Method and system for aircraft flow field prediction using multi-region physically driven neural network

Abstract

The application discloses a method and system for aircraft flow field prediction of a multi-region physical driving neural network, and the method comprises the following steps: constructing a continuous region mask and a high-dimensional physical parameter sampling system, globally sampling the high-dimensional physical parameters by using a Latin hypercube sampling method, and combining a KMeans clustering algorithm to divide the space; inputting the space coordinates, the continuous region mask, the wall distance and the physical condition parameters into an AMPD model to generate a boundary layer mask, a vortex mask and a physical residual; inputting the boundary layer mask and the vortex mask into a loss function system driven by physical constraints to establish a multi-objective residual minimization loss function, which is used for training the AMPD model; training the AMPD model by using a curriculum learning training strategy based on the multi-objective residual minimization loss function and the physical residual; and reconstructing the wing surface flow field by using the trained AMPD model to complete the aircraft flow field prediction, and high-precision and high-efficiency intelligent prediction of the wing flow is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computational fluid dynamics and artificial intelligence, and particularly relates to an aircraft flow field prediction method and system based on a multi-region physically driven neural network. BACKGROUND

[0002] In aerospace engineering, accurate prediction of the flow characteristics around the wing is a key technical basis for aircraft aerodynamic design, performance optimization, and safety evaluation. Accurate understanding of flow field distribution, pressure distribution, lift and drag characteristics, and other parameters under different conditions is of great significance to improve aircraft design efficiency and flight safety. Although traditional computational fluid dynamics (CFD) methods can provide high prediction accuracy, they have the following significant drawbacks: high computational cost, single flow field calculation usually takes several hours to several days; complex grid generation, different wing geometries need to be re-modeled; difficult to meet the engineering needs of real-time prediction and large-scale multi-condition optimization. In recent years, deep learning technology has been increasingly applied in fluid mechanics, but existing methods have the following key technical problems:

[0003] (1) Lack of physical constraints leads to inconsistent predictions: existing patents such as CN202411325265A use U-Net to directly predict flow field images, but its pure data-driven mode lacks explicit constraints on basic physical laws such as conservation of mass, momentum, and energy, and the prediction results may violate the basic principles of fluid mechanics, especially in strong gradient regions such as boundary layers, shock waves, and vortices.

[0004] (2) Insufficient generalization ability for multiple airfoils: patent CN202510266457 proposes an airfoil flow field prediction method based on multi-task learning, which partially solves the gradient conflict problem between airfoil surface and volume area prediction through a multi-head decoder architecture and multi-task loss optimization strategy, but its shared feature extraction layer is still limited by the geometric distribution of the training airfoil family, and its generalization ability for unseen airfoils is limited, making it difficult to adapt to changes in airfoil thickness ratio and camber.

[0005] (3) Conflict between computational complexity and accuracy: patent CN202410781189A improves prediction accuracy by embedding Navier-Stokes equation constraints, but its multi-feature map joint processing mechanism significantly increases computational complexity and greatly increases network parameter quantity, affecting the feasibility of practical applications.

[0006] (4) Limited ability to handle complex flow phenomena: patent CN202411622813A proposes a semi-supervised learning flow field prediction method based on Gaussian Mixture Model (GMM), which reduces the dependence on labeled data through pseudo-label generation and sample selection mechanisms, but the accuracy still needs to be improved when dealing with complex nonlinear flow phenomena such as flow separation, reattachment, and shock boundary layer interference.

[0007] (5) Multiple physical region processing mechanism is imperfect: existing methods such as the machine learning model of CN202411065287A can better capture the evolution of wake geometry, but its training data is limited to fixed airfoils, and there is a lack of effective differentiation and processing mechanism for the characteristics of different physical regions such as solid interior, wall surface, near-wall fluid and free fluid.

[0008] In view of the deficiencies of the prior art, the following key technical problems need to be solved at present:

[0009] Physical consistency guarantee: Establish a neural network architecture that integrates the conservation laws of fluid mechanics to ensure the physical reasonableness of the prediction results;

[0010] Multi-airfoil adaptive capability: Develop a general modeling method with cross-airfoil generalization capability to reduce dependence on specific airfoil training data;

[0011] Accurate capture of complex flow phenomena: Improve the prediction accuracy of the model in complex regions such as shock waves, boundary layers and flow separation;

[0012] Multi-physical region coordination processing: Achieve effective differentiation and unified modeling of different physical region characteristics;

[0013] Balancing of computational efficiency and accuracy: On the premise of ensuring prediction accuracy, significantly improve computational efficiency to meet the real-time application needs of engineering;

[0014] Therefore, it is urgent to develop an intelligent flow field prediction technology that integrates physical prior knowledge, has multi-airfoil adaptability and supports accurate prediction of complex flow phenomena, in order to break through the technical bottlenecks of existing methods and meet the practical application needs of aerospace engineering. SUMMARY

[0015] In view of the above deficiencies in the prior art, the multi-region physical driving neural network aircraft flow field prediction method and system provided by the present application solves the technical problems of existing methods in terms of physical consistency, multi-airfoil generalization capability and complex flow phenomenon capture by integrating physical conservation law constraints, multi-region adaptive modeling and curriculum learning strategies.

[0016] In order to achieve the above invention purpose, the technical scheme adopted by the present application is: a multi-region physical driving neural network aircraft flow field prediction method, comprising the following steps:

[0017] S1, obtain the wing geometry parameters, generate the airfoil profile through the airfoil parameterization equation, and divide the multiple physical regions according to the continuous region mask;

[0018] S2, the Latin hypercube sampling method is used for global sampling of high-dimensional physical parameters, and the parameter space is divided into six representative sub-regions by combining the KMeans clustering algorithm, to obtain the physical condition parameters;

[0019] S3, the space coordinates, continuous region mask, wall distance and physical condition parameters are input into the AMPD (adaptive multi-region physical driving neural network) model, which is mapped to a high-dimensional feature space through dilated convolution and linear transformation, to generate boundary layer mask, vortex mask and physical residual;

[0020] S4, the boundary layer mask and vortex mask are input into the loss function system driven by physical constraints, and a multi-objective residual minimization loss function based on continuity equation, momentum conservation equation, energy equation and ideal gas state equation is established, which is used to train the AMPD model;

[0021] S5, based on the multi-objective residual minimization loss function and the physical residual, the AMPD model is trained by using the curriculum learning training strategy;

[0022] S6, the airfoil type, Mach number, Reynolds number, temperature and angle of attack are input into the trained AMPD model, and the wing surface flow field is reconstructed by the trained AMPD model to complete the aircraft flow field prediction.

[0023] Further, in S1, the method for dividing multiple physical regions according to the continuous region mask is specifically:

[0024] According to the spatial distance and topological discrimination, the airfoil internal region mask value is 1, the airfoil wall region mask value is 2, the near-wall fluid region mask value is 3.0 to 4.0, the external fluid region mask value is 4, the inlet region mask value is 5, the outlet region mask value is 6, and the free boundary region mask value is 7;

[0025] Wherein, the near-wall fluid region mask S ( x ) is described by a sigmoid function for smooth transition from 3.0 to 4.0;

[0026]

[0027] In the formula, is the wall distance, which is the nearest distance from the point x to the airfoil wall, is the adaptive near-wall region scale, is the smooth transition band parameter, is the sigmoid function.

