A physical constraint neural network-based phase change flow full flow field calculation method

Abstract

The application discloses a phase change flow full flow field calculation method based on a physical constraint neural network. The application is based on flow field physical configuration measurement; experimental or simulated point data is collected and normalized to construct a data set; dimensionless phase change control equations and flow control equations are respectively added to a phase change prediction neural network loss function and a flow field prediction neural network loss function as physical constraints, the neural networks are respectively constrained, and the neural networks are trained through the data set; the space-time point coordinates of a to-be-measured point are input into the phase change prediction neural network, the output of the phase change prediction neural network is taken as the input of the flow field prediction neural network, and flow field information of the to-be-measured point is obtained; the application utilizes data of a few detection points and physical constraints to quickly calculate full flow field data, and provides a fast and accurate method for phase change flow full flow field calculation.

Description

Technical Field

[0001] This invention relates to the whole flow field of phase change flow, and more specifically to a method for calculating the whole flow field of phase change flow based on a physically constrained neural network. Background Technology

[0002] Under current technological levels, the main approaches to obtaining full-field information on phase change flows are direct experimental measurements or traditional numerical simulations. However, experimental measurements suffer from drawbacks such as high costs and difficulty in obtaining continuous data. Traditional numerical simulations, on the other hand, are computationally expensive and prone to convergence problems when solving complex flow fields with phase changes. Methods employing machine learning algorithms for flow field prediction require large amounts of sample data, and because they simply apply machine learning algorithms for model training and prediction without considering the underlying physical information, these methods generally perform poorly in flow field calculations, resulting in low computational accuracy. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention proposes a method for calculating the full flow field of phase change flow based on a physical constraint neural network. By utilizing data from a small number of monitoring points and physical constraints, the full flow field can be calculated quickly, providing a fast and accurate method for calculating the full flow field of phase change flow.

[0004] The present invention provides a method for calculating the full flow field of phase change flow based on a physically constrained neural network, comprising the following steps:

[0005] 1) Measure the boundary physical configuration of the entire phase change flow field to obtain geometric parameters, and then determine the computational domain of the entire phase change flow field based on the geometric parameters;

[0006] 2) In the phase change flow field, multiple monitoring points are uniformly set in both the phase change region and the non-phase change region. The number of monitoring points per unit volume in the phase change region should be multiples that in the non-phase change region. The spatiotemporal coordinates (x, y, z, t) of each monitoring point are obtained. The original physical field data of each monitoring point in the experiment or simulation are collected. The physical field data includes liquid volume fraction, velocity field, pressure field, density field and energy field. The original physical field data are normalized to obtain normalized data. All the normalized data constitute the dataset. The normalized liquid volume fraction is used as the first dataset, and the normalized velocity field, pressure field, density field and energy field are used as the second dataset.

[0007] 3) Dimensionless processing is performed on the flow control equations and phase change control equations respectively;

[0008] 4) Establish a phase transition prediction neural network. The dimensionless phase transition control equation is added as a physical constraint to the loss function of the phase transition prediction neural network. The phase transition prediction neural network is trained using the spatiotemporal coordinates (x, y, z, t) from step 2), the first dataset, and the loss function constraint of the phase transition prediction neural network. The input of the phase transition prediction neural network is the spatiotemporal coordinates (x, y, z, t), and the output is the liquid volume fraction.

[0009] 5) Establish a flow field prediction neural network. Add the dimensionless flow control equation as a physical constraint to the loss function of the flow field prediction neural network. Train the flow field prediction neural network using the liquid phase volume fraction output by the phase change prediction neural network, the second dataset obtained in step 2), and the loss function constraint of the flow field prediction neural network to obtain a trained flow field prediction neural network. The input of the flow field prediction neural network is the spatiotemporal coordinates (x, y, z, t) and the liquid phase volume fraction, and the output is the velocity field, pressure field, density field, and energy field.

