Wall surface heat flow prediction method considering material ablation response under semi-supervised learning framework

Abstract

The invention discloses a wall surface heat flow prediction method considering material ablation response under a semi-supervised learning framework, and relates to the field of aerospace, and the method comprises the following steps: constructing a data set; building a neural network model, and introducing the ablation physical information loss item into a loss function of the neural network model; pre-training the neural network model; predicting the label-free data through a pre-trained neural network model, and generating a pseudo label; mixing the labeled data and the unlabeled data, and performing semi-supervised self-training on the pre-trained neural network model to obtain a final neural network model; and wall surface heat flow prediction of material ablation response is carried out through the final neural network model. According to the method, the requirement of the model for a labeled data set is reduced, the model generalization ability under the small sample size training condition is enhanced, the physical interpretability of the network model is improved, and the wall surface heat flow prediction precision is improved.

Description

Technical Field

[0001] This invention relates to the aerospace field, and more specifically to a method for predicting wall heat flux considering material ablation response within a semi-supervised learning framework. Background Technology

[0002] During flight, the heat flux on an aircraft surface increases rapidly, approximately to the cube of its flight speed, leading to increasingly severe aerodynamic heating environments. To protect aircraft from this intense aerodynamic heating, ablative thermal protection systems are often added to the aircraft surface. These systems absorb a significant amount of heat by ejecting ablative gases generated by the ablative material, thereby reducing the heat flux density on the aircraft wall and effectively preventing excessive heat flux that could cause the aircraft to burn up. Different ablative protection materials and structural forms directly affect the heat flux on the aircraft wall; therefore, accurate prediction of the wall heat flux is crucial in designing ablative thermal protection systems during the aircraft design process.

[0003] Currently, commonly used methods for predicting wall heat flow include: ablation wind tunnel tests, empirical formula prediction, CFD-ablation coupled calculations, and fully supervised neural network prediction. Among these methods, ablation wind tunnel testing is too costly and difficult to obtain large amounts of experimental data; empirical formulas predict results with large discrepancies from actual results, failing to accurately predict the heat flow data of the aircraft wall; CFD-ablation coupled calculation can achieve refined prediction of ablation heat flow of the aircraft wall, but this method is costly and time-consuming, often requiring 1-2 months or even longer to calculate along the flight trajectory (which consists of hundreds or thousands of points, each with different inflow conditions, making it difficult to obtain results through coupled calculation); therefore, using a fully supervised neural network for rapid prediction has become a common method for predicting wall ablation heat flow. This method uses wall heat flow without considering ablation, boundary layer outer edge parameters, inflow parameters, and ablation mass flow rate as input features of the fully supervised neural network, and wall heat flow considering ablation as the output feature, constructing a fully connected neural network model. After training on a large amount of data, it can obtain wall heat flow data under ablation conditions by inputting only data without ablation and mass flow rate data. However, long-term research has revealed several drawbacks to this method. First, it suffers from significant data dependency. Model training requires massive training datasets with strict correspondences (no-ablation heat flux, ablation mass flow rate, and target ablation heat flux must correspond one-to-one, i.e., labeled datasets). Each data sample must be obtained through computational fluid dynamics (CFD) coupled with the ablation model simulation, resulting in a single set of data taking weeks to generate, and the cost of building large-scale datasets increases exponentially. Second, it exhibits weak generalization ability with small samples. When the amount of training data is too small, the model prediction error generally exceeds the 15% threshold allowed in engineering, easily leading to distortion in localized heat flux predictions, making it difficult to meet the requirements of high-precision aerodynamic thermal protection design. Third, it lacks physical interpretability. Although the existing network architecture has strong nonlinear fitting capabilities, its "black box" nature makes it impossible to analyze the physical relationship between mass flow rate and heat flux, severely restricting the reliability verification of the model in the design of new aircraft. Therefore, it is urgent to build a prediction model with the following characteristics: (1) reduce the model’s need for labeled datasets; (2) reduce training data while ensuring prediction accuracy, i.e. enhance the model’s generalization ability under small sample training conditions; and (3) increase the physical interpretability of the network model. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, the wall heat flow prediction method considering material ablation response within a semi-supervised learning framework provided by this invention solves the problems of existing fully supervised network models having a large demand for labeled data, weak generalization ability with small sample sizes, and poor physical interpretability.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for predicting wall heat flux considering material ablation response under a semi-supervised learning framework is provided, which includes the following steps: The heat flux distribution data and ablation mass flow rate of the smooth wall were obtained through CFD and ablation calculations. Partial heat flux distribution data of the ablation wall were obtained through CFD-ablation coupled calculations and used as the true labels. When there is a corresponding ablation heat flux distribution data for the smooth wall and the ablation mass flow rate, the corresponding smooth wall heat flux distribution data and ablation mass flow rate are recorded as labeled data; otherwise, the corresponding smooth wall heat flux distribution data and ablation mass flow rate are recorded as unlabeled data. A neural network model is constructed, the loss term of ablation physical information is calculated, and the sum of this term and the mean square error loss function is used as the loss function of the neural network model. Based on the loss function of the neural network model, labeled data is used to pre-train the neural network model to obtain a pre-trained neural network model. Predictions are made on unlabeled data using a pre-trained neural network model to generate pseudo-labels; Based on the loss function of the neural network model, the pre-trained neural network model is semi-supervised and self-trained using a dataset consisting of labeled data, ablation wall heat flow distribution data, unlabeled data, and pseudo-labels to obtain the final neural network model. The wall heat flow of the material ablation response is predicted using the final neural network model.