[0028] The beneficial effects of the above-mentioned further solutions are as follows: This invention proposes a continuous region masking mechanism to achieve smooth division and adaptive processing of multiple physical regions such as solids, walls, near-wall fluids, and free fluids.

[0029] Furthermore: In S2, the objective function of the KMeans clustering algorithm The specific expression is:

[0030]

[0031] In the formula, It is the first One cluster, It is a cluster center. For physical parameters, The square of the Euclidean distance;

[0032] The high-dimensional physical parameters sampled are normalized to ensure that each physical parameter has the same scale during training.

[0033]

[0034] In the formula, Physical parameters The mean, Physical parameters standard deviation These are the normalized physical parameters.

[0035] Furthermore: In S3, the AMPD model includes an input embedding layer, a feature fusion backbone, and a physical constraint and explicit residual module connected in sequence;

[0036] Input embedding layer: Input spatial coordinates, continuous region masks, wall distances, and physical condition parameters. Spatial coordinates are extracted using a dilated convolution module for multi-scale feature extraction and mapped to a 32-dimensional feature space, with dilation rates of 1, 2, and 4, generating spatial coordinate features. Continuous region masks and wall distances are embedded into a 16-dimensional feature space through independent linear transformations, generating region mask features and wall distance features, respectively. Physical condition parameters are encoded into physical condition vectors using a multilayer perceptron, which are used to train the AMPD model using the course learning training strategy.

[0037] Feature fusion backbone: Spatial coordinate features, region mask features and wall distance features are combined to form a 64-dimensional fusion feature. The fusion feature is then input into the residual backbone network and attention mechanism module to obtain flow field data, including velocity field, density field, temperature field and pressure field.

[0038] The residual backbone network consists of six stacked residual neural blocks. Each residual neural block includes a bilinear transform layer, layer normalization, Leaky ReLU activation function, and residual connections. The attention mechanism module is used to dynamically adjust the weights of features at different spatial locations and improve the model's ability to represent complex flow fields through deep feature fusion.

[0039] The physical constraint and explicit residual module includes a gradient feature extraction unit, a physical region identification unit, a residual calculation unit, and a boundary condition enforcement unit;

[0040] Gradient feature extraction unit: Input flow field data and calculate the spatial gradient, gradient magnitude and higher-order derivative features of vorticity of each physical quantity through adaptive finite difference method;

[0041] Physical region identification unit: Dynamically identifies key physical regions in the flow field based on gradient characteristics, calculates boundary layer masks and eddy current masks, and provides region prior information for subsequent physical constraints;

[0042] Residual calculation unit: Calculates physical residuals using the basic governing equations of fluid mechanics. Physical residuals include residuals of the continuity equation, momentum equation, energy equation, and ideal gas law equation.

[0043] Boundary condition enforcement unit: Receives flow field data, physical residuals, continuous region mask, boundary layer mask, and eddy current mask. Inputs the spatial gradient, gradient magnitude, and higher-order derivative features of vorticity of each physical quantity calculated by the gradient feature extraction unit. Dynamically calculates gradients and divergences using an adaptive finite difference method, and enhances the physical residuals using an adaptive weight network. Corrects systemic biases in density, velocity, temperature, and pressure through residual calculations, and enforces corresponding physical constraints in different boundary regions. Based on the different physical regions corresponding to the continuous region mask values—inside the airfoil, on the airfoil wall, near-wall fluid, external fluid, inlet, outlet, and free boundary—it enforces corresponding constraints for no-slip boundary conditions, specified inflow conditions, pressure or velocity gradient conditions, and far-field boundary conditions, adjusting the flow field data and outputting the corrected flow field data.

[0044] The beneficial effects of the above-mentioned further solutions are as follows: the present invention constructs a physical constraint-driven neural network architecture, and ensures the physical consistency of flow field prediction through explicit residual correction.

[0045] Further: In S3, the boundary layer mask is calculated. The specific expression is:

[0046]

[0047] In the formula, For the velocity gradient, For gradient threshold, For characteristic attenuation length, d The distance to the solid boundary, k For gradient sensitivity parameters;

[0048] Calculate the eddy current mask The specific expression is:

[0049]

[0050] In the formula, For the preset threshold, The clustering label is 1 for valid clusters and 0 otherwise;

[0051] In the process of dynamically calculating the gradient using the adaptive finite difference method, the following formula is used:

[0052]

[0053]

[0054] In the formula, and For grid spacing, and For grid node indexing, f Let be the physical field function to be differentiated. For grid points ( The function value at () For grid points ( The function value at () For grid points ( The function value at () For grid points ( The function value at ().

[0055] Laplace operator The second derivative is approximated using a five-point scheme:

[0056]

[0057] The specific expressions for the boundary condition enforcement element that enforces the corresponding physical constraints in different boundary regions are as follows:

[0058]

[0059] In the formula, The velocity field within the solid region. This is the velocity field at the wall boundary, ensuring no-slip conditions between the solid surface and the wall.

[0060] Further, in S4, the multi-objective residual minimization loss function of the physical constraint driven loss function system The expression of the continuity loss is specifically:

[0061]

[0062] In the formula, is a loss term covering continuity loss, momentum loss, energy loss, ideal gas state loss, inlet region loss, boundary condition constraint and penalty loss, is a physical quantity gradient norm regularization term, is a parameter weight, is a physical quantity change rate constraint loss, is a change rate constraint weight parameter, is a gradient regularization weight parameter;

[0063] The continuity loss The expression of the continuity loss is specifically:

[0064]

[0065] In the formula, is an adaptive weighting function calculated based on the key region weight synthesis of the continuous region mask, the near-wall distance and the physical shock, is the total number of sampling points in the fluid domain, is the sampling point, is the fluid calculation domain, is the sampling point, is the continuity equation residual at the sampling point;

[0066] In response to the predicted velocity vector , the density , the temperature , the pressure P , and the spatial coordinates , the expression of the mass conservation residual is specifically:

[0067]

[0068] In the formula, is a partial derivative symbol, x is the x axis coordinate, y is the y axis coordinate, u is the x direction velocity component, v is the y direction velocity component;

[0069] Momentum loss includes x Momentum loss in the axial direction and y Momentum loss in the axial direction ;

[0070]

[0071]

[0072] wherein, and is the standard deviation of the momentum residual, is x the momentum residual in the axial direction, is y the momentum residual in the axial direction;

[0073]

[0074]

[0075] wherein, is the Laplacian operator, and is the viscous stress divergence term;

[0076]

[0077]

[0078] wherein, is the fluid dynamic viscosity;

[0079] Energy loss is specifically expressed as:

[0080]

[0081] wherein, is a normalization coefficient, is a numerical stability regularization term, is the energy conservation equation residual at the th sampling point;

[0082] Energy conservation equation residual is specifically expressed as:

[0083]

[0084] wherein, is the constant-pressure specific heat capacity, is the thermal conductivity, is the viscous dissipation term;

[0085]

[0086] wherein, is the coefficient of volume viscosity;

[0087] Ideal gas state loss is specifically expressed as:

[0088]

[0089] wherein, is the ideal gas state equation residual;

[0090]

[0091] wherein, R is the ideal gas constant, is the density value at position x is the product of the ideal gas constant and temperature, is the predicted pressure value at position x Inlet region loss is specifically expressed as:

[0092]

[0093]

[0094] wherein, is the standard inlet physical parameter, is the number of inlet nodes, is the inlet boundary region, q is a single physical quantity in the physical quantity variable set is the predicted physical quantity value at the th inlet node; The boundary condition constraint produces a continuous value weight through a tanh activation function, which is used to reflect the relative importance of each node in the abnormal physical region smoothly, and the continuous value weight

[0095] is specifically expressed as:

[0096]

[0097] wherein, is the shock region weight coefficient, is the vortex region weight coefficient, is the shear region weight coefficient, is the shock recognition mask, is the vortex recognition mask, is the shear layer recognition mask.​​​

[0098] penalty loss The expression of the penalty loss is specifically:

[0099]

[0100] In the formula, is a safety threshold, is a penalty factor, is a ReLU activation function.

