CN112084727A - A Transition Prediction Method Based on Neural Network - Google Patents

Abstract

The invention discloses a transition prediction method based on a neural network, which comprises the following steps: step A, obtaining a plurality of known intermittent factors and local average characteristic quantity as a training set; step B, training a neural network and obtaining an intermittent factor mapping model by taking the local average characteristic quantity in the training set and the corresponding intermittent factor as an input value and an output value respectively; step C, a computational fluid mechanics solver carries out flow field iterative computation until the flow field computation result is iteratively converged, and a transition flow field prediction result is output; in each iteration step, the computational fluid dynamics solver provides the local flow field average for the pause factor mapping model, and the pause factor mapping model provides the pause factor for the computational fluid dynamics solver. According to the method, the number of partial differential equations needing to be calculated for transition prediction is reduced, the calculation time is greatly shortened, the contradiction that the precision and the calculation efficiency cannot coexist is solved, the calculation efficiency is high, and the calculation precision is high; the invention has good universality because of not depending on empirical formulas.

Description

A Transition Prediction Method Based on Neural Network

technical field

The invention belongs to the technical field of computational fluid dynamics, in particular to a transition prediction method based on a neural network.

Background technique

Transition refers to the transition process of fluid from stratified stable flow to chaotic turbulent flow. Boundary layer transitions generally exist inside or on the surface of fluid machinery, which is a challenging problem to be solved urgently in classical physics. During the operation of the aircraft, the fluid is easily transformed from stratified stable flow to chaotic turbulent flow due to the influence of many factors including incoming turbulence, surface roughness, Mach number and so on. This transition process is accompanied by abrupt changes in wall friction and heat transfer characteristics. Therefore, fast and accurate computational fluid dynamics (CFD) simulation of transitional flow is of great significance to improve the performance and design efficiency of aircraft.

Transition prediction can be done by using Direct Numerical Simulation (DNS) that solves full-scale turbulent pulsation or Large-Eddy Simulation (LES) method that only solves large-scale pulsation, but the calculation amount increases exponentially with the Reynolds number Re. The current development of computer technology is still difficult to meet the computing needs of its engineering applications. The Reynolds Averaged Navier-Stokes (RANS) method combined with the transition model still plays an important role in engineering practice due to its ease of use and high efficiency.

The correlation intermittent transition model can quantitatively describe the turbulence generation by introducing the "intermittent factor", which can predict the transition problem more accurately, and is the most popular method in the current engineering transition prediction. In transitional flow, the flow field will intermittently present laminar or turbulent flow at the same spatial position, which is called intermittent phenomenon. If a function is used to describe this phenomenon, and the function value is defined as 0 for laminar flow and 1 for turbulent flow, then the intermittent factor γ is the time average of the function.

The traditional approach to the model is to obtain the desired intermittent factor either by explicit equations or by solving additional transport equations. In the early days, Dhawan et al. used an empirical formula to describe the flow direction distribution of the intermittent factor, but the obtained intermittent factor only depends on the incoming flow and the surface conditions, and does not consider the flow field structure, and is only suitable for simple flows. Libby first introduced the transport equation with intermittent factor to calculate turbulent flow, and many subsequent researches focused on improving the prediction effect of transport equation on various flows such as free shear flow and boundary layer transition. However, these models often judge the onset of transition through non-local variables, that is, when calculating the intermittent factor of local grid points, the flow physical quantities in the upstream and downstream regions of the flow field need to be used. This kind of calculation method is difficult to achieve parallel calculation. The grid is also difficult to apply because the grid is not arranged in sequence. Menter et al. used the base strain rate Reynolds number instead of the momentum thickness Reynolds number Re _θ , and developed a four-equation SST-γ-Re _θ transition model based entirely on local variables. The model calculates the intermittent factor by introducing two additional partial differential equations with the intermittent factor γ and the momentum thickness Reynolds number Re _θ as variables respectively, and then couples the SST turbulence model to simulate the transition process. Many studies at home and abroad show that the model has more accurate prediction ability for transitional flow, but the additional differential equation introduced reduces the solution efficiency of this method. Bas et al. proposed an algebraic transition model (or called BC model) based on the local mean quantity, which is more computationally efficient than the former, but cannot accurately simulate the turbulence in the (0.5, 2) interval. In addition, traditional transition models are bound by empirical formulas and are sensitive to computing platforms. Model parameters need to be re-calibrated for different platforms, which often does not have good versatility.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a neural network-based transition prediction method for the shortcomings of low computational efficiency, low computational accuracy, dependence on empirical formulas, and poor versatility in the existing transition prediction methods, which greatly shorten the calculation time. Time, high calculation efficiency, high calculation accuracy, does not rely on empirical formulas, and has good versatility.