[0006] Furthermore, the heat flux distribution data of the smooth wall includes the boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, incoming total enthalpy, wall temperature, wall pressure, and smooth wall heat flux; the ablation mass flow rate is obtained through ablation calculation.

[0007] Furthermore, the neural network model is a deep neural network model.

[0008] Furthermore, the expression for the loss function of the neural network model is:

[0009]

[0010] in This is the loss function for the neural network model; Let the mean squared error loss function be used. This refers to the loss of physical information due to ablation. n This represents the total number of labeled data. The label prediction results of the neural network model on labeled data; This is a real label.

[0011] Furthermore, the physical information loss due to ablation The calculation expression is:

[0012] in No heat flow is generated on the ablated wall surface; The mass flow rate of the gas ejector; The static enthalpy at the outer edge of the boundary layer; The coefficient of recovery; The velocity is the outer edge velocity of the boundary layer.

[0013] Furthermore, the specific method for pre-training the neural network model using labeled data is as follows: Labeled data is used as input to a deep neural network model. The deep neural network model predicts the label of the input labeled data. The predicted label and the corresponding real label are used as the basis for calculating the mean squared error loss function. The ablation physical information loss term is calculated, and then the loss function value of the neural network model is obtained. The deep neural network model is optimized using the loss function value of the neural network model to complete the pre-training.

[0014] Furthermore, the specific method for semi-supervised self-training of a pre-trained neural network model using a dataset composed of labeled data, ablation wall heat flux distribution data, unlabeled data, and pseudo-labels includes the following steps: Labeled and unlabeled data are mixed to obtain mixed data; The current neural network model is trained using mixed data, and the normalized loss for unlabeled data is calculated; the current neural network model before training with mixed data is a pre-trained neural network model. Unlabeled data with a normalization loss of less than 0.5 are classified as labeled data and the mixed data is updated; the current neural network model is updated using the loss function value of the neural network model. The semi-supervised self-training of the current neural network model is repeated with mixed data until the amount of remaining unlabeled data is less than one-fifth of the initial total amount of unlabeled data or the number of training rounds is reached, thus obtaining the final neural network model.

[0015] Furthermore, the specific method for predicting the wall heat flux of the material ablation response using the final neural network model is as follows: The heat flux distribution data and ablation mass flow rate of the smooth wall surface of the object to be predicted are obtained by CFD calculation and ablation calculation methods and used as input to the final neural network model. The heat flux distribution data of the ablated wall surface output by the final neural network model is used as the wall heat flux prediction result of the material ablation response.

[0016] A computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs a method for predicting wall heat flux considering material ablation response under a semi-supervised learning framework.

[0017] A computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, causes the processor to perform a wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0018] The beneficial effects of this invention are as follows: 1. Reduced reliance on labeled datasets. This neural network model makes full use of unlabeled data, eliminating the need for CFD-ablation coupling calculations for each flight path, thus significantly reducing the cost of constructing datasets.

[0019] 2. Enhanced model generalization ability under small sample training conditions. Compared with traditional fully supervised neural network models, the wall heat flow prediction accuracy of this neural network model trained with a small sample size is improved by 30%.

[0020] 3. Enhance the physical interpretability of the network model. By incorporating the ablation physical information loss term into the loss function of the neural network model, the relationship between the heat flux of the non-ablated wall, the ablation mass flow rate, and the heat flux of the ablated wall is accurately described, thus increasing the physical meaning of the network model. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a schematic diagram of the Stardust calculation model; Figure 3 This is a comparison chart of the calculation results of T66 and the prediction results of machine learning; Figure 4 This is a comparison chart of the T76 calculation results and the machine learning prediction results. Detailed Implementation