[0101] Further, in S5, the method for training the AMPD model by adopting the curriculum learning training strategy is specifically:

[0102] The first stage: the attack angle range is limited within 20% of the total range, the upper limit of the Reynolds number is 5*10 5 , the upper limit of the Mach number is 0.3, the number of geometric samples is 15, and the number of physical conditions is 5;

[0103] The second stage: the attack angle range is expanded to 40%, the upper limit of the Reynolds number is 6*10 5 , the upper limit of the Mach number is 0.5, the number of geometric samples is 18, and the number of physical conditions is 6;

[0104] The third stage: the attack angle range is expanded to 60%, the upper limit of the Reynolds number is 7*10 5 , the upper limit of the Mach number is 0.7, the number of geometric samples is 20, and the number of physical conditions is 7;

[0105] The fourth stage: the attack angle range is expanded to 80%, the upper limit of the Reynolds number is 8*10 5 , the upper limit of the Mach number is 0.85, the number of geometric samples is 23, and the number of physical conditions is 8;

[0106] The fifth stage: the attack angle range reaches 100%, the upper limit of the Reynolds number is 10 6 , the upper limit of the Mach number is 0.95, the number of geometric samples is 25, and the number of physical conditions is 10;

[0107] In each stage, a loss threshold is set, and the initial value is 500, which is decreased to 10 according to the increase of the training period, and in each stage, when the validation loss is lower than the loss threshold and the training period is at least 20, the next stage is entered.

[0108] The beneficial effects of the above further scheme are that the curriculum learning training strategy is designed to gradually improve the prediction ability of the model for complex flow phenomena.

[0109] The aircraft flow field prediction system of the multi-region physical driving neural network comprises:

[0110] A data preprocessing module is configured to generate an airfoil profile, construct a continuous region mask, and implement high-dimensional physical parameter sampling.

[0111] a neural network training module comprising an AMPD model architecture, a physically constrained loss function system, and a curriculum learning training strategy;

[0112] a flow field prediction module comprising the trained AMPD model, inputting airfoil type, Mach number, Reynolds number, temperature, and angle of attack, and outputting complete flow field distribution data;

[0113] a result visualization module providing velocity vector diagrams, pressure contour maps, density contour maps, velocity amplitude maps, and surface pressure coefficient curve display modes.

[0114] The present application has the following advantages:

[0115] (1) The present application provides an aircraft flow field prediction method and system based on a multi-region physically driven neural network, proposes a continuous region mask and a high-dimensional physical parameter sampling system, realizes smooth division of multiple physical regions through a sigmoid function, establishes a Latin hypercube sampling and KMeans clustering strategy, globally samples and spatially divides parameters such as Mach number, Reynolds number, and temperature, constructs an AMPD model comprising an input embedding layer, a feature fusion backbone, and a physical constraint and explicit residual module, designs a physically constrained loss function system, and combines an adaptive weighting mechanism to strengthen key flow regions, and implements a five-stage increasing complexity curriculum learning training strategy. The present application solves the technical problems of existing methods in terms of physical consistency, multi-airfoil generalization ability, and complex flow phenomenon capture by fusing physical constraints, multi-region adaptive modeling, and curriculum learning, realizes high-precision and high-efficiency intelligent prediction of wing flow, and has a speedup of more than 3,000 times compared to traditional CFD methods, and the prediction accuracy meets the requirements of engineering applications.

[0116] (2) The adaptive multi-region physically driven neural network method of the present application exhibits excellent prediction performance under various airfoil types and different flow conditions:

[0117] Prediction accuracy: average error of pressure coefficient distribution < 3%, lift and drag coefficient error < 5%, meeting the requirements of engineering applications;

[0118] Physical consistency: strictly satisfies the conservation laws of fluid mechanics, and has no non-physical phenomena;

[0119] Computing efficiency: more than 3,000 times faster than traditional CFD methods, with a single prediction time of less than 10 seconds;

[0120] Strong adaptability: supports various airfoil geometries and a wide range of working conditions, with outstanding generalization ability;

[0121] Complex flow capture: accurately predicts complex flow phenomena such as separation and shock, with significant technological advancement.

[0122] The method provides an efficient, accurate and intelligent solution for rapid flow field analysis in the field of aerospace, and has important engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0123] Figure 1 A flow chart of the aircraft flow field prediction method of the multi-region physical driving neural network.

[0124] Figure 2 A schematic diagram of the aircraft flow field prediction method of the multi-region physical driving neural network.

[0125] Figure 3 The aircraft flow field prediction system of the multi-region physical driving neural network. DETAILED DESCRIPTION

[0126] The specific embodiments of the present application are described below to facilitate understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all applications utilizing the concept of the present application are within the scope of protection.

[0127] As shown in Figure 1 and Figure 2 In one embodiment of the present application, the aircraft flow field prediction method of the multi-region physical driving neural network comprises the following steps:

[0128] S1, obtain the wing geometry parameters, generate the airfoil profile through the airfoil parameterization equation, and divide the multi-physical region according to the continuous region mask;

[0129] S2, use the Latin hypercube sampling method to perform global sampling on the high-dimensional physical parameters, and combine the KMeans clustering algorithm to divide the parameter space into six representative sub-regions to obtain the physical condition parameters;

[0130] S3, input the spatial coordinates, the continuous region mask, the wall distance and the physical condition parameters into the AMPD model, map them to the high-dimensional feature space through the dilated convolution and linear transformation, generate the boundary layer mask, the vortex mask and the physical residual;

[0131] S4, input the boundary layer mask and the vortex mask into the loss function system driven by the physical constraint, establish a multi-objective residual minimization loss function based on the continuity equation, the momentum conservation equation, the energy equation and the ideal gas state equation, and use it to train the AMPD model;

[0132] S5, based on the multi-objective residual minimization loss function and the physical residual, use the curriculum learning training strategy to train the AMPD model;

[0133] S6, input the airfoil type, Mach number, Reynolds number, temperature and attack angle into the trained AMPD model, reconstruct the wing surface flow field through the trained AMPD model, and complete the aircraft flow field prediction.