[0010] 6) Input the spatiotemporal coordinates of the point to be measured in the entire flow field of phase change flow into the phase change prediction neural network. The phase change prediction neural network outputs the liquid volume fraction of the point to be measured. Input the spatiotemporal coordinates of the point to be measured and the liquid volume fraction of the point to be measured output by the phase change prediction neural network into the flow field prediction neural network. The flow field prediction neural network outputs the velocity field, pressure field, density field and energy field of the point to be measured, thereby obtaining the flow field information of the point to be measured.

[0011] 7) Repeat step 6) above to obtain the flow field information of all test points in the phase change flow field, thereby obtaining the phase change flow field.

[0012] In step 1), the geometric parameters include the inlet and outlet dimensions of the flow field, the wall variation curve, and the axis of symmetry.

[0013] In step 2), the number of monitoring points per unit volume in the phase change region shall not be less than one, and the number of monitoring points per unit volume in the phase change region shall be 2 to 3 times that in the non-phase change region.

[0014] To ensure that the dataset-driven loss term and the physical constraint loss term are weighted in the same dimension, normalization is performed on the first and second datasets respectively:

[0015]

[0016] Among them, Z n For normalized data, Z represents the original physical field data. max Z represents the maximum value of the original physical field data. min This represents the minimum value of the original physical field data.

[0017] In step 3), the flow control equations are constructed based on the Euler equations and then made dimensionless:

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025] Where: x, y, and z are the length, width, and height of the flow field, respectively; u is the x-axis velocity, v is the y-axis velocity, w is the z-axis velocity, t is time, p is pressure, ρ is density, E is energy, and L... * U is the characteristic length. * Let S' be the characteristic velocity, ρ* be the characteristic density, and S′ be the characteristic velocity. m For the dimensionless mass source term, S e ′ represents the dimensionless energy source term; x′, y′, z′, u′, v′, w′, t′, p′, ρ′ and E′ are the dimensionless quantities of flow field length, width, height, x-axis velocity, y-axis velocity, z-axis velocity, time, pressure, density and energy, respectively.

[0026] The phase transition control equations are constructed based on the Lee model and then made dimensionless. The dimensionless mass source term and energy source term are then substituted into the flow control equations:

[0027]

[0028] When p′≥p′ sat This refers to the evaporation process:

[0029]

[0030] S e ′=h′S′ m

[0031] When p′<p′ sat This is the condensation process:

[0032]

[0033] S e ′=-h′S′ m

[0034] Where: p is the pressure, p sat For saturation pressure, S′ m S is a dimensionless mass source term. e ' is a dimensionless energy source term, K is the evaporation-condensation coefficient, and α is the energy source term. v α is the gas phase volume fraction. l ρ is the liquid volume fraction. v ρ is the gas density. l Where p' is the liquid phase density and h is the latent heat of vaporization; p′, p′ sat p′ v p′ l h and h′ are dimensionless quantities of pressure, saturation pressure, gas density, liquid density, and latent heat of vaporization, respectively.

[0035] In step 4), the phase transition prediction neural network adopts a fully connected form, including an input layer, a hidden layer, and an output layer.

[0036] The loss function for the phase transition prediction neural network is:

[0037] Loss pc =loss d +loss f =MSE f +MSE d

[0038]

[0039]

[0040] Among them: Loss pc For phase transition prediction neural networks, the loss function is... d For data-driven loss terms, loss f For the physical constraint loss term, MSE f The mean square error (MSE) of the physical constraint equations. d For the mean squared error of the data-driven loss term, r i (x i ,y i ,t i ) represents the residual of the physical loss term at the i-th monitoring point, s i Let s be the liquid volume fraction at the i-th monitoring point. i (x i ,y i ,t i ) represents the model-predicted liquid volume fraction for the i-th monitoring point, where N represents the number of monitoring points, and x i ,y i ,z i ,ti This represents the two-dimensional coordinates and time of the i-th monitoring point.

[0041] In step 5), the flow field prediction neural network adopts a fully connected form, including an input layer, a hidden layer, and an output layer.