[0022] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0023] like Figure 1 As shown, the wall heat flux prediction method considering material ablation response under this semi-supervised learning framework includes the following steps: S1. Obtain heat flux distribution data and ablation mass flow rate of smooth wall surface through CFD calculation and ablation calculation methods; obtain partial heat flux distribution data of ablation wall surface through CFD-ablation coupled calculation and use it as the true label; when there is a corresponding ablation wall surface heat flux distribution data for smooth wall surface heat flux distribution data and ablation mass flow rate, the corresponding smooth wall surface heat flux distribution data and ablation mass flow rate are recorded as labeled data, otherwise the corresponding smooth wall surface heat flux distribution data and ablation mass flow rate are recorded as unlabeled data; S2. Build a neural network model, calculate the ablation physical information loss term and use the sum of it and the mean square error loss function as the loss function of the neural network model. S3. Based on the loss function of the neural network model, the neural network model is pre-trained using labeled data to obtain the pre-trained neural network model. S4. Predict unlabeled data using a pre-trained neural network model to generate pseudo-labels; S5. Based on the loss function of the neural network model, the pre-trained neural network model is semi-supervised and self-trained using a dataset consisting of labeled data, ablation wall heat flow distribution data, unlabeled data, and pseudo-labels to obtain the final neural network model. S6. Predict the wall heat flow of the material ablation response using the final neural network model.

[0024] In the specific implementation process, the heat flux distribution data of the smooth wall includes the boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, incoming total enthalpy, wall temperature, wall pressure, and smooth wall heat flux; the ablation mass flow rate is obtained through ablation calculations. The neural network model is a deep neural network (DNN) model. The parameters input to the neural network model are used to participate in the neural network model's prediction of labels and affect the calculation of the mean squared error loss function, thereby improving the accuracy of the neural network model's label prediction.

[0025] The expression for the loss function of the neural network model in step S2 is:

[0026]

[0027] in This is the loss function for the neural network model; Let the mean squared error loss function be used. This refers to the loss of physical information due to ablation. n This represents the total number of labeled data. The label prediction results of the neural network model on labeled data; This is a real label.

[0028] Ablation physical information loss item The calculation expression is:

[0029] in No heat flow is generated on the ablated wall surface; The mass flow rate of the gas ejector; The static enthalpy at the outer edge of the boundary layer; The coefficient of recovery; The velocity is the outer edge velocity of the boundary layer.

[0030] The specific method for pre-training the neural network model using labeled data in step S3 is as follows: Labeled data is used as input to a deep neural network model. The deep neural network model predicts the label of the input labeled data. The predicted label and the corresponding real label are used as the basis for calculating the mean squared error loss function. The ablation physical information loss term is calculated, and then the loss function value of the neural network model is obtained. The deep neural network model is optimized using the loss function value of the neural network model to complete the pre-training.

[0031] In this embodiment, the specific method for semi-supervised self-training of a pre-trained neural network model using a dataset composed of labeled data, ablation wall heat flux distribution data, unlabeled data, and pseudo-labels includes the following steps: S5-1. Mix labeled and unlabeled data to obtain mixed data; S5-2. Train the current neural network model using mixed data and calculate the normalized loss for unlabeled data; where the current neural network model before training with mixed data is a pre-trained neural network model. S5-3. Divide unlabeled data with normalization loss less than 0.5 into labeled data and update the mixed data; update the current neural network model using the loss function value of the neural network model; S5-4. Repeatedly mix the data to semi-supervisedly train the current neural network model until the remaining amount of unlabeled data is less than one-fifth of the initial total amount of unlabeled data or reaches the preset number of training rounds, to obtain the final neural network model.

[0032] The specific method for predicting the wall heat flow of the material ablation response using the final neural network model in step S6 is as follows: The heat flux distribution data and ablation mass flow rate of the smooth wall surface of the object to be predicted are obtained by CFD calculation and ablation calculation methods and used as input to the final neural network model. The heat flux distribution data of the ablated wall surface output by the final neural network model is used as the wall heat flux prediction result of the material ablation response.

[0033] This embodiment also provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes a wall heat flow prediction method that considers the material ablation response under a semi-supervised learning framework.

[0034] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform a wall heat flux prediction method considering material ablation response under a semi-supervised learning framework.

[0035] In one embodiment of the present invention, the accuracy of the method is verified using the Stardust model as an example. A schematic diagram of the Stardust model is shown below. Figure 2 As shown in Table 1, the incoming flow conditions and the calculation methods are as follows.

[0036] Table 1: Incoming Flow Conditions and Calculation Methods

[0037] Among them, the heat flux of the ablation wall was obtained by CFD-ablation coupling calculation for operating conditions T34 and T60, that is, there is ablation heat flux data that can be used as the output data of the neural network. Therefore, operating conditions T34 and T60 are called labeled data and are divided into training set; T42, T48 and T54 are divided into unlabeled datasets; T66 and T76 are divided into test datasets.

[0038] The results of CFD-ablation coupling calculations, traditional fully supervised neural network predictions, and the ablation heat flux of this method are compared as follows: Figure 3 and Figure 4 As shown in the figure, the ablation heat flux results obtained by this method are in high agreement with the CFD-ablation coupling calculation results. Under the same training data (a small amount of training data), the prediction error is within 5%, and the prediction accuracy is significantly better than the traditional fully supervised neural network model.