[0134] S1 is specifically:

[0135] Airfoil geometry generation and grid discretization: taking NACA series wings as the flow field geometry basis, the airfoil profile is defined by a parameterized equation, and a standardized spatial point cloud is generated by two-dimensional uniform grid discretization in the generalized coordinate domain. The typical grid scale is 100x100 nodes;

[0136] Accurate division of multiple physical regions: for different attack angles and geometric configurations, multiple region division is realized through spatial distance and topological discrimination, and the specific allocation rules are shown in Table 1:

[0137] Table 1 Mask allocation rules

[0138]

[0139] In S1, the method of dividing multiple physical regions according to the continuous region mask is specifically:

[0140] According to the spatial distance and topological discrimination, the airfoil internal region mask value is 1, the airfoil wall region mask value is 2, the near-wall fluid region mask value is 3.0 to 4.0, the external fluid region mask value is 4, the inlet region mask value is 5, the outlet region mask value is 6, and the free boundary region mask value is 7;

[0141] Wherein, the near-wall fluid region mask S ( x ) is described by a sigmoid function to describe the smooth transition of 3.0 to 4.0;

[0142]

[0143] In the formula, is the wall distance, which is the nearest distance from the point x to the airfoil wall, is the adaptive near-wall region scale, is the smooth transition band parameter, is the sigmoid function.

[0144] This embodiment distinguishes the near-wall fluid and free fluid region by introducing a smooth function form to determine whether each node is close to the wall through this mechanism. Through the parameterized mask generation mechanism, the geometric changes of different NACA series airfoils are adaptively processed. For different airfoil parameters (maximum thickness, camber, camber position), the wall distance function and the adaptive near-wall region scale are dynamically adjusted to achieve accurate adaptation of various airfoil geometries.

[0145] In S2, the high-dimensional physical parameters are the basis for describing the physical characteristics of the flow field, as shown in Table 2:

[0146] Table 2 Physical parameters and parameter ranges

[0147]

[0148] Latin hypercube sampling method implementation: the Latin hypercube sampling method is used to sample , and to generate multiple sets of physical boundary condition samples to cover different flow field conditions.

[0149] Among them, determines the strength of the compression effect such as shock wave and expansion wave in the flow field, which has a significant impact on flow separation and pressure distribution, especially at high angles of attack. Re represents the ratio of inertial force to viscous force of the fluid, which directly affects the behavior of the boundary layer and the vortex characteristics. In the range from low to high Reynolds number, the viscous effect of the flow field and the formation of separation bubbles change with Re. Indirectly regulate the compressibility and viscosity characteristics of the flow field, especially the thermodynamic behavior of the gas at different temperatures.

[0150] Physical parameter sampling uses the Latin hypercube sampling method: Mach number Ma is sampled in the range of [0.3, 0.95] with 50 values; Reynolds number is sampled in the range of [1×10 4 , 1×10 6 ] with 50 values; Free stream temperature is sampled in the range of [250K, 350K] with 30 values, a total of 75,000 working condition combinations. The KMeans clustering algorithm is used to divide the parameter space into 6 representative sub-regions, and the cluster centers correspond to: low speed low Reynolds number, medium speed medium Reynolds number, high speed low Reynolds number, low speed high Reynolds number, high speed high Reynolds number, transonic complex working condition, etc. Typical flow states.

[0151] The KMeans clustering algorithm is used to generate physical condition parameters, and the expression of the objective function of the KMeans clustering algorithm is as follows:

[0152]

[0153] In the formula, It is the first One cluster, It is a cluster center. For physical parameters, The square of the Euclidean distance is used to divide the parameter space into six representative sub-regions, effectively covering the flow field behavior from laminar weak compressibility to turbulent strong compressibility, providing diverse training samples for the model.

[0154] The high-dimensional physical parameters sampled are normalized to ensure that each physical parameter has the same scale during training.

[0155]

[0156] In the formula, Physical parameters The mean, Physical parameters standard deviation These are the normalized physical parameters.

[0157] The AMPD model consists of an input embedding layer, a feature fusion backbone, and a physical constraint and explicit residual module connected in sequence.

[0158] Input embedding layer: Input spatial coordinates, continuous region mask, wall distance, and physical condition parameters;

[0159] Spatial coordinates are extracted and mapped to a 32-dimensional feature space through a dilated convolution module with dilation rates of 1, 2 and 4. Local and global spatial information are fused through multi-scale receptive fields to generate spatial coordinate features.

[0160] The continuous region mask is embedded into the 16-dimensional feature space through independent linear transformations, preserving the continuity features of different physical regions (solid, wall, near-wall fluid, free fluid);

[0161] After normalization, the wall distance is embedded into the 16-dimensional feature space through linear transformation to provide geometric prior information for boundary layer recognition.

[0162] Physical condition parameters are encoded into physical condition vectors through a multilayer perceptron, which are used to train the AMPD model using course learning training strategies, and guide the AMPD model to adapt to different working conditions.

[0163] Feature fusion backbone: Spatial coordinate features, region mask features and wall distance features are combined to form a 64-dimensional fusion feature. The fusion feature is then input into the residual backbone network and attention mechanism module to obtain flow field data, including velocity field, density field, temperature field and pressure field.

[0164] The residual backbone network includes six stacked residual neural blocks, each of which includes a bilinear transformation layer (64->128->64 dimensions), layer normalization, a Leaky ReLU activation function (negative slope 0.01), and a residual connection, which solves the gradient disappearance problem of a deep network and improves the representation ability of the model; the attention mechanism module is used for dynamically adjusting the weights of features at different spatial positions, and the representation ability of the model for a complex flow field is improved through deep feature fusion;

[0165] The physical constraint and explicit residual module fuses the fluid mechanics conservation law constraint to ensure the physical consistency and rationality of the neural network prediction result, and the physical constraint and explicit residual module includes a gradient feature extraction unit, a physical region identification unit, a residual calculation unit, and a boundary condition forcing unit;

[0166] In the embodiment, the physical constraint and explicit residual module calculates the spatial gradient and high-order derivative features of the physical field through an adaptive finite difference method, dynamically identifies key flow regions such as boundary layers and vortices based on a velocity gradient and a DBSCAN clustering algorithm, calculates physical residuals of continuity, momentum, energy, and state equations using basic fluid mechanics control equations, and forces the execution of corresponding physical constraint conditions in different boundary regions to ensure that the prediction result strictly satisfies the fluid mechanics conservation law.

[0167] The gradient feature extraction unit: input flow field data, calculate the spatial gradient, gradient amplitude, and vorticity high-order derivative features of each physical quantity through an adaptive finite difference method;

[0168] In the embodiment, for any physical field , the gradient amplitude is calculated by the following formula:

[0169]

[0170] In the formula, the spatial partial derivative and are solved by a numerical solution method based on a local least square method through a linear equation system.

[0171] The physical region identification unit: dynamically identifies key physical regions in the flow field according to the gradient features, calculates a boundary layer mask and a vortex mask, and provides region prior information for subsequent physical constraints;

[0172] The residual calculation unit: calculates physical residuals using basic fluid mechanics control equations, and the physical residuals include continuity equation residuals, momentum equation residuals, energy equation residuals, and ideal gas state equation residuals;

[0173] In this embodiment, the control equation adopted by the residual calculation unit is consistent with the physical constraint equation defined by the loss function in S4, ensuring the physical consistency of the training process and the inference process. This unit identifies the part of the prediction result that violates the physical law through residual analysis, providing a quantitative basis for subsequent correction.

