CN114676522A - Aerodynamic shape optimization design method, system and equipment integrating GAN and transfer learning - Google Patents

Abstract

The aerodynamic shape optimization design method, system and equipment integrating GAN and transfer learning include the following steps: collecting design parameters to establish a computational fluid dynamics simulation process; obtaining sample information that has completed similar optimization tasks; Variable design space; collect the corresponding sample set based on the latent variable design space; search for the coordinates where the EI function maximizes to obtain new coordinates; evaluate the performance of the geometric design of the new coordinates, and add the two groups of samples obtained after the evaluation to the sample set; The steps of building a surrogate model, searching for new samples, and adding new samples are performed cyclically for the updated sample set until the optimization stopping condition is satisfied. The invention solves the problem of incompatibility of parameterized spaces of different design tasks, accelerates the solution of target tasks by using the solutions of similar tasks that have been completed, breaks through the limitation of "starting from scratch" for each specific optimization problem of the classical design optimization method, and improves new task optimization efficiency.

Description

Aerodynamic shape optimization design method, system and equipment integrating GAN and transfer learning

technical field

The invention belongs to the field of aerodynamic shape design optimization, and particularly relates to an aerodynamic shape optimization design method, system and equipment integrating GAN and transfer learning.

Background technique

In recent years, in the field of aerodynamic shape engineering optimization of mechanical components, automatic optimization design methods have been more and more widely used. The automatic optimization design method needs to manually determine a design space that can be parameterized, and each sample in the space corresponds to a specific geometric design scheme. Taking the results of computer simulation calculations as the optimization goal, the geometry of components with excellent aerodynamic performance can be automatically designed with the help of specific optimization algorithms. The use of automated design methods can effectively reduce the need for designer experience and complete designs quickly and with high quality.

However, high-precision CFD sample calculation often consumes a lot of time and computing resources, so it is of great engineering significance to reduce the number of sample calculations in the optimization design process and improve the efficiency of the optimization design.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an aerodynamic shape optimization design method, system and device integrating GAN and transfer learning, which can be realized by flexibly utilizing the sample information of similar optimization tasks that have been completed.

To achieve the above object, the present invention adopts the following technical solutions:

An aerodynamic shape optimization design method integrating GAN and transfer learning, including the following steps:

Collect the design parameters of the design object to establish a computational fluid dynamics simulation process, and at the same time determine the optimal target performance of the optimal design;

Get sample information on similar optimization tasks that have been completed. Similar optimization tasks are defined here as optimization tasks with similar flow characteristics, such as two cascades that also work under subsonic conditions but with different operating parameters such as Mach number and Reynolds number. Samples from similar tasks as described above will be collected and the same scatter representation will be used to characterize the profile data.

Denote these sample data as {X _data , Y _data }, assuming that the number of existing samples is n _data , each sample contains geometric design data X _data and performance result data Y _data representing the optimization goal;

Using the geometric design data X _data to build a generative adversarial neural network GAN, a new latent variable design space is obtained, and the number of design variables in the design space is d;

Through the trained adversarial neural network, input d-dimensional data and output a geometric design, and denote the obtained latent variable design space as Z; use the GA algorithm to project the geometric design source data samples {X _data , Y _data } to the space Z , denoted as {Z _S ,Y _S };

The number of initial samples in the space Z is n _ini =6d, the sample coordinate set in the Z space is denoted as Z _T , and the obtained samples are evaluated to obtain the corresponding evaluation value Y _T ;

Use the sample set {Z _T , Y _T } to establish surrogate models respectively, and search for the coordinates that maximize the EI function to obtain new coordinates

对于的几何设计进行性能评估，将评估后得到的两组样本加入样本集合{Z_T,Y_T}中；respectively for the new coordinates

For the performance evaluation of the geometric design, the two groups of samples obtained after the evaluation are added to the sample set {Z _T , Y _T };

The steps of building a surrogate model, searching for new samples, and adding new samples are performed cyclically for the updated sample set until the optimization stopping condition is satisfied.

