Search

CN-121997803-A - Rapid prediction method for wing profile bypass flow field based on Fourier neural operator

CN121997803ACN 121997803 ACN121997803 ACN 121997803ACN-121997803-A

Abstract

The invention provides a rapid prediction method of an airfoil bypass flow field based on a Fourier neural operator, and belongs to the field of intelligent fluid mechanics. The invention integrates machine learning and computational fluid mechanics depth, and rapidly predicts the future flow field state through a series of input and output data, and the computational efficiency is far higher than that of the traditional numerical method. The invention can shorten the calculation task of the traditional computational fluid mechanics simulation for hours or even days to the second level, greatly accelerate the optimization iteration period of the appearance design of the aircraft, and obviously reduce the dependence and cost of research and development on wind tunnel tests and high-performance calculation.

Inventors

  • MENG DEYING

Assignees

  • 中国航天空气动力技术研究院

Dates

Publication Date
20260508
Application Date
20251225

Claims (8)

  1. 1. A fast prediction method of an airfoil bypass flow field based on a Fourier neural operator is characterized by comprising the following steps: Acquiring unsteady flow field data of airfoil streaming, wherein the flow field data at least comprises flow direction speed, normal speed and pressure; Constructing a Fourier neural operator model comprising a plurality of Fourier layers; inputting flow field data of the first N time steps as a model, and using flow field data of the (n+1) th time step as a label of the model output to construct a training data set; training a Fourier neural operator model using the training dataset to minimize a multivariate relative error loss function between the predicted flow field and the real flow field; And rapidly predicting the future flow field state according to the input flow field sequence by using the trained model.
  2. 2. The rapid prediction method of the airfoil bypass flow field based on the Fourier neural operator of claim 1, wherein the Fourier neural operator model comprises 4 Fourier layers, and the Fourier mode number of each Fourier layer is 12.
  3. 3. The method for rapidly predicting an airfoil bypass flow field based on a Fourier neural operator according to claim 1, wherein the multivariate relative error loss function is characterized by Is defined as follows: Where m is the size of the model output structure, Respectively represents the flow field flow direction speed, the normal speed and the pressure intensity predicted by the Fourier neural operator model, Representing the flow direction velocity, normal velocity and pressure of the real flow field respectively, Is the physical quantity of the i-th grid point.
  4. 4. The rapid prediction method of the airfoil bypass flow field based on the Fourier neural operator according to claim 1, wherein GELU functions are used as activation functions in the training process of the Fourier neural operator model.
  5. 5. The rapid prediction method of the airfoil bypass flow field based on the Fourier neural operator according to claim 1, wherein an Adam algorithm is used as an optimization algorithm in the training process of the Fourier neural operator model.
  6. 6. The rapid prediction method for the airfoil bypass flow field based on the Fourier neural operator is characterized in that prediction of future flow field states is achieved based on instantaneous prediction and iterative prediction strategies, the instantaneous prediction is achieved by directly mapping input of a current time step of a model into prediction output of a next time step, and the iterative prediction is achieved by taking the prediction output of the current time step of the model as a part of the input and is used for predicting flow fields of the next time step, so that sequence prediction of unsteady flow fields is achieved.
  7. 7. A terminal device, comprising: A memory for storing programs executed by the at least one processor; a processor for executing a computer program stored in a memory, implementing the method according to any of claims 1-6.
  8. 8. A computer readable storage medium, characterized in that the computer readable storage medium stores a computer program which, when run on a computer, causes the computer to perform the method according to any of claims 1-6.

Description

Rapid prediction method for wing profile bypass flow field based on Fourier neural operator Technical Field The invention belongs to the field of intelligent fluid mechanics, relates to a method for quickly predicting a flow field, and particularly relates to a method for quickly predicting an airfoil bypass flow field based on a Fourier neural operator. Background In the design and research and development process of an aircraft, flow field calculation and evaluation of the aircraft are key links for optimizing the appearance of the aircraft and improving the performance of the aircraft. However, although the traditional numerical calculation method and wind tunnel test method can provide high-precision flow field data, the calculation cost and experimental cost are too high, the time consumption is too long, and the requirements of rapid optimization and iteration on the appearance design of the aircraft in engineering practice are difficult to meet. Therefore, combining with advanced machine learning methods, achieving rapid prediction of aircraft flow fields is not only an important goal for aerodynamic development, but also an urgent need for aerospace technology development. Machine learning methods refer to algorithms in which a computer generates a predictive model based on a large amount of empirical data. Machine learning can accomplish classification and recursive tasks based on input data. The core idea of machine learning is to optimize model parameters through a training algorithm and training data, and predict unknown input data. In recent years, with the development of cloud computing and the increase of computing resources, the field of hydrodynamics has accumulated a large amount of flow data in different scenarios, which can be used for training of machine learning models. Therefore, many machine learning methods have been widely used in hydrodynamic problems, including turbulence model modeling, hypersonic flow transition, flow field prediction, and aerodynamic shape optimization of aircraft. The traditional computational fluid dynamics method and experimental method have high cost and long time consumption, and the flow prediction method based on machine learning can realize real-time, rapid and accurate prediction of the flow field, and remarkably shorten the time required by flow field calculation. The machine learning method has strong nonlinear computing capability and strong advantages when processing nonlinear problems, can capture nonlinear function relations of the flow field from a large amount of experimental or numerical simulation data, and learns nonlinear rules of the flow field, thereby reducing dependence on traditional experience models. Most neural network models in terms of flow field prediction learn the mapping between finite dimensional euclidean spaces. The generalization of these models is limited by different parameters, initial conditions and boundary conditions. The traditional flow field prediction model based on the convolutional neural network has weak parameterization capability, and is difficult to accurately predict flow fields with different parameters flexibly. Meanwhile, a flow field prediction model based on a convolutional neural network is usually fixed on the grid which is identical to training data, and the dimensions of input and output tensors are fixed and cannot be generalized to grids with different resolutions or different topological structures. Traditional flow field prediction models based on convolutional neural networks depend on stacking convolutional layers to enlarge the receptive field, so that prediction accuracy is improved. However, stacking convolution layers can greatly increase the parameter number of the model, reduce the model training efficiency and increase the hardware cost of the model training. Disclosure of Invention The invention solves the technical problems of overcoming the defects of the prior art, providing the rapid prediction method of the wing profile bypass flow field based on the Fourier neural operator, realizing rapid and accurate prediction of the wing profile bypass flow field and remarkably reducing the cost consumption of wind tunnel test and numerical simulation in the wing profile optimization process. The technical scheme of the invention is as follows: in a first aspect, the present invention provides a fast prediction method for an airfoil bypass flow field based on a fourier nerve operator, including: Acquiring unsteady flow field data of airfoil streaming, wherein the flow field data at least comprises flow direction speed, normal speed and pressure; Constructing a Fourier neural operator model comprising a plurality of Fourier layers; inputting flow field data of the first N time steps as a model, and using flow field data of the (n+1) th time step as a label of the model output to construct a training data set; training a Fourier neural operator model using the training dataset to minimize a