[0174] The boundary condition enforcement unit receives the flow field data, physical residual, continuous region mask, boundary layer mask, and vortex mask, and inputs the spatial gradients of each physical quantity calculated by the gradient feature extraction unit, the gradient amplitude, and the high-order derivative features of the vorticity, dynamically calculates the gradient and divergence through the adaptive finite difference method, and combines the adaptive weight network to enhance the physical residual. The system bias of density, velocity, temperature, and pressure is corrected through residual calculation, and the corresponding physical constraint conditions are enforced in different boundary regions. According to the different physical regions of the inside of the airfoil, the airfoil wall surface, the near-wall fluid, the external fluid, the inlet, the outlet, and the free boundary corresponding to the continuous region mask value, the corresponding constraints of the no-slip boundary condition, the specified inflow condition, the pressure or velocity gradient condition, and the far-field boundary condition are enforced respectively to adjust the flow field data. The corrected flow field data, including the density field, velocity field, temperature field, and pressure field, are output to ensure that the physical constraint conditions of each boundary region are satisfied, eliminate the system bias in flow field prediction, and improve the physical consistency and accuracy of the prediction result.

[0175] The system bias of density, velocity, temperature, and pressure is corrected through residual calculation, and the corresponding physical constraint conditions are enforced in different boundary regions. The flow field data is adjusted to ensure the consistency of multiple physical regions, especially in the wall The no-slip condition is enforced, and velocity and pressure constraints are applied at the boundaries , , to generate a flow field prediction that conforms to the physical law.

[0176] In S3, the expression for calculating the boundary layer mask is as follows:

[0177]

[0178] where is the velocity gradient, is the gradient threshold, which is set to the 95th percentile in this embodiment, is the characteristic decay length, which is estimated based on the scale of the flow field, d is the distance to the solid boundary, kis the gradient sensitivity parameter; the vortex mask identifies potential vortex points by high value regions of velocity gradient (over 90% quantile) and applies DBSCAN clustering algorithm (set neighborhood radius 50 mm, minimum sample number 10) to group connected regions, and the vortex mask is calculated The expression of the formula is specifically:

[0179]

[0180] In the formula, is a preset threshold value, is a cluster label, and the valid cluster is 1, and otherwise 0;

[0181] In this embodiment, through the weighting mechanism, these masks guide the attention of the model to the key flow regions. Subsequently, these masks are used in the physical constraint driven loss function system, and by adaptively weighting the residual term, the contributions of high gradient regions such as shock waves, vortices and boundary layers are amplified, thereby improving the performance of the model in capturing extreme nonlinear phenomena.

[0182] The core numerical calculation methods adopted by the physical constraint and explicit residual module include the adaptive finite difference method, the Laplace operator, the viscous stress divergence and the boundary region enforcement, which provide a numerical basis for the accurate calculation of the physical residual, and are specifically as follows:

[0183] In the process of dynamically calculating the gradient by the adaptive finite difference method, the following formula is used:

[0184]

[0185]

[0186] In the formula, and are grid spacings, and are grid node indices, f is a physical field function to be derived, is the function value at the grid point ( ), is the function value at the grid point ( ), is the function value at the grid point ( ), is the function value at the grid point ( );

[0187] Laplace operator The five-point format is used to approximate the second-order derivative:

[0188]

[0189] The boundary condition enforcement unit enforces the expressions of the corresponding physical constraint conditions in different boundary regions, specifically:

[0190]

[0191] wherein, is the velocity field in the solid region, is the velocity field at the wall boundary, ensuring the no-slip condition between the solid surface and the wall.

[0192] Viscous stress divergence: the divergence of the stress tensor components is calculated using the finite difference method :

[0193]

[0194] wherein, is the shear stress component, representing the stress acting in the x direction on a plane perpendicular to the y axis, is the normal stress component, representing the stress acting in the x direction on a plane perpendicular to the x axis, is the stress vector in the x direction;

[0195] Through the synergistic effect of the above four methods, the physical constraint and explicit residual module can effectively identify the key physical regions in the flow field, calculate accurate physical residuals, and enforce boundary conditions, ultimately output the predicted results of the flow field that satisfy the basic laws of fluid mechanics, significantly improving the physical consistency and prediction accuracy of the model in predicting complex flow phenomena.

[0196] In S4, the multi-objective residual minimization loss function of the physical constraint driven loss function system includes:

[0197] Continuity loss The expression is specifically:

[0198]

[0199] wherein, is an adaptive weighting function calculated based on the key region weight synthesis of the continuity region mask, near-wall distance, and physical shock, is the total number of sampling points in the fluid domain, is the th sampling point, is the fluid calculation domain, is the th sampling point, is the continuity equation residual at the th sampling point;

[0200] In response to the predicted velocity vector , the density is , the temperature is , the pressure is P , the spatial coordinates are , the mass conservation residual is , and the expression of the mass conservation residual is specifically:

[0201]

[0202] In the formula, is a partial derivative symbol, x is an x axis coordinate, y is an y axis coordinate, u is an x directional velocity component, v is an y directional velocity component;

[0203] The momentum loss includes x momentum loss in the axis direction and y momentum loss in the axis direction;

[0204]

[0205]

[0206] In the formula, and are standard deviations of momentum residuals, used for normalization processing of different physical quantity scales, avoiding loss term weight deviation, and improving training stability, is a momentum residual in the x axis direction, is a momentum residual in the y axis direction;

[0207]

[0208]

[0209] In the formula, is a Laplace operator, and are viscous stress divergence terms;

[0210]

[0211]

[0212] In the formula, is a fluid dynamic viscosity;

[0213] Energy loss The expression of the energy loss is specifically:

[0214]

[0215] In the formula, is a normalization coefficient, is a numerical stability regularization term, is an energy conservation equation residual at the i-th sampling point; Energy conservation equation residual

[0216] The expression of the energy conservation equation residual is specifically:

[0217]

[0218] In the formula, is a constant-pressure specific heat capacity, is a thermal conductivity, is a viscous dissipation term;

[0219]

[0220] In the formula, is a bulk viscosity coefficient;

[0221] Ideal gas state loss The expression of the ideal gas state loss is specifically:

[0222]

[0223] In the formula, is an ideal gas state equation residual, and the ideal gas state loss serves as a constraint to ensure that the predicted state variable is physically reasonable;

[0224]

[0225] In the formula, R is an ideal gas constant, is a density value at the position i, x is a product of the ideal gas constant and the temperature, is a predicted pressure value at the position i; x Inlet region loss The expression of the inlet region loss is specifically:

[0226]

[0227]

[0228] In the formula, is a standard inlet physical parameter, is the number of inlet nodes, is an inlet boundary region,​​​q is a single physical quantity in the set of physical quantity variables , is a predicted physical quantity value at the th inlet node;

[0229] The boundary condition constraint produces a continuous value weight through a tanh activation function, which is used to reflect the relative importance of each node in the abnormal physical region smoothly, and the expression of the continuous value weight is specifically:

[0230]

[0231] In the formula, is a shock region weight coefficient, is a vortex region weight coefficient, is a shear layer weight coefficient, is a shock identification mask, is a vortex identification mask, is a shear layer identification mask; this mechanism amplifies the residual of the key region and improves the attention of the vortex and boundary layer during training.