For solving the above-mentioned technical problems, the technical scheme adopted in the present invention is:

A neural network-based transition prediction method is characterized by including the following steps:

Step A, obtaining known multiple groups of intermittent factors and corresponding local average feature quantities as a training set;

Step B, using the local average feature quantity in the training set as the input value, and using the intermittent factor corresponding to the local average feature quantity as the input value in the training set as the output value, train the neural network and obtain the intermittent factor mapping model;

Step C, the computational fluid dynamics solver performs the iterative calculation of the flow field until the iteratively convergent flow field calculation result, and outputs the transition flow field prediction result; wherein, the iterative calculation process of the flow field includes:

Step C1, the computational fluid dynamics solver outputs the local average characteristic quantity;

Step C2, inputting the local average feature quantity obtained in Step C1 into the intermittent factor mapping model;

Step C3, the intermittent factor mapping model is solved to obtain the intermittent factor and input the intermittent factor into the computational fluid dynamics solver;

Step C4, based on the intermittent factor obtained in step C3, determine whether the flow field calculation result of the fluid mechanics solver converges, if so, terminate the iterative process; if not, then the computational fluid dynamics solver is based on the intermittent factor value obtained in step C3. Update the local average feature quantity and jump to step C1.

With the above method, the present invention reconstructs the pure data-driven "black box" mapping model between the local average feature quantity and the corresponding intermittent factor based on the known training set data and obtains the intermittent factor mapping model by using the neural network method. The intermittent factor is solved by the neural network (algebraic) method, which has higher computational efficiency than the traditional SST-γ-Re _θ model that introduces additional differential equations under the premise of ensuring accuracy. It truly realizes the unification of transition prediction accuracy and efficiency. Since the invention does not need to solve additional differential equations, the calculation time is greatly shortened, and the calculation efficiency is high; at the same time, because it does not rely on empirical formulas, the universality is good.

As a preferred way, in the step A, the intermittent factor is derived from known experimental or high-precision transition model calculation data; the local average characteristic quantity corresponding to the intermittent factor is derived from the SA combined with the existing intermittent factor and corrected by the data. Calculated data of the turbulence model; in the step C, the computational fluid dynamics solver is a computational fluid dynamics solver based on the SA turbulence model. The invention combines known data and uses the SA turbulence model to control the generation of turbulence at each spatial position of the flow field, and designs a data-driven algebraic SA-NN transition model.

As a preferred way, in the step A, the local average characteristic quantities in the local average characteristic quantities include flow field density, minimum wall distance, free turbulent flow degree, kinematic viscosity coefficient, Q criterion, normalized strain rate , one or more of SA model flow field variables, eddy Reynolds numbers, dimensionless terms like turbulent viscosity, and pressure gradients along streamlines. The invention utilizes a series of local average characteristic quantities closely related to the intermittent factor, and models the mapping relationship between the local average characteristic quantity and the intermittent factor by introducing the neural network technology, thereby providing a numerical simulation for the transition prediction problem. Efficient and accurate method.

As a preferred manner, the neural network includes six fully connected layers, two residual blocks containing eight hidden layers and eight fully connected layers arranged in sequence from shallow to deep, wherein each hidden layer has 24 Neuron nodes, and the activation function of each hidden layer is a linear rectification function (RELU function). After repeated experiments to balance the calculation cost and the loss function, this form of neural network has the best effect.

As a preferred manner, in the step C, the result of the flow field calculation is the flow field density, the flow field velocity or the flow field pressure.

As a preferred way, in the step C4, the method for judging whether the flow field calculation result is convergent is: if the difference between the current flow field calculation result and the last flow field calculation result is less than 10 ^-5 , then judge the convergence , otherwise the judgment does not converge.

As a preferred method, the process of obtaining the local average feature quantity corresponding to the intermittent factor includes: step A1, substituting the intermittent factor obtained from the existing data into the SA turbulence model and freezing; step A2, the computational fluid dynamics solver performs iterative calculation Until the flow field calculation result converges; in step A3, the computational fluid dynamics solver outputs the local average characteristic quantity.

