifically set to 0.999, 0.997 or 0.998, etc.

[0066] In some embodiments of the present invention, in the step of determining the correction sub-function of each sub-model for each device parameter in the sub-interval corresponding to each sub-model based on each device parameter:

[0067] Determine whether there are adjacent sub-intervals in the shrinking direction and the increasing direction in the sub-interval of the device parameter corresponding to the sub-model;

[0068] If the sub-model does not have an adjacent sub-interval in the shrinking direction but has an adjacent sub-interval in the increasing direction, then construct the correction sub-function based on the first adjustment method;

[0069] If the sub-model has an adjacent sub-interval in the shrinking direction but does not have an adjacent sub-interval in the increasing direction, then construct the correction sub-function based on the second adjustment method;

[0070] If the sub-model has an adjacent sub-interval in the shrinking direction and has an adjacent sub-interval in the increasing direction, then construct the correction sub-function based on the third adjustment method.

[0071] In some embodiments of the present invention, the first adjustment method is to use the sigmoid function S corresponding to the overlapping interval of the sub-interval as the correction sub-function; the second adjustment method is to calculate 1 - S for the sigmoid function S corresponding to the overlapping interval of the sub-interval as the correction sub-function; the third adjustment method is to use the two sigmoid functions corresponding to the overlapping interval of the sub-interval. For the sigmoid function S1 corresponding to the left overlapping interval and the sigmoid function S2 corresponding to the right overlapping interval in the two overlapping intervals, calculate S1(1 - S2) as the correction sub-function.

[0072] As Figure 4 and 5 shown, Figure 5 where S(x) is the original sigmoid function and h(x) = 0.01 , in the specific implementation process, when the adjacent sub-model of the j-th model is on its right side, in order to keep the internal value output of the j-th sub-model unchanged and the value of the adjacent sub-model does not affect the output of the j-th sub-model, multiply by (1 - S). On the contrary, when the adjacent sub-model of the j-th model is on its left side, in order to keep the internal value output of the j-th sub-model unchanged and the value of the adjacent sub-model does not affect the output of the j-th sub-model, multiply by S. When there are adjacent sub-models on both the left and right sides of the j-th sub-model, multiply by S*(1 - S).

[0073] Adopting the above scheme, the characteristic of the sigmoid function is that the left side of the processing range remains unchanged and the values on the right side of the processing range become 0. The 1 - sigmoid function is the opposite.

[0074] In the specific implementation process, this scheme can process the boundary part of the sub-model function while keeping the internal part unchanged, and the sigmoid function has such characteristics. As Figure 4 shown, it can be seen that by constraining the device parameters and , the output function fD that does not connect multiple sub-models only works in the boundary part of the sub-model function, while the internal part remains unchanged. This scheme uses the sigmoid function to modify the trained sub-model function. For each segmentation region at each geometric input parameter, a sigmoid function is required to perform the modification process, so that the internal output value of the sub-model remains unchanged, the values at the boundary are modified, and the values outside the sub-model range are 0, which will not affect the output values of other models.

[0075] In some embodiments of the present invention, in the step of constructing the correction function based on all the correction sub-functions of the sub-model, multiply the correction sub-functions corresponding to the sub-model for each device parameter to obtain the correction function corresponding to each sub-model.

[0076] In some embodiments of the present invention, in the step of outputting the electromagnetic frequency domain characteristic value corresponding to the device parameter group to be calculated based on the large-range model:

[0077] Input the device parameter group to be calculated into the large-range model;

[0078] Correct the output of each sub-model in the large-range model through the corresponding correction function to obtain the value of the correction function;

[0079] Superimpose the values of the correction functions of all sub-models to obtain the final electromagnetic frequency domain characteristic value.