Further, the design parameters of the design object are collected to establish a computational fluid dynamics simulation process.

Further, sample information that has completed similar optimization tasks is obtained from the database.

Further, the number d of design variables is a self-determined integer. The larger the value, the larger the variation range of geometric design in space, and the richer the design details.

Further, in the trained adversarial neural network, the input variable z of the generator is a d-dimensional random variable uniformly distributed in the [0,1] interval, and the number of training iterations in the GAN neural network is set to 50,000 steps.

Further, the LHS method is used for initial sampling in space Z.

对于所建立的代理模型，搜索EI函数最大化的坐标得到新坐标

Further, use the sample set {Z _T , Y _T } to establish a single-fidelity Kriging surrogate model

For the established surrogate model, search for the coordinates where the EI function maximizes to get new coordinates

对于所建立的代理模型，搜索EI函数最大化的坐标得到新坐标

Further, a multi-fidelity co-Kriging surrogate model is established using the sample set {Z _T , Y _T } as high-precision samples and the sample set {Z _S , Y _S } as low-precision samples

For the established surrogate model, search for the coordinates where the EI function maximizes to get new coordinates

Further, aerodynamic shape optimization systems that integrate GAN and transfer learning include:

The fluid dynamics simulation process establishment module is used to collect the parameters of the design object to establish the computational fluid dynamics simulation process, and at the same time determine the optimized target performance of the optimized design;

The data acquisition module is used to obtain sample information that has completed similar optimization tasks, record these sample data as {X _data , Y _data }, assuming that the number of existing samples is n _data , each sample contains geometric design data X _data and The performance result data Y _data representing the optimization objective;

The adversarial neural network building module is used to build a generative adversarial neural network GAN using the geometric design data X _data , and obtain a new latent variable design space. After the data, a geometric design is output, and the obtained latent variable design space is represented by Z; the geometric design source data samples {X _data , Y _data } are projected into the space Z, denoted as {Z _S , Y _S }; In the process, the GA algorithm is optimized to approximate the corresponding coordinate Z _S of the geometric model in the space Z. The input of the optimization search is the coordinate in the Z space. The goal of optimization is to make the coordinate in the Z space pass through the generator. The error of the actual model in _data is the smallest; the value of Y _S is the same as that of Y _data .

The evaluation value acquisition module is used for the number of initial samples in the space Z to be n _ini =6d, the sample coordinate set in the Z space is denoted as Z _T , and the obtained samples are evaluated to obtain the corresponding evaluation value Y _T ;

The new coordinate acquisition module is used to use the sample set {Z _T , Y _T } to establish surrogate models respectively, and search for the coordinates that maximize the EI function to obtain new coordinates

对于的几何设计进行性能评估，将评估后得到的两组样本加入样本集合{Z_T,Y_T}中；对于更新后的样本集合循环执行建立代理模型，搜索新样本，添加新样本的步骤直到满足优化停止条件。Evaluation module for separately evaluating the new coordinates

For the performance evaluation of the geometric design, the two groups of samples obtained after evaluation are added to the sample set {Z _T , Y _T }; for the updated sample set, the steps of establishing a surrogate model, searching for new samples, and adding new samples are performed cyclically until The optimization stop condition is satisfied.

Further, a computer device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, when the processor executes the computer program, the integration of GAN and transfer learning is realized. Steps of aerodynamic shape optimization design method.

Compared with the prior art, the present invention has the following technical effects:

The invention uses the deep learning technology GAN network to establish a new design space parameterized modeling method, this method solves the problem that the parameterized space of different design tasks is incompatible, so that the data of the completed design tasks in the database can be used without Barriers can be applied to new design tasks, thereby improving the optimization efficiency of new tasks.