[0232] The expression of the penalty loss is specifically:

[0233]

[0234] In the formula, is a safety threshold, is a penalty factor, is a ReLU activation function.

[0235] To further improve the robustness and convergence stability of training, an exponentially moving average (EMA) mechanism is introduced to perform dynamic normalization on the statistical values of each loss, avoiding the influence of local extreme values on training. In addition, for gradient anomaly regions such as shock regions, the residual variance is calculated as a loss stability indicator to prevent training from falling into distortion divergence. In summary, the expression of the multi-objective residual minimization loss function of the loss function system driven by physical constraints is specifically:

[0236]

[0237] In the formula, is a loss term covering the continuity loss, momentum loss, energy loss, ideal gas state loss, inlet region loss, boundary condition constraint and penalty loss, is a physical quantity gradient norm regularization term, is a parameter weight, is a physical quantity change rate constraint loss, is a change rate constraint weight parameter, to regularize the weight parameters of the gradient;

[0238] In S5, the method for training the AMPD model using the curriculum learning training strategy is specifically:

[0239] First stage: the angle of attack range is limited within 20% of the total range, the upper limit of the Reynolds number is 5x10 5 , the upper limit of the Mach number is 0.3, the number of geometric samples is 15, and the number of physical conditions is 5;

[0240] Second stage: the angle of attack range is expanded to 40%, the upper limit of the Reynolds number is 6x10 5 , the upper limit of the Mach number is 0.5, the number of geometric samples is 18, and the number of physical conditions is 6;

[0241] Third stage: the angle of attack range is expanded to 60%, the upper limit of the Reynolds number is 7x10 5 , the upper limit of the Mach number is 0.7, the number of geometric samples is 20, and the number of physical conditions is 7;

[0242] Fourth stage: the angle of attack range is expanded to 80%, the upper limit of the Reynolds number is 8x10 5 , the upper limit of the Mach number is 0.85, the number of geometric samples is 23, and the number of physical conditions is 8;

[0243] Fifth stage: the angle of attack range reaches 100%, the upper limit of the Reynolds number is 10 6 , the upper limit of the Mach number is 0.95, the number of geometric samples is 25, and the number of physical conditions is 10;

[0244] In each stage, a loss threshold is set, which is initially 500 and decreases to 10 as the training period increases. In each stage, when the validation loss is lower than the loss threshold and the training period is at least 20, the next stage is entered.

[0245] In this embodiment, the curriculum learning training strategy further includes:

[0246] Stage early stopping mechanism: a loss threshold is set for each stage (decreasing from 500 in the first stage to 10 in the fifth stage), and switching is triggered when the validation loss is lower than the threshold and the training period is at least 20; the new stage retains part of the samples of the previous stage, and the period is dynamically adjusted according to the validation loss (extended by 2 periods when the loss exceeds 500); global early stopping terminates training when overall performance stagnates;

[0247] Sample selection and refresh strategy: the training set and the validation set are generated by different random seeds to avoid overlap; geometric samples are uniformly sampled from LHS results according to the angle of attack threshold in each stage, and physical conditions are randomly selected from the candidate pool that meets the upper limit of the Reynolds number and the Mach number; samples are refreshed every 5 periods to prevent overfitting;

[0248] Physical clustering synergistic optimization: Each stage sample not only increases difficulty, but also covers clustering diversity, avoids single condition overfitting, and enhances model robustness.

[0249] As Figure 3 shown, the aircraft flow field prediction system of the multi-region physical driving neural network is applied to the aircraft flow field prediction method of the multi-region physical driving neural network, and the system comprises:

[0250] A data preprocessing module is configured to generate an airfoil profile, construct a continuous region mask, and implement high-dimensional physical parameter sampling.

[0251] A neural network training module includes an AMPD model architecture, a physically constrained loss function system, and a curriculum learning training strategy.

[0252] A flow field prediction module includes a trained AMPD model, which inputs airfoil type, Mach number, Reynolds number, temperature, and attack angle, and outputs complete flow field distribution data.

[0253] A result visualization module provides velocity vector diagrams, pressure contour maps, density contour maps, velocity amplitude maps, and surface pressure coefficient curve display methods.

[0254] The neural network training module adopts a distributed deployment architecture, specifically including:

[0255] Service architecture: a microservice architecture based on RESTful API design provides flow field prediction service interfaces through standard HTTP / HTTPS protocols, supports JSON format data exchange, and facilitates integration with different client applications and third-party systems;

[0256] Concurrent processing capability: supports multi-user concurrent access, a single server node can handle not less than 50 concurrent requests at the same time, and through request queue management and asynchronous processing mechanism ensures that the system response time is controlled within 2 seconds;

[0257] Load balancing and fault tolerance: configure an automatic load balancer to intelligently distribute requests based on server load, and when a single node fails, it can automatically switch to a standby node within 30 seconds to ensure service continuity; the system has automatic monitoring and fault detection functions, supports automatic restart and recovery of services;

[0258] Extensibility design: adopts containerized deployment, supports horizontal expansion according to load demand, and can dynamically add or reduce computing nodes to ensure stable service quality during peak periods.

[0259] To verify the flow field prediction results of the trained AMPD model, the following cases are provided, including the prediction performance of different airfoils under different working conditions:

[0260] Case one:

[0261] Test case: NACA 0012 airfoil, Ma = 0.5, Re = 3 x 10 5 , a = 6°, = 288 K.

[0262] Prediction accuracy verification: Compared with CFD (Fluent solver, k- w SST turbulence model) results, the average relative error of pressure coefficient distribution is 2.1%; the error of skin friction coefficient is less than 4.5%; the predicted value of lift coefficient CL is 0.667, the CFD reference value is 0.655, and the relative error is 1.8%; the predicted value of drag coefficient CD is 0.0089, the CFD reference value is 0.0085, and the relative error is 4.7%. Flow field feature capture: successfully predicts the accelerated area on the upper surface and the stagnation area on the lower surface of the airfoil; accurately captures the leading edge stagnation point and the trailing edge wake area; the predicted boundary layer thickness is consistent with the theoretical calculation, and the error of 99% boundary layer thickness is less than 3%.

[0263] Computational efficiency: the time consumption of a single prediction is 5.8 seconds, which is more than 3,700 times faster than the traditional CFD method (about 6 hours).

[0264] Case two:

[0265] This embodiment verifies the prediction performance of NACA 2412 airfoil under different working conditions. The maximum thickness of the airfoil is 12%, and the maximum camber is 2% at 40% chord length.

[0266] The test working condition is set as: Mach number 0.6, Reynolds number 3 x 10 5 , attack angle 6°, free stream temperature 288 K. The training process follows a five-stage course learning strategy, and the loss convergence of each stage is: 428.7, 156.3, 78.9, 38.2, and 7.4, respectively.

[0267] The comparison of prediction results and wind tunnel experimental data shows that: the average relative error of airfoil surface pressure coefficient distribution is 2.8%; the prediction error of lift coefficient is 1.5%; the prediction error of drag coefficient is 4.2%. In the key positions such as the leading edge stagnation point, the upper surface pressure peak point and the trailing edge, the prediction accuracy is maintained within 3%.