[0039] In summary, this invention proposes a wall heat flow prediction method that considers material ablation response within a semi-supervised learning framework. This method reduces the model's requirement for labeled datasets, enhances the model's generalization ability under small sample training conditions, increases the physical interpretability of the network model, and improves the accuracy of wall heat flow prediction.

Claims

1. A method for predicting wall heat flux considering material ablation response within a semi-supervised learning framework, characterized in that, Includes the following steps: The heat flux distribution data and ablation mass flow rate of the smooth wall were obtained through CFD and ablation calculations. Partial heat flux distribution data of the ablation wall were obtained through CFD-ablation coupled calculations and used as the true labels. When there is a corresponding ablation heat flux distribution data for the smooth wall and the ablation mass flow rate, the corresponding smooth wall heat flux distribution data and ablation mass flow rate are recorded as labeled data; otherwise, the corresponding smooth wall heat flux distribution data and ablation mass flow rate are recorded as unlabeled data. A neural network model is constructed, the loss term of ablation physical information is calculated, and the sum of this term and the mean square error loss function is used as the loss function of the neural network model. Based on the loss function of the neural network model, labeled data is used to pre-train the neural network model to obtain a pre-trained neural network model. Predictions are made on unlabeled data using a pre-trained neural network model to generate pseudo-labels; Based on the loss function of the neural network model, the pre-trained neural network model is semi-supervised and self-trained using a dataset consisting of labeled data, ablation wall heat flow distribution data, unlabeled data, and pseudo-labels to obtain the final neural network model. The wall heat flow of the material ablation response is predicted using the final neural network model.

2. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 1, characterized in that, The heat flux distribution data for the smooth wall includes the boundary layer outer edge Mach number, boundary layer outer edge density, boundary layer outer edge velocity, boundary layer outer edge temperature, total enthalpy of incoming flow, wall temperature, wall pressure, and heat flux over the smooth wall; the ablation mass flow rate is obtained through ablation calculations.

3. The wall heat flux prediction method considering material ablation response within a semi-supervised learning framework according to claim 1, characterized in that, The neural network model is a deep neural network model.

4. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 3, characterized in that, The expression for the loss function of a neural network model is: in This is the loss function for the neural network model; Let the mean squared error loss function be used. This refers to the loss of physical information due to ablation. n This represents the total number of labeled data. The label prediction results of the neural network model on labeled data; This is a real label.

5. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 4, characterized in that, Ablation physical information loss item The calculation expression is: in No heat flow is generated on the ablated wall surface; The mass flow rate of the gas ejector; The static enthalpy at the outer edge of the boundary layer; The coefficient of recovery; The velocity is the outer edge velocity of the boundary layer.

6. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 5, characterized in that, The specific method for pre-training a neural network model using labeled data is as follows: Labeled data is used as input to a deep neural network model. The deep neural network model predicts the label of the input labeled data. The predicted label and the corresponding real label are used as the basis for calculating the mean squared error loss function. The ablation physical information loss term is calculated, and then the loss function value of the neural network model is obtained. The deep neural network model is optimized using the loss function value of the neural network model to complete the pre-training.

7. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 1, characterized in that, The specific method for semi-supervised self-training of a pre-trained neural network model using a dataset consisting of labeled data, ablation wall heat flux distribution data, unlabeled data, and pseudo-labeled data includes the following steps: Labeled and unlabeled data are mixed to obtain mixed data; The current neural network model is trained using mixed data, and the normalized loss for unlabeled data is calculated; the current neural network model before training with mixed data is a pre-trained neural network model. Unlabeled data with a normalization loss of less than 0.5 are classified as labeled data and the mixed data is updated; the current neural network model is updated using the loss function value of the neural network model. The semi-supervised self-training of the current neural network model is repeated with mixed data until the amount of remaining unlabeled data is less than one-fifth of the initial total amount of unlabeled data or the number of training rounds is reached, thus obtaining the final neural network model.

8. The wall heat flux prediction method considering material ablation response under the semi-supervised learning framework according to claim 1, characterized in that, The specific method for predicting wall heat flow in material ablation response using the final neural network model is as follows: The heat flux distribution data and ablation mass flow rate of the smooth wall surface of the object to be predicted are obtained by CFD calculation and ablation calculation methods and used as input to the final neural network model. The heat flux distribution data of the ablated wall surface output by the final neural network model is used as the wall heat flux prediction result of the material ablation response.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform a wall heat flow prediction method considering material ablation response under the semi-supervised learning framework described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer program is stored therein, and when the computer program is executed by a processor, the processor performs a method for predicting wall heat flow considering material ablation response under the semi-supervised learning framework described in any one of claims 1 to 8.

Images