Compared with the prior art, the present invention not only has the same or even higher prediction accuracy, but also greatly shortens the calculation time by reducing the number of partial differential equations that need to be calculated for the transition prediction, and solves the contradiction that the accuracy and the calculation efficiency cannot coexist. , the calculation efficiency is high, and the calculation precision is high; because the present invention does not rely on empirical formulas, the universality is good.

Description of drawings

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

Figure 2 is a schematic diagram of the neural network structure.

Figure 3 is a grid diagram of the T3A-transition plate calculation.

Figure 4 is a comparison diagram of the T3A-transition plate wall friction curve.

Figure 5 is a grid diagram of the S809 airfoil calculation.

Figure 6 is a comparison chart of the calculated and experimental values of the aerodynamic characteristics of the S809 airfoil. Among them, Figure 6(a) is a comparison diagram of the relationship between the lift coefficient and the angle of attack; Figure 6(b) is a comparison diagram of the relationship between the drag coefficient and the angle of attack.

Figure 7 is a comparison diagram of the friction coefficient distribution of the airfoil surface under different attack angles of the S809 airfoil. Among them, the angle of attack in Fig. 7(a) is 1°; the angle of attack in Fig. 7(b) is -5°; and the angle of attack in Fig. 7(c) is 9°.

Figure 8 is the residual convergence curve of T3A-transition plate and S809 airfoil at 0° angle of attack.

Detailed ways

Specifically, the present invention includes three stages: the first is the data preparation stage, which collects accurate calculation data of basic flow problems obtained by various experiments or model methods to form a training set, including but not limited to flat plate transition flow, flow around a large separation cylinder , flow around the square column, periodic mountain flow, back-step flow, square tube flow, airfoil flow and free shear flow; then the learning phase, using the training set to train the neural network model through supervised learning to form a flow The intermittent factor mapping model in which the average characteristic quantity of the field is the input and the intermittent factor is the output; the last is the coupling solution stage. In each iteration step of the flow field calculation, the intermittent factor is given by the intermittent factor mapping model, and no additional solution is required. The differential equation of , greatly improves the computational efficiency of the transition prediction problem.

The invention combines the neural network technology and the SA full turbulence model to design a data-driven SA-NN transition prediction method. The model building process can be divided into two parts: offline learning and coupled solution. The offline learning part mainly includes training data selection, neural network model framework and parameter optimization. In the coupling solution part, the trained intermittent factor mapping model will be embedded into the CFD (Computational Fluid Dynamics) solver, and the intermittent factor is calculated by the local flow field average characteristic quantity q(q ₁ ,q ₂ ,…,q _n ) calculation, and pass it to the CFD solver, the specific process is shown in Figure 1.

Specifically, the neural network-based transition prediction method of the present invention includes the following steps:

Step A, obtaining known multiple groups of intermittent factors and corresponding local average feature quantities as a training set;

Step B, using the local average feature quantity in the training set as the input value, and using the intermittent factor corresponding to the local average feature quantity as the input value in the training set as the output value, train the neural network and obtain the intermittent factor mapping model;

Step C, the computational fluid dynamics solver performs the iterative calculation of the flow field until the iteratively convergent flow field calculation result, and outputs the transition flow field prediction result; wherein, the iterative calculation process of the flow field includes:

Step C1, the computational fluid dynamics solver outputs the local average characteristic quantity;

Step C2, inputting the local average feature quantity obtained in Step C1 into the intermittent factor mapping model;

Step C3, the intermittent factor mapping model is solved to obtain the intermittent factor and input the intermittent factor into the computational fluid dynamics solver;

Step C4, based on the intermittent factor γ obtained in step C3, determine whether the flow field calculation result of the fluid mechanics solver converges, if so, terminate the iterative process; if not, then the computational fluid dynamics solver is based on the intermittent factor obtained in step C3. value to update the local average feature quantity and jump to step C1.

In the step C, the calculation result of the flow field is the flow field density, the flow field velocity or the flow field pressure. In the comparative test of the embodiment, the flow field density is selected to judge whether the flow field converges.

In the step C4, the method for judging whether the flow field calculation result is convergent is: if the difference (residual) between the current flow field calculation result and the last flow field calculation result is less than 10 ^-4 , then judge the convergence, otherwise Judgment does not converge.