[0042] The loss function for the flow field prediction neural network is:

[0043] Loss = loss d +loss f =MSE f +MSE d

[0044]

[0045] MSE d =MSE ρ +MSE p +MSE u +MSE v +MSE w +MSE e

[0046]

[0047] Where: Loss is the loss function of the flow field prediction neural network, and MSE is... ρ MSE p MSE u MSE v MSE w and MSE e The mean square errors of the data loss terms for density, pressure, x-axis velocity, y-axis velocity, z-axis velocity, and energy, respectively; ρ i Let p be the density of the i-th monitoring point. i Let u be the pressure at the i-th monitoring point. i Let v be the x-axis velocity of the i-th monitoring point. i Let w be the y-axis velocity of the i-th monitoring point. i Let E be the z-axis velocity of the i-th monitoring point. i Let ρ be the energy at the i-th monitoring point; i (x i ,y i ,t i Let p be the model prediction density for the i-th monitoring point. i (x i ,y i ,t i ) represents the model-predicted pressure at the i-th monitoring point, u i (x i ,y i ,ti ) is the model prediction of the x-axis velocity for the i-th monitoring point, v i (x i ,y i ,t i ) is the model prediction of the y-axis velocity for the i-th monitoring point, w i (x i ,y i ,t i E is the model prediction of the z-axis velocity for the i-th monitoring point. i (x i ,y i ,t i ) is the model predicted energy for the i-th monitoring point.

[0048] Advantages of this invention:

[0049] This invention measures flow field based on physical configuration; it collects experimental or simulated point data and normalizes it to construct a dataset; it incorporates dimensionless phase change control equations and flow control equations as physical constraints into the loss functions of the phase change prediction neural network and the flow field prediction neural network, respectively, to constrain the neural networks, and trains the neural networks using the dataset; it inputs the spatiotemporal coordinates of the measured point into the phase change prediction neural network, and uses the output of the phase change prediction neural network as the input of the flow field prediction neural network to obtain the flow field information of the measured point; this invention utilizes data from a small number of detection points and physical constraints to quickly calculate the full flow field data, providing a fast and accurate method for calculating the full flow field of phase change flow. Attached Figure Description

[0050] Figure 1 This is a flowchart of the phase change flow full flow field calculation method based on physical constraint neural network of the present invention. Detailed Implementation

[0051] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0052] like Figure 1 As shown, the phase change flow full-field calculation method based on physical constraint neural network in this embodiment includes the following steps:

[0053] 1) Measure the boundary physical configuration of the entire phase change flow field to obtain geometric parameters, including the inlet and outlet dimensions, wall variation curves, and axis of symmetry; then determine the computational domain of the entire phase change flow field based on the geometric parameters.

[0054] 2) In the phase change flow field, multiple monitoring points are uniformly set in both the phase change region and the non-phase change region. The number of monitoring points per unit volume in the phase change region should not be less than one. The more monitoring points, the more accurate the results. The number of monitoring points in the phase change region should be three times that in the non-phase change region. Obtain the spatiotemporal coordinates (x, y, z, t) of each monitoring point. Simulate the original physical field data of each monitoring point. The physical field data includes liquid phase volume fraction, velocity field, pressure field, density field, and energy field. To ensure that the loss term driven by the dataset and the physical constraint loss term are weighted in the same dimension, the first and second datasets are normalized using the following formula:

[0055]

[0056] Where: Z n For normalized data, Z represents the original physical field data; Z max This represents the maximum value of the original physical field data;

[0057] Z min This represents the minimum value of the original physical field data;

[0058] All the normalized data constitute the dataset. The normalized liquid volume fraction is used as the first dataset, and the normalized velocity field, pressure field, density field, and energy field are used as the second dataset.

[0059] 3) For the flow control equations and phase change control equations, construct the flow control equations based on the Euler equations and perform dimensionless transformation:

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] Where: x, y, and z are the length, width, and height of the flow field, respectively; u is the x-axis velocity; v is the y-axis velocity; w is the z-axis velocity; t is time; p is pressure; ρ is density; E is energy; L * U is the characteristic length; * The characteristic velocity is ρ. * S′ represents the characteristic density; mFor the dimensionless mass source term, S e ′ represents the dimensionless energy source term; x′, y′, z′, u′, v′, w′, t′, p′, ρ′ and E′ are the dimensionless quantities of flow field length, width, height, x-axis velocity, y-axis velocity, z-axis velocity, time, pressure, density and energy, respectively.