This method uses the deep learning technology GAN network to complete the optimized parametric modeling. Compared with the traditional modeling method, this method is not limited to a fixed reference design, and can construct reasonable components within a large range of variation scales. Geometric modeling broadens the exploration scope of optimal design.

This method proposes a GTO optimization algorithm in the optimization process. By constructing a single-fidelity surrogate model and a multi-fidelity surrogate model at the same time, the robustness of the optimization is fidelity and the accumulated data in the database is effectively used to improve the optimization performance. efficiency.

Description of drawings

FIG. 1 is a schematic diagram of an embodiment of the present invention.

FIG. 2 is a schematic diagram of a specific flow of a latent space construction method according to an embodiment of the present invention.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is further described:

An aerodynamic shape design optimization method integrating GAN and transfer learning, including the following steps:

According to the design requirements and optimization objectives of the design object, the computational fluid dynamics simulation process is established, and the optimization target performance of the optimization design is determined at the same time. For each input geometric design, a unique aerodynamic performance parameter output can be obtained;

Obtain sample information of similar optimization tasks that have been completed from the database, and denote these sample data as {X _data , Y _data }, assuming that the number of existing samples is n _data , each sample contains geometric design data that can be represented by point clouds X _data and performance result data Y _data representing the optimization objective.

For a specific type of design task, there is often a database that records the previous design results. The source of the information in the database can be the design geometry information of mature products at home and abroad, or the previous design data accumulated by users themselves.

A generative adversarial neural network (GAN) is built using the geometric design data X _data , thereby obtaining a new latent variable design space.

Determine the number d of design variables in the constructed design space. d is an integer determined by the user. The larger the value of d, the greater the variation range of the geometric design in the space, and the richer the design details, but it will also increase the difficulty of optimization search.

As shown in Figure 2, the input variable z of the generator is a d-dimensional random variable uniformly distributed in the [0,1] interval. The specific settings of the generator and discriminator in the GAN neural network are shown in Table 1. Set the number of iterations for training to 50000 steps.

Table 1GAN network setup details

Through the trained neural network, a geometric design can be output after inputting d-dimensional data. The obtained latent variable design space is denoted by Z. The geometric design source data samples {X _data , Y _data } can be projected into the space Z, denoted as {Z _S , Y _S }. In the projection process, the corresponding coordinate Z _S of the geometric model in the space Z is approximated by the GA algorithm optimization. The input of the optimization search is the coordinate in the Z space. The goal of optimization is to make the coordinates in the Z space pass through the generator. Generated geometric model The error with the actual model in X _data is the smallest; the value of Y _S is the same as that in Y _data .

The LHS method performs initial sampling in space Z, the number of initial samples is n _ini =6d, and the set of sample coordinates in Z space is denoted as Z _T . The obtained samples are evaluated to obtain the corresponding evaluation value Y _T .

对于所建立的代理模型，搜索EI函数最大化的坐标得到新坐标

Use the sample set {Z _T ,Y _T } to build a single-fidelity Kriging surrogate model

For the established surrogate model, search for the coordinates where the EI function maximizes to get new coordinates

对于所建立的代理模型，搜索EI函数最大化的坐标得到新坐标

Use the sample set {Z _T , Y _T } as high-precision samples and the sample set {Z _S , Y _S } as low-precision samples to establish a multi-fidelity co-Kriging surrogate model

For the established surrogate model, search for the coordinates where the EI function maximizes to get new coordinates

对于的几何设计进行性能评估，将评估后得到的两组样本加入样本集合{Z_T,Y_T}中。respectively for the new coordinates

For the performance evaluation of the geometric design, the two groups of samples obtained after the evaluation are added to the sample set {Z _T , Y _T }.

The steps of building a surrogate model, searching for new samples, and adding new samples are performed cyclically for the updated sample set until the optimization stopping condition is satisfied.