In the step A, the intermittent factor is derived from known experimental or transition model calculation data; the local average characteristic quantity corresponding to the intermittent factor is derived from the calculation data of the SA turbulence model combined with the existing intermittent factor and corrected by the data; the In step C, the computational fluid dynamics solver is a computational fluid dynamics solver based on the SA turbulence model.

The governing equation of the one-equation SA model is as follows:

表示与湍流涡粘系数相关的修正涡粘系数，P_ν为生成项，D_ν为破坏项，C_b2和σ为常数。where ν is the molecular kinematic viscosity coefficient,

Represents the modified eddy viscosity related to the turbulent eddy viscosity, where P _ν is the generation term, D _ν is the destruction term, and C _b2 and σ are constants.

Referring to the idea of the correlation intermittent transition model, the present invention applies the intermittent factor to the turbulence model to suppress the generation of turbulence before and during the transition, so that the original full turbulence results are corrected to the transition flow field. Considering the objective definition of intermittent factor describing the intermittent state at a certain point in the flow field, it is reasonable to describe its potential function relationship through neural network and combine it with SA turbulence model. For the transport equation of the SA turbulence model, the intermittent factor is used to suppress the generation and destruction terms of the eddy viscosity:

Among them, P _ν and D _ν are the generation and destruction terms of the original SA model, respectively, γ _pre is the intermittent factor predicted by the neural network model, and β is the correction coefficient of the destruction term.

The acquisition process of the local average characteristic quantity corresponding to the intermittent factor includes: step A1, the intermittent factor obtained from the existing data is substituted into the SA turbulence model in the form of formula (2) and frozen, that is, the frozen intermittent factor is calculated in the subsequent calculation. Its value does not change; in step A2, the computational fluid dynamics solver performs iterative calculation until the flow field calculation result converges; in step A3, the computational fluid dynamics solver outputs the local average characteristic quantity.

In the step A, the local average characteristic quantities in the local average characteristic quantities include flow field density, minimum wall distance, free turbulent flow degree, kinematic viscosity coefficient, Q criterion, normalized strain rate, and SA model flow field variables. , eddy Reynolds number, dimensionless term similar to turbulent viscosity, pressure gradient along the streamline, and local average characteristic quantities are shown in Table 1.

Table 1. Local average feature quantities as neural network input

The invention utilizes a series of local average characteristic quantities closely related to the intermittent factor, and models the mapping relationship between the local average characteristic quantity and the intermittent factor by introducing the neural network technology, thereby providing a numerical simulation for the transition prediction problem. Efficient and accurate method.

The invention adopts the neural network to construct the intermittent factor mapping model.

可表示为：The general neural network structure can be composed of an input layer, several hidden layers, an output layer and a weighted connection, as shown in Figure 2. The network layers surrounded by cross-layer connections are called residual blocks. The network layers are fully connected. The input layer consists of a set of local mean quantities q=(q ₁ , q ₂ , . . . , q _n ) representing different properties of the flow field. These inputs go through a weighted connection to the next hidden layer, where the new input value is compared to a threshold at the node and then transformed through a nonlinear activation function. Each time the feature information passes through a hidden layer, it will be transformed into a new feature space until the output result. For a new input x _i , the output components of the first hidden layer

can be expressed as:

Wherein, the superscript (m) represents the mth hidden layer, φ is the nonlinear activation function, and the RELU function is used in this application; w _ij is the connection weight between the upper layer and this layer; c _j is the threshold at the neuron node; The subscript i represents the hidden layer of this layer, and the subscript j represents the hidden layer of the upper layer. Similarly, the output components of the second and third hidden layers can be written as:

where m ₁ and m ₂ represent the number of neuron nodes in the first and second hidden layers, respectively. It can be seen that the output of the first hidden layer is added to the unactivated output of the third hidden layer through the cross-layer connection, and the combined data flow is passed down through the activation function. This cross-layer connection is actually between layers. Adding an identity map, the final output component of the fourth hidden layer can be expressed as:

Thanks to the above-mentioned multi-level and nonlinear activation transformation, the network has the ability to describe the deep nonlinear mapping between input features and output quantities.