[0080] Adopting the above - mentioned solution, after obtaining the models of each sub - region through pre - training, since the sub - models are independently trained, there will be a discontinuity problem at the boundaries of adjacent models. To reduce the difference between the outputs of adjacent sub - models at the boundaries, this solution proposes to set an overlapping region for adjacent sub - models along each dimension as described above, and adjacent sub - models will all be trained on this part of the data. After the sub - model training process, adjacent sub - models have very similar output responses at the boundaries. Then, use the mathematical properties of the sigmoid function to correct each dimension of each sub - model to solve the multi - dimensional discontinuity problem; the sigmoid function can keep the data within the sub - model unchanged, make the data outside the sub - model be 0, and make reasonable modifications to the data on the boundary. After specially processing the boundary of the overall model with the sigmoid function, a continuous large - model will be obtained. Finally, the output responses of each sub - model are superimposed to obtain the output response of the total model.

[0081] In some embodiments of the present invention, in the step of constructing sub - models based on multiple sub - intervals of multiple device parameters, for each type of device parameter, the sub - intervals are sequentially numbered based on the positions where the sub - intervals are located, and each sub - model corresponds to a combination of numbers of sub - intervals of each device parameter.

[0082] For example, if the device parameters include A and B, and the interval of each device parameter is (0, 1), and each device parameter is divided into two sub - intervals (0, 0.6) and (0.4, 1), then (0, 0.6) is numbered 1, (0.4, 1) is numbered 2. Then the corresponding sub - models include the 1st sub - model corresponding to the sub - interval with A parameter numbered 1 and B parameter numbered 1; the 2nd sub - model corresponding to the sub - interval with A parameter numbered 1 and B parameter numbered 2; the 3rd sub - model corresponding to the sub - interval with A parameter numbered 2 and B parameter numbered 1; the 4th sub - model corresponding to the sub - interval with A parameter numbered 2 and B parameter numbered 2.

[0083] In summary, during the current process of radio - frequency and microwave device modeling, both designers and users hope that the model can work within a large range, that is, when the design variables change within a large range, the model always remains accurate. Establishing an accurate large - range parameterized model for the electromagnetic response of radio - frequency and microwave devices in an electromagnetic environment can increase the scope of use of the model, provide accurate and fast output responses within a large range; at the same time, it can reduce the number of iterations during the optimal design of radio - frequency and microwave devices and avoid falling into local optimal solutions. However, as the modeling range, that is, the range of change of the design variables, increases, the complexity and non - linearity of the electromagnetic response increase rapidly.

[0084] To establish an accurate large-scale model, the most commonly used methods in the prior art are the knowledge-based neural network modeling method and the space mapping method combined with neural networks. These two methods increase the working range of the model by introducing information such as the equivalent circuit of the device or the prior physical knowledge and mathematical models of the device. However, both of these methods rely on empirical models or prior knowledge, so there are certain application limitations. When there is no applicable empirical model or prior knowledge for a new type of radio frequency and microwave device, these two methods cannot be used to model the device. Moreover, although the above two methods can expand the working range of the model, they do not really solve the problem of large-scale modeling. In the process of using the existing single-neural-network-based modeling method for large-scale parametric modeling, there are problems such as a large demand for data volume, high model complexity, high calculation cost, and low model accuracy.

[0085] Therefore, this solution proposes an adjoint neural network large-scale parametric model for microwave devices based on decomposition technology to quickly establish an accurate parametric model for microwave devices with geometric variables changing in a large range. This solution is applicable to the design of microwave radio frequency filters, such as cavity filters, microstrip line filters, interdigital filters and other filter designs. Compared with the existing large-scale parametric modeling methods for microwave filters, the model proposed in this solution has better generalization performance, requires less training data volume, speeds up the modeling speed, and can also predict the derivatives of the output response with respect to each geometric parameter, greatly solving the difficulty of large-scale modeling of microwave filters and accelerating the design and optimization of microwave radio frequency filters.

[0086] The embodiment of the present invention also provides a large-scale modeling system for microwave devices based on adjoint neural networks. The system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0087] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps implemented by the aforementioned large-scale modeling method for microwave devices based on adjoint neural networks. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0088] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in connection with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement it in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, and so on. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium or transmitted through a data signal carried in a carrier wave over a transmission medium or a communication link.