As shown in Figure 1, this embodiment provides an aerodynamic shape design optimization method integrating GAN and transfer learning and applies it to the aerodynamic shape optimization design, which specifically includes the following steps:

1. Establishment of optimization design tasks

In this embodiment, the most commonly used low-speed airfoil is selected as the design object. The optimization goal is to maximize the lift-drag ratio L/D of the airfoil. The outflow conditions of the airfoil design are: Reynolds number Re=1.8×10 ⁶ , Mach number Ma=0.01, and the attack angle of the airfoil is set to 0 degrees.

2. Establishment of performance evaluation model

The XFOIL software is used to automatically evaluate the obtained two-dimensional geometric model, and the output is the lift-drag ratio of the airfoil.

3. Establishment of design space

The optimization design space of the present invention is automatically established by using the GAN method according to the completed task samples. In this embodiment, a set of optimized samples {X _data , Y _data } are used as the source samples, and the outflow of the completed optimization task is used. Mach number Ma=0.45. Using the point cloud of geometric samples as input, set the number of design space variables d = 13, and build a GAN neural network as shown in Table 1. The latent variable design space Z is established by using the trained GAN network as a parametric modeling method. The geometric design source data samples {X _data , Y _data } are projected into space Z, denoted as {Z _S , Y _S }; in the projection process, the GA algorithm is optimized to approximate the corresponding geometric model in space Z Coordinate Z _S , the input of the optimization search is the coordinates in the Z space, and the goal of optimization is to minimize the error between the geometric model generated by the generator in the Z space and the actual model in the X _data ; the value of Y _S is the same as that of the Y _data .

4. The specific process of optimizing the design

Referring to Figure 1, the specific process is as follows:

4a. In the established design space Z, use the LHS method to obtain the coordinates of 78 samples with a relatively uniform distribution, and use the established performance evaluation model to evaluate the performance to obtain the lift-to-drag ratio of the 78 design samples, and to obtain the 78 samples. The set {Z _T ,Y _T }.

(由于默认优化过程中的目标为最小值，样本值设置为升阻比乘以-1)4b. Use the sample coordinates Z _T and sample values Y _T obtained in 4a to establish a single-fidelity Kriging surrogate model, and obtain new coordinates by maximizing EI search on the established single-fidelity Kriging surrogate model

(Since the target in the default optimization process is the minimum value, the sample value is set to the lift-drag ratio multiplied by -1)

4c. Use the sample set {Z _T , Y _T } obtained in 4a as the high-precision source and the transformed completed task data {Z _S , Y _S } as the low-precision source to build a multi-fidelity co-Kriging agent model, obtain new coordinates by maximizing EI search on the established multi-fidelity co-Kriging surrogate model

对应的几何设计进行性能评估获得其升阻比，将这两个样本的数据加入集合{Z_T,Y_T}。4d. Use the established performance evaluation model for the new coordinates

The corresponding geometric design is evaluated for performance to obtain its lift-to-drag ratio, and the data of these two samples are added to the set {Z _T , Y _T }.

4f. Repeat steps 4b to 4d until the optimized stopping conditions are met.

5. Results of the optimized design:

Repeat the above steps for 10 independent optimizations, and the results of the airfoil optimization design are shown in Table 2.

Table 2 Optimization results details

lift to drag ratio reference design 137.047 optimized design 240.153 Lift ratio 75.23% Standard deviation of optimization results 5.911

The principle of the present invention is as follows:

The biggest feature and innovation of the present invention is reflected in the application of GAN neural network to obtain two parts, the latent variable design space generation method and the GTO optimization algorithm for realizing aerodynamic shape transfer optimization, which can be reused for multiple similar design tasks. The joint use of these two methods can apply the data of similar optimization tasks to the new optimization design task, which can effectively improve the optimization efficiency of the new task and save money.

The latent variable design space generation method uses a GAN neural network algorithm to build a parametric space that conforms to the geometric design characteristics of the source task.