[0268] In terms of flow separation prediction, this method successfully captures the laminar separation bubble phenomenon near the trailing edge of the upper surface of the airfoil, and the error of separation point position compared with experimental observation is less than 2% chord length, and the error of reattachment point position is less than 3% chord length.

[0269] Case three:

[0270] This embodiment tests the prediction ability of complex flow under high attack angle and large angle of attack conditions. NACA 0012 airfoil is selected, the attack angle is set to 15°, the Mach number is 0.4, and the Reynolds number is 5 x 105 .

[0271] In this case, a large-scale flow separation occurs on the upper surface of the airfoil, forming a clear separation vortex structure. The prediction results show that the method accurately captures the separation point location (at 45% chord length), with an error of less than 3% chord length compared with the CFD calculation results. The backflow phenomenon and vorticity distribution in the separation zone are also accurately predicted.

[0272] The error of the predicted lift coefficient compared with the experimental data is 3.7%, considering the strong nonlinearity and instability of the flow under high attack angle conditions, this level of accuracy indicates that the method has good complex flow capturing ability.

[0273] In terms of boundary layer prediction, the method accurately predicts the trend of boundary layer thickness along the chord length through the accurate division of continuous regional masks, with a prediction error of 99% boundary layer thickness within 5%.

[0274] Case Four:

[0275] This embodiment verifies the shock wave capturing ability under near transonic conditions. The test condition is NACA 0012 airfoil, Mach number 0.85, Reynolds number 8x10 5 , attack angle 2°.

[0276] Under this condition, a supersonic region is formed on the upper surface of the airfoil, and a positive shock wave is generated at the rear. The prediction results show that the shock wave position is located at 70% chord length, with an error of 1.8% chord length compared with the experimental data. The pressure jump and density change before and after the shock wave are accurately predicted, with a pressure ratio prediction error of less than 5%.

[0277] The shock wave boundary layer interference phenomenon is also effectively captured, and the boundary layer thickening and possible local separation induced by the shock wave are reflected in the prediction results. This indicates that the method has good numerical stability and physical accuracy in dealing with strong discontinuities such as shock waves.

[0278] Through the verification of the above four cases, the adaptive multi-region physical driving neural network method of the present application has excellent prediction performance under various airfoils and different flow conditions:

[0279] (1) Prediction accuracy: average error of pressure coefficient distribution <3%, error of lift and drag coefficients <5%, meeting the requirements of engineering applications;

[0280] (2) Physical consistency: strictly satisfies the conservation laws of fluid mechanics, without non-physical phenomena;

[0281] (3) Calculation efficiency: more than 3,000 times faster than traditional CFD methods, with a single prediction time of less than 10 seconds;

[0282] (4) Strong adaptability: support multiple airfoil geometries and wide operating conditions, generalization ability is outstanding;

[0283] (5) Complex flow capture: accurately predict complex flow phenomena such as separation, shock, etc., and the technology is advanced.

[0284] The method provides an efficient, accurate and intelligent solution for rapid flow field analysis in the field of aerospace, and has important engineering application value.

[0285] In the description of the present application, it should be understood that the orientation or positional relationship indicated by the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features. Therefore, the features limited by "first", "second", "third" can explicitly or implicitly include one or more features.

Claims

1. A method for aircraft flow field prediction using a multi-region physically driven neural network, characterized in that, The method comprises the following steps: S1, acquiring wing geometric parameters, generating an airfoil profile through an airfoil parameterization equation, and dividing multiple physical regions according to a continuous region mask; S2, performing global sampling on high-dimensional physical parameters by using a Latin hypercube sampling method, and dividing the parameter space into six representative sub-regions by combining a KMeans clustering algorithm to obtain physical condition parameters; S3, inputting the spatial coordinates, the continuous region mask, the wall distance and the physical condition parameters into the AMPD model, mapping to a high-dimensional feature space through dilated convolution and linear transformation to generate a boundary layer mask, a vortex mask and a physical residual; S4, inputting the boundary layer mask and the vortex mask into a loss function system driven by physical constraints, establishing a multi-objective residual minimization loss function based on the continuity equation, the momentum conservation equation, the energy equation and the ideal gas state equation, and using the loss function for training the AMPD model; S5, training the AMPD model by using a curriculum learning training strategy based on the multi-objective residual minimization loss function and the physical residual; S6, inputting the airfoil type, the Mach number, the Reynolds number, the temperature and the angle of attack into the trained AMPD model, reconstructing the wing surface flow field through the trained AMPD model, and completing the aircraft flow field prediction; In S1, the method for dividing multiple physical regions according to the continuous region mask is as follows: The multiple physical regions are divided according to the spatial distance and topological discrimination, the mask value of the internal region of the airfoil is 1, the mask value of the airfoil wall region is 2, the mask value of the near-wall fluid region is 3.0 to 4.0, the mask value of the external fluid region is 4, the mask value of the inlet region is 5, the mask value of the outlet region is 6, and the mask value of the free boundary region is 7; wherein the near-wall fluid region mask S ( x ) employs a sigmoid function to describe a smooth transition from 3.0 to 4.

0. wherein is the wall distance, which is the value of the point x is the minimum distance to the airfoil wall, is the adaptive near-wall region scale, is the smooth transition band parameter, is the sigmoid function; In S3, the boundary layer mask is calculated The expression of the boundary layer mask is given by wherein is a velocity gradient, is a gradient threshold, is a characteristic decay length, d is a distance to a solid boundary, k is a gradient sensitivity parameter; Computing a vortex mask The expression for the vortex mask is given by In the formula, is a preset threshold value, is a cluster label, 1 for a valid cluster, and 0 otherwise; The AMPD model comprises a boundary condition forced unit, and the following formula is used in the process of dynamically calculating the gradient by using the adaptive finite difference method: wherein and is the grid spacing, and is the grid node index, f is the physical field function to be differentiated, is the function value at the grid point , is the function value at the grid point , is the function value at the grid point , is the function value at the grid point ; laplacian The second order derivatives are approximated using a five-point stencil: The expression of the boundary condition forced unit for forcibly executing the corresponding physical constraint condition in different boundary regions is as follows: where is the velocity field in the solid region, is the velocity field at the wall boundary, ensuring the no-slip condition between the solid surface and the wall.

2. The method of claim 1, wherein, In S2, the objective function of the KMeans clustering algorithm The expression of the objective function is specifically: wherein is the first cluster, is the cluster center, is the physical parameter, is the squared Euclidean distance; The sampled high-dimensional physical parameters are subjected to distribution normalization processing, so that each physical parameter has the same scale in the training process; wherein is the mean value of the physical parameter is the mean value of the physical parameter is the standard deviation of the physical parameter is the standard deviation of the physical parameter is the normalized physical parameter.