[0068] The flow control equations are constructed based on the Lee model and then made dimensionless:

[0069]

[0070] Table 1. Phase Change Control Equations

[0071]

[0072] Where: p is the pressure, p sat For saturation pressure, S′ m S is a dimensionless mass source term. e ' is a dimensionless energy source term, K is the evaporation-condensation coefficient, and α is the energy source term. v α is the gas phase volume fraction. l ρ is the liquid volume fraction. v ρ is the gas density. l Where p' is the liquid phase density and h is the latent heat of vaporization; p′, p′ sat p′ v p l h′ and h′ are dimensionless quantities of pressure, saturation pressure, gas density, liquid density, and latent heat of vaporization, respectively.

[0073] 4) Establish a phase transition prediction neural network. The phase transition prediction neural network adopts a fully connected form. The size of the neural network is determined according to the size of the flow field and the phase transition region. It is generally five to seven layers, including one input layer, three to five hidden layers and one output layer. Each hidden layer contains more than 20 neurons. The input layer has a dimension of three and the output layer has a dimension of one.

[0074] The dimensionless phase transition governing equations are incorporated as physical constraints into the loss function of the phase transition prediction neural network. When the pressure is greater than the saturation pressure, the liquid begins to vaporize. The mass source term equals the product of the evaporation-condensation coefficient, the liquid volume fraction, the density, and the pressure difference coefficient, while the energy source term equals the product of the mass source term and the latent heat of vaporization. When the pressure is less than the saturation pressure, the gas begins to liquefy. The mass source term and the energy source term have opposite signs to those in the vaporization process, as shown in Table 1, p′≤p′. sat As shown in the row;

[0075] The loss function for the phase transition prediction neural network is:

[0076] Loss pc =loss d+loss f =MSE f +MSE d

[0077]

[0078]

[0079] Among them: Loss pc For phase transition prediction neural networks, the loss function is... d For data-driven loss terms, loss f For the physical constraint loss term, MSE f The mean square error (MSE) of the physical constraint equations. d For the mean squared error of the data-driven loss term, r i (x i ,y i ,t i ) represents the residual of the physical loss term at the i-th monitoring point, s i Let s be the liquid volume fraction at the i-th monitoring point. i (x i ,y i ,t i ) represents the model-predicted liquid volume fraction for the i-th monitoring point, where N represents the number of monitoring points, and x i ,y i ,z i ,t i This represents the two-dimensional coordinates and time of the i-th monitoring point;

[0080] Using the spatiotemporal coordinates (x, y, z, t) from step 2) and the first dataset, along with the loss function constraints of the phase transition prediction neural network, the phase transition prediction neural network is trained to obtain a trained phase transition prediction neural network; the input of the phase transition prediction neural network is the spatiotemporal coordinates (x, y, z, t), and the output is the liquid volume fraction.

[0081] 5) Establish a flow field prediction neural network. The flow field prediction neural network adopts a fully connected form. The size of the neural network is determined according to the size of the flow field and the phase transition region, generally seven to ten layers, including one input layer, five to eight hidden layers and one output layer. Each hidden layer contains more than 20 neurons. The input layer has a dimension of three and the output layer has a dimension of five.

[0082] The dimensionless flow control equations are incorporated as physical constraints into the loss function of the flow field prediction neural network.