[0089] It should be clear that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0090] In the present invention, the features described and / or illustrated for one embodiment can be used in the same or a similar manner in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0091] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A large-scale modeling method for microwave devices based on adjoint neural network, characterized in that: The steps of the method include: Acquire the range of multiple device parameters of the microwave device, wherein the device parameters include the height, length, width of the device and the material properties of the device itself, divide the range of the device parameters into multiple sub-intervals, and there are overlapping intervals between adjacent sub-intervals, and construct sub-models based on the multiple sub-intervals of the multiple device parameters, and each sub-model is correspondingly provided with an accompanying neural network; Acquire training data based on the subinterval of the device parameter corresponding to each sub-model, pre-train each sub-model based on the training data, and perform auxiliary training through the accompanying neural network corresponding to each sub-model during the training process; For the upper and lower boundaries of each overlapping interval, construct a sigmoid function corresponding to each overlapping interval; Based on the subintervals of each device parameter corresponding to each submodel, determine the correction subfunction of each submodel for each device parameter, construct a correction function based on all the correction subfunctions of the submodel, and determine whether the subintervals of the device parameters corresponding to the submodel have adjacent subintervals in the shrinking direction and the increasing direction; if the submodel does not have adjacent subintervals in the shrinking direction, but has adjacent subintervals in the increasing direction, then construct a correction subfunction based on the first adjustment method, and the first adjustment method is to use the sigmoid function S corresponding to the overlapping interval of the subinterval as the correction subfunction; if the submodel has adjacent subintervals in the shrinking direction, but does not have adjacent subintervals in the increasing direction interval, a correction sub-function is constructed based on the second adjustment method, and the second adjustment method is to calculate 1-S as the correction sub-function for the sigmoid function S corresponding to the overlapping interval of the sub-interval; if the sub-model has adjacent sub-intervals in the shrinking direction and adjacent sub-intervals in the increasing direction, a correction sub-function is constructed based on the third adjustment method, and the third adjustment method is to correspond the overlapping interval of the sub-interval to two sigmoid functions, and for the sigmoid function S1 corresponding to the overlapping interval on the left and the sigmoid function S2 corresponding to the overlapping interval on the right of the two overlapping intervals, S1 (1-S2) is calculated as the correction sub-function; All sub-models provided with correction functions are constructed into a large-scale model, and electromagnetic frequency domain characteristic values ​​corresponding to the device parameter group to be calculated are output based on the large-scale model.

2. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 1 is characterized in that: In the step of constructing a sigmoid function corresponding to each overlapping interval for the upper and lower boundaries of each overlapping interval, the sigmoid function is calculated based on the upper boundary of the overlapping interval. Parameters and The parameter value of the parameter constructs the sigmoid function corresponding to the overlapping interval.

3. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 2 is characterized in that: Calculate the sigmoid function based on the upper boundary of the overlapping interval Parameters and In the step of calculating the parameter value of the parameter, the sigmoid function is calculated based on the following formula Parameters and parameter: Among them, r is the preset calculation parameter, is the value of the upper boundary of the overlapping interval, is the value of the lower boundary of the overlapping interval.

4. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 3 is characterized in that: The parameter r is calculated according to the following formula: in, The preset adjustment width value.

5. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 1 is characterized in that: In the step of constructing a correction function based on all correction sub-functions of the sub-model, the correction sub-functions corresponding to each device parameter of the sub-model are multiplied together to obtain a correction function corresponding to each sub-model.

6. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 1 is characterized in that: In the step of outputting electromagnetic frequency domain characteristic values ​​corresponding to the device parameter group to be calculated based on the large-scale model: Inputting a device parameter set to be calculated into the large-scale model; Correcting the output of each sub-model in the large-scale model by using a corresponding correction function to obtain a value of the correction function; The values ​​of the correction functions of all sub-models are superimposed to obtain the final electromagnetic frequency domain characteristic value.

7. The large-scale modeling method of microwave devices based on adjoint neural network according to claim 1 is characterized in that: In the step of constructing a sub-model based on multiple sub-intervals of multiple device parameters, the sub-intervals of each device parameter are sequentially numbered based on the positions of the sub-intervals, and each sub-model corresponds to a number combination of the sub-intervals of each device parameter.

8. A microwave device large-scale modeling system based on adjoint neural network, characterized in that: The system includes a computer device, which includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 7.

Images