The loss function is an important indicator that directly reflects the accuracy of the model, and the loss function adopted in the present invention is set as follows:

表示训练集中对应的真实标签，γ_k(q_i)表示神经网络的输出结果，N为用于训练的数据点总数。in,

represents the corresponding true labels in the training set, γ _k (q _i ) represents the output of the neural network, and N is the total number of data points used for training.

Finally, the weight w _ij and the threshold c _j of the network nodes are optimized by the gradient descent method:

Among them, η is the optimization step size.

After repeated experiments to balance the calculation cost and the loss function, the neural network adopted in the present invention includes six fully connected layers set sequentially from shallow to deep, two residual blocks containing eight hidden layers, and eight fully connected layers. The connection layer, wherein each hidden layer has 24 neuron nodes, and the activation function of each hidden layer is a linear rectification function (RELU function).

The correlation intermittent transition model actually controls the generation of turbulence in the flow field through the intermittent factor, a parameter that varies with space, so as to achieve the purpose of simulating transition. For example, for a point in space, if the flow state here is laminar flow, the intermittent factor should be 0, then in the governing equation, the generation term of turbulence will be multiplied by 0, that is, no turbulence will be generated at this point. Then the problem is transformed into how to calculate the intermittent factor. The traditional method uses empirical formulas or introduces additional differential equations to calculate the intermittent factor, which may introduce calculation errors or reduce efficiency. The invention uses the neural network to build a model, and calculates the intermittent factor through the local average characteristic quantities of 10 flow fields, without using an empirical formula, reducing the uncertainty introduced by humans, so as to improve the prediction accuracy, and the model is an algebraic model, ensuring that Transition prediction efficiency.

Compared with the commonly used correlation intermittent transition model, the data-driven transition model proposed by the present invention has the same or even higher prediction accuracy, and greatly shortens the calculation time, does not rely on empirical formulas, and has stronger versatility.

The advantages of the method of the present invention are verified by calculation examples below.

1.T3A-Transfer Plate

The T3A-transition plate experiment will be used to test the predictive power of the above data-driven intermittent factor model for transition. The calculation grid used is shown in Figure 3, the grid is 324 × 108 (flow direction × normal direction), and 291 grid cells are distributed on the plate. The incoming Mach number is 0.0577, the Reynolds number is 1.4×10 ⁶ , and the turbulence degree is set to 0.843.

The turbulent boundary layer has greater friction, so the transition of the boundary layer can be judged by the friction curve. Figure 4 shows the comparison results of the wall friction. It can be found that the SST-γ-Re _θ model and the SA-NN model proposed by the present invention can predict the natural transition well, while the SA model fails to predict the transition, and The wall friction calculated for fully turbulent flow is very different from the experimental value. The BC model is able to predict the transition phenomenon, but prematurely predicts the transition location for the T3A-slab case. It can be seen from the wall friction curve of the T3A-transition plate (turbulent degree of 0.843) shown in Fig. 4 that the transition position simulated by the BC model has a large deviation from the experimental value, and the present invention conforms well. The reason is that the existing models largely rely on empirical formulas calibrated by humans, which increases the uncertainty of the model within a certain range and affects the applicability. The SA-NN model, which is also a single equation, avoids the above problems while ensuring the solution efficiency.

2. S809 airfoil

The S809 airfoil is a laminar airfoil with a thickness of 21%c, which is specially designed for transverse axis wind turbines and is a typical example to verify the transition model. The calculation grid is shown in Figure 5. The C-type topology is adopted, and a total of about 66,000 grid cells are divided. The grid distance of the first layer of the wall is 1×10 ^-6 c, the far-field boundary is 120c, and the inlet Mach number is 0.1. The Reynolds number was 2.0×10 ⁶ , and the airfoil leading edge turbulence was set to 0.2%.

Figure 6 compares the calculated and experimental results of the aerodynamic characteristics of the S809 airfoil. It can be seen from the figure that the lift-drag characteristics of the SA-NN model and the SST-γ-Re _θ model are more consistent with the experimental values. The lift coefficient predicted by the original SA model without considering the transition is too small, and the drag is always too much predicted. The SA-NN model largely corrects this situation.

Fig. 7 shows the friction coefficient distribution of the airfoil surface at different angles of attack. It can be seen from the figure that the SA model fails to predict the transition at the angles of attack of 1°, -5° and 9°. The SA-NN model is very close to the results of the SST-γ-Re _θ transition model. At 1° angle of attack, the upper and lower airfoil surfaces of the airfoil undergo separation transitions around 0.550 and 0.526, respectively. The transition position of the upper airfoil moves forward continuously with the increase of the angle of attack. When the angle of attack is 9°, the transition position reaches the vicinity of 0.01, the wall friction decreases continuously along the flow direction, and the flow becomes laminar again. The transition position of the lower airfoil continuously moves towards the trailing edge with the angle of attack.