3. The method of claim 1, wherein, In S3, the AMPD model comprises an input embedding layer, a feature fusion main body and a physical constraint and explicit residual module connected in sequence; The input embedding layer: inputting the spatial coordinates, the continuous region mask, the wall distance and the physical condition parameters, performing multi-scale feature extraction on the spatial coordinates through a dilated convolution module and mapping to a 32-dimensional feature space, generating spatial coordinate features, embedding the continuous region mask and the wall distance into 16-dimensional feature spaces through independent linear transformation respectively, and generating region mask features and wall distance features; the physical condition parameters are encoded into a physical condition vector by using a multilayer perceptron, and used for training the AMPD model by using a curriculum learning training strategy; The feature fusion main body: combining the spatial coordinate features, the region mask features and the wall distance features to form 64-dimensional fusion features, inputting the fusion features into a residual main network and an attention mechanism module in sequence, and obtaining flow field data, including a velocity field, a density field, a temperature field and a pressure field; The residual backbone network comprises six stacked residual neural blocks, each of which comprises a bilinear transformation layer, layer normalization, a Leaky ReLU activation function and a residual connection, and the attention mechanism module is used for dynamically adjusting the weight of features at different spatial positions, and the model's representation ability for a complex flow field is improved through deep feature fusion. The physical constraint and explicit residual module comprises a gradient feature extraction unit, a physical region identification unit, a residual calculation unit and a boundary condition forcing unit. The gradient feature extraction unit: input flow field data, and calculate the spatial gradient of each physical quantity, the gradient amplitude and the high-order derivative features of vorticity through an adaptive finite difference method. The physical region identification unit: dynamically identify key physical regions in the flow field according to the gradient features, and calculate a boundary layer mask and a vortex mask to provide regional prior information for subsequent physical constraints. The residual calculation unit: calculate physical residuals by using the basic control equations of fluid mechanics, wherein the physical residuals comprise a continuity equation residual, a momentum equation residual, an energy equation residual and an ideal gas state equation residual. The boundary condition forcing unit: receives the flow field data, the physical residuals, the continuous region mask, the boundary layer mask and the vortex mask, and inputs the spatial gradient of each physical quantity, the gradient amplitude and the high-order derivative features of vorticity calculated by the gradient feature extraction unit, dynamically calculates the gradient and divergence through an adaptive finite difference method, and enhances the physical residuals through an adaptive weight network, corrects the system deviation of density, velocity, temperature and pressure through residual calculation, and forces the execution of corresponding physical constraint conditions in different boundary regions, according to the different physical regions of the inside of the airfoil, the airfoil wall surface, the near-wall fluid, the external fluid, the inlet, the outlet and the free boundary corresponding to the continuous region mask value, respectively forces the execution of the corresponding constraints of the no-slip boundary condition, the specified inflow condition, the pressure or velocity gradient condition and the far-field boundary condition, adjusts the flow field data, and outputs the corrected flow field data.

4. The method of claim 3, wherein, In S4, the multi-objective residual minimization loss function of the physical constraint driven loss function system The expression is specifically: wherein, is a loss term covering continuity loss, momentum loss, energy loss, ideal gas state loss, inlet region loss, boundary condition constraint, and penalty loss, is a physical quantity gradient norm regularization term, is a parameter weight, is a physical quantity change rate constraint loss, is a change rate constraint weight parameter, is a gradient regularization weight parameter; loss of continuity The expression of the continuity loss is specifically: In the formula, An adaptive weighting function is calculated based on the key region weights of continuous region masking, near-wall distance, and physical shock wave. This represents the total number of sampling points within the fluid domain. For the first One sampling point, For the fluid computation domain, For the first The continuity equation residuals at each sampling point For the first Spatial coordinates of each sampling point; In response to the predicted velocity vector being , the density being , the temperature being , the pressure being P , and the spatial coordinates being , the expression for the mass conservation residual is wherein is the partial derivative symbol, x is x axis coordinates, y is y axis coordinates, u is x directional velocity component, v is y directional velocity component; Momentum losses include x Momentum losses in the axial direction And y Momentum losses in the axial direction ; In the formula, and is a standard deviation of the momentum residual, is x is a momentum residual in the axial direction, is y is a momentum residual in the axial direction; wherein is the Laplacian operator, and is the viscous stress divergence term; In the formula, η is the fluid dynamic viscosity; Energy loss The expression for the energy loss is given by: wherein is a normalization coefficient, is a numerical stability regularization term, is the energy conservation equation residual at the th sampling point; Energy conservation equation residual The expression for the energy conservation equation residual is given by: wherein Cp is the constant pressure specific heat capacity, k is the thermal conductivity, is the viscous dissipation term; wherein is the coefficient of volume viscosity; Ideal gas state loss The expression for the ideal gas state loss is given by: In the formula, is the ideal gas equation of state residual; In the formula, R Let be the ideal gas constant. For position x Density value at that location, It is the product of the ideal gas constant and temperature. For position x Predicted pressure value at the location; inlet zone losses The expression for the inlet zone losses is given by: In the formula, For standard inlet physical parameters, The number of entry nodes. This is the entrance boundary area. q Set of physical quantity variables A single physical quantity in it. For the first Predicted physical quantity values ​​at each entry node; The boundary condition restricts the continuous value weight generated by the tanh activation function to smoothly reflect the relative importance of each node in the abnormal physical region, the continuous value weight The expression is specifically: wherein is a shock region weight coefficient, is a vortex region weight coefficient, is a shear region weight coefficient, is a shock identification mask, is a vortex identification mask, is a shear layer identification mask. penalized loss The expression of the penalized loss is given by wherein is a safety threshold, is a penalty factor, is a ReLU activation function.

5. The method of claim 1, wherein, In S5, the method for training the AMPD model by using the curriculum learning training strategy is specifically as follows: First stage: the angle of attack range is limited within 20% of the total range, the upper limit of Reynolds number is 5 x 10 5 , the upper limit of Mach number is 0.3, the number of geometric samples is 15, and the number of physical conditions is 5; Second stage: the angle of attack range is extended to 40%, the upper limit of Reynolds number is 6x10 5 , the upper limit of Mach number is 0.5, the number of geometric samples is 18, and the number of physical conditions is 6; Stage 3: The angle of attack range is extended to 60%, the upper limit of Reynolds number is 7 x 10 5 , the upper limit of Mach number is 0.7, the number of geometric samples is 20, and the number of physical conditions is 7. Phase IV: The angle of attack range is extended to 80%, the upper limit of Reynolds number is 8 x 10 5 , the upper limit of Mach number is 0.85, the number of geometric samples is 23, and the number of physical conditions is 8. Fifth stage: angle of attack range reaches 100%, upper limit of Reynolds number 10 6 , upper limit of Mach number 0.95, number of geometric samples 25, number of physical conditions 10; In each stage, the loss threshold is set, and the initial value is 500, which decreases to 10 as the training period increases, and in each stage, when the validation loss is lower than the loss threshold and the training period is at least 20, the next stage is entered.

6. An aircraft flow field prediction system of multi-region physically driven neural network, applied to the aircraft flow field prediction method of multi-region physically driven neural network as claimed in any one of claims 1-5, characterized in that the system It comprises: a data preprocessing module for generating an airfoil profile, constructing a continuous region mask and implementing high-dimensional physical parameter sampling; a neural network training module comprising an AMPD model architecture, a physical constraint driven loss function system and a curriculum learning training strategy; a flow field prediction module comprising a trained AMPD model, which inputs an airfoil type, a Mach number, a Reynolds number, a temperature and an attack angle, and outputs complete flow field distribution data; a result visualization module providing a velocity vector diagram, a pressure contour map, a density contour map, a velocity amplitude map and a surface pressure coefficient curve display mode.

Images