The GTO migration optimization algorithm completes optimization points by simultaneously building a multi-fidelity surrogate model that includes source task migration information and a single-fidelity surrogate model that does not include migration information. Compared with the EGO algorithm that only builds a single-fidelity model, the GTO algorithm has higher efficiency due to the use of migration information; compared with the STO algorithm that only builds a multi-fidelity model, the GTO algorithm has stronger robustness. Optimization stalls due to negative transfer can also be avoided.

In one embodiment of the present invention, an aerodynamic shape design optimization system integrating GAN and transfer learning is provided, which can be used to realize the above-mentioned aerodynamic shape design optimization method integrating GAN and transfer learning. The aerodynamic shape design optimization system includes:

The fluid dynamics simulation process establishment module uses the design parameters of the design object to establish the computational fluid dynamics simulation process, and at the same time determines the optimal target performance of the optimal design;

The data acquisition module is used to obtain sample information that has completed similar optimization tasks, record these sample data as {X _data , Y _data }, assuming that the number of existing samples is n _data , each sample contains geometric design data X _data and The performance result data Y _data representing the optimization objective;

The adversarial neural network building module is used to build a generative adversarial neural network GAN using the geometric design data X _data to obtain a new geometric design space. The number of design variables in the design space is d; for the adversarial neural network completed through training, input d-dimensional data Then output a geometric design, and denote the obtained latent variable design space as Z; use the GA algorithm to project the geometric design source data sample {X _data , Y _data } into the space Z, denoted as { Z _S , Y _S };

The evaluation value acquisition module is used for the number of initial samples in the space Z to be n _ini =6d, the sample coordinate set in the Z space is denoted as Z _T , and the obtained samples are evaluated to obtain the corresponding evaluation value Y _T ;

The new coordinate acquisition module is used to use the sample set {Z _T , Y _T } to establish surrogate models respectively, and search for the coordinates that maximize the EI function to obtain new coordinates

对于的几何设计进行性能评估，将评估后得到的两组样本加入样本集合{Z_T,Y_T}中；对于更新后的样本集合循环执行建立代理模型，搜索新样本，添加新样本的步骤直到满足优化停止条件。Evaluation module for separately evaluating the new coordinates

For the performance evaluation of the geometric design, the two groups of samples obtained after evaluation are added to the sample set {Z _T , Y _T }; for the updated sample set, the steps of establishing a surrogate model, searching for new samples, and adding new samples are performed cyclically until The optimization stop condition is satisfied.

In yet another embodiment of the present invention, a computer device is provided, the computer device includes a processor and a memory, the memory is used for storing a computer program, the computer program includes program instructions, and the processor is used for executing the computer Program instructions stored in the storage medium. The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (Digital Signal Processors, DSP), application specific integrated circuits (ASICs), and off-the-shelf programmable gate arrays. (Field-Programmable GateArray, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, specifically suitable for One or more instructions in the computer storage medium are loaded and executed to realize the corresponding method process or corresponding function; the processor according to the embodiment of the present invention can be used for the operation of integrating GAN and transfer learning for aerodynamic shape design optimization.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: the present invention can still be Modifications or equivalent replacements are made to the specific embodiments of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (10)

1. The pneumatic shape optimization design method fused with GAN and transfer learning is characterized by comprising the following steps of:

collecting design parameters of a design object to establish a computational fluid dynamics simulation flow, and simultaneously determining the optimization target performance of the optimization design;

obtaining sample information of completed similar optimization task, and recording the sample data as { X_data,Y_dataLet n be the number of existing samples_dataEach sample containing geometric design data X represented by scatter_dataWith performance result data Y representing optimization objectives_data；

Using geometric design data X_dataEstablishing and generating a confrontation neural network GAN to obtain a new latent variable design space, wherein the number of design variables of the design space is d;

inputting d-dimensional data and outputting a geometric design by a generator obtained through training, wherein the obtained latent variable design space is represented by Z; geometric design source data samples { X ] using GA algorithm_data,Y_dataAll project into space Z, denoted as { Z }_S,Y_S}；