The above results verify the transition prediction ability of the SA-NN model. On this basis, another advantage of the model is the solution efficiency. Compared with the four-equation SST-γ-Re _θ transition model, the SA-NN model gives the intermittent factor distribution through the algebraic "black box" model, without introducing additional differential equations, reducing the computational time. The efficiency improvement is more obvious in the case with large mesh amount. Figure 8 lists the residual convergence curves at 0° angle of attack for the T3A-transition plate and the S809 airfoil. When converging to the same accuracy, the computation time of the SA-NN model is reduced by more than 35% compared to the SST-γ-Re _theta transition model. The efficiency advantage brought by the algebraic transition model is even more significant when the flow is extended to three dimensions.

The embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than limiting. Under the inspiration of the present invention, without departing from the scope of protection of the spirit of the present invention and the claims, many forms can be made, which all fall within the protection scope of the present invention.

Claims (7)

1. A transition prediction framework based on a neural network is characterized by comprising the following steps:

step A, acquiring a plurality of known intermittent factors and corresponding local average characteristic quantities as a training set;

step B, taking the local average characteristic quantity in the training set as an input value, taking an intermittent factor corresponding to the local average characteristic quantity taken as the input value in the training set as an output value, training a neural network and obtaining an intermittent factor mapping model;

step C, a computational fluid mechanics solver carries out flow field iterative computation until the flow field computation result is iteratively converged, and a transition flow field prediction result is output; wherein, the flow field iterative computation process comprises:

step C1, calculating the local average characteristic quantity output by the fluid mechanics solver;

a step C2 of inputting the local average feature quantity obtained in the step C1 to the intermittent factor mapping model;

step C3, solving the intermittent factor mapping model to obtain intermittent factors, and inputting the intermittent factors into a computational fluid mechanics solver;

step C4, judging whether the flow field calculation result of the fluid mechanics solver converges or not based on the intermittent factor obtained in the step C3, and if so, terminating the iteration process; if not, the computational fluid dynamics solver updates the local average feature quantity according to the pause factor value obtained in step C3 and jumps to step C1.

2. The method for predicting transition based on neural network of claim 1, wherein in step a, the pause factor is derived from known experimental or high-precision transition model calculation data; the local average characteristic quantity corresponding to the intermittent factor is derived from the calculation data of the SA turbulence model which is combined with the existing intermittent factor and subjected to data correction; in the step C, the computational fluid mechanics solver is a computational fluid mechanics solver based on the SA turbulence model.

3. The neural network-based transition prediction method of claim 2, wherein in the step a, the local average feature quantities in the local average feature quantities include one or more of flow field density, minimum wall distance, free turbulence flow degree, kinematic viscosity coefficient, Q criterion, normalized strain rate, SA model flow field variable, vortex reynolds number, dimensionless terms like turbulence viscosity, and pressure gradient along a flow line.

4. The method according to any one of claims 1 to 3, wherein the neural network comprises six fully-connected layers, two residual blocks containing eight hidden layers, and eight fully-connected layers sequentially arranged from shallow to deep, wherein each hidden layer has 24 neuron nodes, and an activation function of each hidden layer is a linear rectification function.

5. The transition prediction method based on the neural network as claimed in any one of claims 1 to 3, wherein in the step C, the flow field calculation result is a flow field density, a flow field velocity or a flow field pressure.

6. The transition prediction method based on a neural network as claimed in any one of claims 1 to 3, wherein in the step C4, the method for determining whether the flow field calculation result converges includes: if the difference value between the current flow field calculation result and the last flow field calculation result is less than 10^-4If not, the convergence is judged.

7. The method for predicting transition based on a neural network of claim 2, wherein the obtaining of the local average feature quantity corresponding to the pause factor comprises: step A1, substituting the pause factor obtained from the existing data into an SA turbulence model and freezing; step A2, performing iterative computation by a computational fluid dynamics solver until a flow field computation result is converged; in step a3, the computational fluid dynamics solver outputs the local average feature quantity.

Images