[0083] The loss function for the field prediction neural network is:

[0084] Loss = loss d+loss f =MSE f +MSE d

[0085]

[0086] MSE d =MSE ρ +MSE p +MSE u +MSE v +MSE e

[0087]

[0088] Where: Loss is the loss function of the flow field prediction neural network, and MSE is... ρ MSE p MSE u MSE v MSE w and MSE e The mean square errors of the data loss terms for density, pressure, x-axis velocity, y-axis velocity, z-axis velocity, and energy, respectively; ρ i Let p be the density of the i-th monitoring point. i Let u be the pressure at the i-th monitoring point. i Let v be the x-axis velocity of the i-th monitoring point. i Let w be the y-axis velocity of the i-th monitoring point. i Let E be the z-axis velocity of the i-th monitoring point. i Let ρ be the energy at the i-th monitoring point; i (x i ,y i ,t i Let p be the model prediction density for the i-th monitoring point. i (x i ,y i ,t i ) represents the model-predicted pressure at the i-th monitoring point, u i (x i ,y i ,t i ) is the model prediction of the x-axis velocity for the i-th monitoring point, v i (x i ,y i ,t i ) is the model prediction of the y-axis velocity for the i-th monitoring point, w i (x i ,y i ,t i E is the model prediction of the z-axis velocity for the i-th monitoring point. i (xi ,y i ,t i ) is the model-predicted energy for the i-th monitoring point;

[0089] The flow field prediction neural network is trained using the liquid phase volume fraction output by the phase change prediction neural network, the second dataset obtained in step 2), and the loss function constraints of the flow field prediction neural network. The input of the flow field prediction neural network is the spatiotemporal coordinates (x, y, z, t) and the liquid phase volume fraction, and the output is the velocity field, pressure field, density field, and energy field.

[0090] 6) Input the spatiotemporal coordinates of the point to be measured in the entire flow field of phase change flow into the phase change prediction neural network. The phase change prediction neural network outputs the liquid volume fraction of the point to be measured. Input the spatiotemporal coordinates of the point to be measured and the liquid volume fraction of the point to be measured output by the phase change prediction neural network into the flow field prediction neural network. The flow field prediction neural network outputs the velocity field, pressure field, density field and energy field of the point to be measured, thereby obtaining the flow field information of the point to be measured.

[0091] 7) Repeat step 6) above to obtain the flow field information of all test points in the phase change flow field, thereby obtaining the phase change flow field.

[0092] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A method for calculating the entire flow field of phase change flow based on a physically constrained neural network, characterized in that, The method for calculating the entire flow field of phase change flow includes the following steps: 1) Measure the boundary physical configuration of the entire phase change flow field to obtain geometric parameters, and then determine the computational domain of the entire phase change flow field based on the geometric parameters; 2) In the phase change flow field, multiple monitoring points are uniformly set in both the phase change region and the non-phase change region. The number of monitoring points per unit volume in the phase change region should be multiple times that in the non-phase change region. The spatiotemporal coordinates (x, y, z, t) of each monitoring point are obtained. The original physical field data of each monitoring point in the experiment or simulation are collected. The physical field data includes liquid volume fraction, velocity field, pressure field, density field and energy field. The original physical field data are normalized to obtain normalized data. All the normalized data constitute the dataset. The normalized liquid volume fraction is used as the first dataset, and the normalized velocity field, pressure field, density field and energy field are used as the second dataset. 3) The flow control equations and phase change control equations are dimensionless. 4) Establish a phase transition prediction neural network. The dimensionless phase transition control equation is added as a physical constraint to the loss function of the phase transition prediction neural network. Using the spatiotemporal coordinates (x, y, z, t) from step 2) and the first dataset, as well as the loss function constraint of the phase transition prediction neural network, train the phase transition prediction neural network to obtain the trained phase transition prediction neural network. The input of the phase transition prediction neural network is the spatiotemporal coordinates (x, y, z, t), and the output is the liquid phase volume fraction. 5) Establish a flow field prediction neural network. Add the dimensionless flow control equation as a physical constraint to the loss function of the flow field prediction neural network. Train the flow field prediction neural network using the liquid phase volume fraction output by the phase change prediction neural network, the second dataset obtained in step 2), and the loss function constraint of the flow field prediction neural network to obtain a trained flow field prediction neural network. The input of the flow field prediction neural network is the spatiotemporal coordinates (x, y, z, t) and the liquid phase volume fraction, and the output is the velocity field, pressure field, density field, and energy field. 6) Input the spatiotemporal coordinates of the test point in the entire phase change flow field into the phase change prediction neural network, and the phase change prediction neural network outputs the liquid phase volume fraction of the test point; input the spatiotemporal coordinates of the test point and the liquid phase volume fraction of the test point output by the phase change prediction neural network into the flow field prediction neural network, and the flow field prediction neural network outputs the velocity field, pressure field, density field and energy field of the test point, thereby obtaining the flow field information of the test point; 7) Repeat step 6 above to obtain the flow field information of all test points in the phase change flow field, thereby obtaining the phase change flow field.

2. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 1), the geometric parameters include the inlet and outlet dimensions of the flow field, the wall variation curve, and the axis of symmetry.

3. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 2), the number of monitoring points per unit volume in the phase change region should be 2 to 3 times that in the non-phase change region.

4. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 2), the first and second datasets are normalized respectively: in: For normalized data, Z represents the original physical field data. This represents the maximum value of the original physical field data. This represents the minimum value of the original physical field data.

5. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 3), the flow control equations are constructed based on the Euler equations and then made dimensionless: Where: x, y, and z are the length, width, and height of the flow field, respectively; u is the x-axis velocity, v is the y-axis velocity, w is the z-axis velocity, t is time, p is pressure, ρ is density, E is energy, and L... * U is the characteristic length. * For characteristic velocity, ρ * For characteristic density, For the dimensionless mass source term, The energy source term after dimensionless designation; These are dimensionless quantities representing the flow field length, width, height, x-axis velocity, y-axis velocity, z-axis velocity, time, pressure, density, and energy, respectively.

6. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 3), the phase transition control equations are constructed based on the Lie model and made dimensionless. Then, the dimensionless mass source term and energy source term are substituted into the flow control equations: when This refers to the evaporation process: when This is the condensation process: Where: p is the pressure, p sat For saturation pressure, For dimensionless mass source terms, Here, K is the dimensionless energy source term, and K is the evaporation-condensation coefficient. This refers to the gas phase volume fraction. It is the liquid volume fraction. For gas density, Where is the liquid density, and h is the latent heat of vaporization; These are dimensionless quantities representing pressure, saturation pressure, gas density, liquid density, and latent heat of vaporization, respectively.

7. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 4), the phase transition prediction neural network adopts a fully connected form, including an input layer, a hidden layer, and an output layer.

8. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 4), the loss function of the phase transition prediction neural network is: Among them: Loss pc For phase transition prediction neural networks, the loss function is... d For data-driven loss terms, loss f For the physical constraint loss term, MSE f The mean square error (MSE) of the physical constraint equations. d The mean squared error of the data-driven loss term. The residual of the physical loss term at the i-th monitoring point, Let be the liquid volume fraction at the i-th monitoring point. The model predicts the liquid volume fraction for the i-th monitoring point, where N represents the number of monitoring points. This represents the three-dimensional coordinates and time of the i-th monitoring point.

9. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 5), the flow field prediction neural network adopts a fully connected form, including an input layer, a hidden layer, and an output layer.

10. The method for calculating the entire flow field of phase change flow as described in claim 1, characterized in that, In step 5), the loss function of the flow field prediction neural network is: Where: Loss is the loss function of the flow field prediction neural network. These are the mean square errors of the data loss terms for density, pressure, x-axis velocity, y-axis velocity, z-axis velocity, and energy, respectively. Let the density be the density of the i-th monitoring point. Let the pressure at the i-th monitoring point be... Let x be the x-axis velocity of the i-th monitoring point. Let be the y-axis velocity of the i-th monitoring point. Let be the z-axis velocity of the i-th monitoring point. Let i be the energy of the i-th monitoring point; The model predicts the density for the i-th monitoring point. For the model to predict the pressure at the i-th monitoring point, Predict the x-axis velocity for the model at the i-th monitoring point. Predict the y-axis velocity for the model at the i-th monitoring point. Predict the z-axis velocity for the model at the i-th monitoring point. Predict the energy for the i-th monitoring point using the model.

Images