The number of initial samples in space Z is n_ini6d, the set of sample coordinates in Z space is denoted as Z_TEvaluating the obtained sample to obtain corresponding evaluation value Y_T；

Using a sample set { Z_T,Y_TRespectively establishing proxy models, searching the coordinates with the maximized EI function to obtain new coordinates

Respectively to the new coordinates

Performing performance evaluation on the geometric design, and adding two groups of samples obtained after evaluation into a sample set { Z_T,Y_TIn (1) };

and circularly executing the steps of establishing a proxy model for the updated sample set, searching for a new sample, and adding the new sample until an optimization stopping condition is met.

2. The pneumatic shape optimization design method based on the fusion of the GAN and the transfer learning of claim 1, wherein a computational fluid dynamics evaluation model for obtaining the optimized target performance parameters is established.

3. The pneumatic shape optimization design method based on the fusion of the GAN and the transfer learning of claim 1, wherein the sample information of the similar optimization task is obtained from a database.

4. The pneumatic shape optimization design method integrating the GAN and the transfer learning as claimed in claim 1, wherein the number d of design variables is a self-determined integer, and the larger the value is, the larger the variation range of the geometric design in space is, and the richer the design details are.

5. The pneumatic shape optimization design method integrating GAN and transfer learning as claimed in claim 1, wherein in the trained GAN neural network, the input variable z of the generator is a d-dimensional random variable uniformly distributed in the [0,1] interval, and the number of training iterations in the GAN neural network is set to 50000 steps.

6. The pneumatic shape optimization design method based on the fusion of GAN and the migration learning of claim 1, wherein the initial sampling is performed in the space Z by using LHS method.

7. The method of claim 1, wherein a sample set { Z } is used_T,Y_T}, establish single fidelityKriging agent model

For the established proxy model, searching the coordinates with maximized EI function to obtain new coordinates

8. The method of claim 1, wherein a sample set { Z } is used_T,Y_TAs high precision samples, use the sample set { Z }_S,Y_SEstablishing a multi-fidelity co-Kriging agent model as a low-precision sample

For the established proxy model, searching the coordinates with maximized EI function to obtain new coordinates

9. Pneumatic shape optimal design system that fuses GAN and migration learning, its characterized in that includes:

the fluid dynamics simulation flow establishing module is used for establishing a computational fluid dynamics simulation flow by using parameters of a design object and determining the optimization target performance of the optimization design;

the data acquisition module is used for acquiring sample information of completed similar optimization tasks, and remembering the sample data as { X_data,Y_dataLet n be the number of existing samples_dataEach sample containing geometric design data X_dataWith performance result data Y representing optimization objectives_data；

A GAN neural network building block for using geometric design data X_dataEstablishing a GAN to generate an antagonistic neural network to obtain a new latent variable design space with design variablesThe number is d; inputting d-dimensional data and outputting a geometric design through a generator of the trained GAN neural network, and expressing an obtained latent variable design space by Z; geometric design source data samples { X ] using GA algorithm_data,Y_dataProject into space Z, noted as Z_S,Y_S}；

An initial sample evaluation module for n number of initial samples in space Z_ini6d, the set of sample coordinates in Z space is denoted as Z_TEvaluating the obtained sample to obtain corresponding evaluation value Y_T；

A new design search evaluation module for using the sample set { Z_T,Y_TRespectively establishing proxy models, searching the coordinates with the maximized EI function to obtain new coordinates

For new coordinate

Performing performance evaluation on the corresponding geometric design, and adding two groups of samples obtained after evaluation into a sample set { Z_T,Y_TIn (j) }; and circularly executing the steps of establishing a proxy model for the updated sample set, searching for a new sample, and adding the new sample until an optimization stopping condition is met.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor when executing the computer program implements the steps of the method for pneumatic shape-optimized design with fusion of GAN and migration learning according to any of claims 1 to 8.

Images