Two-phase flow prediction method based on branch neural network

Abstract

The invention discloses a two-phase flow prediction method based on a branch neural network, and the method comprises the steps: respectively learning the speed, density, pressure or other important physical attributes of two phases and a phase boundary change rule through different branches, and sharing features and transmitting features through feature coupling. According to the method, the output of each branch can be integrated by using a proper fusion strategy according to the correlation between different branches and mutually complementary feature information, so that a final two-phase flow prediction result is obtained, and the defects that a traditional computational fluid mechanics method is high in iteration frequency in the calculation process, and the calculation efficiency is low are overcome. And the problems of poor two-phase flow prediction precision, boundary divergence and the like of a common neural network are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of two-phase flow prediction, and particularly to a two-phase flow prediction method based on a branched neural network. Background Art

[0002] Computational Fluid Dynamics (CFD) plays an important role in scientific research and industrial applications. It uses a computer as a tool and applies various discretized mathematical methods to conduct numerical experiments, computational simulations, and analytical studies on various problems of fluid mechanics. Among them, the Lattice Boltzmann method (LBM), as a new computational method, has achieved great development in the past thirty years. This method is a mesoscopic-level numerical simulation method between microscopic molecular dynamics simulation methods and macroscopic traditional numerical simulation methods. The Lattice Boltzmann method focuses on the distribution function of fluid particles and describes the spatio-temporal evolution of the distribution function according to the corresponding control equations. The macroscopic motion information of the fluid is obtained through the relationship between macroscopic physical quantities and the distribution function. Compared with traditional numerical simulation methods, the mesoscopic physical background and clear particle motion image of the Lattice Boltzmann method enable it to easily describe the interactions inside the fluid and between the fluid and the solid from the bottom layer, thus having significant advantages in the simulation of complex flows such as multiphase multicomponent flows and porous media flows.

[0003] Currently, a variety of LBM multiphase flow models have been established and used for the simulation of multiphase flow phenomena. Most of these models can be divided into four categories: color models, pseudo-potential models, free energy models, and phase field models. Due to the simplicity and applicability of the pseudo-potential model, it is widely used in multiphase flow simulation. In this type of model, non-local interaction forces are introduced between particles at each adjacent lattice point. The model can simulate multiphase multicomponent flows, and each component has an interaction potential function to control the state equation of the fluid of that component. The occurrence of phase separation depends on the value of the interaction potential. Compared with traditional methods, the pseudo-potential model can automatically achieve the process of phase separation without an additional interface capture function, and the pseudo-potential model describes the interaction between phases in the form of intuitive inter-particle forces, which can better reflect the physical essence of phase separation.

[0004] Due to the complexity and diversity of multiphase flow behavior, a single neural network may not be able to generalize well under different multiphase flow scenarios and conditions, resulting in a decline in prediction performance. Therefore, a branched neural network is introduced to accelerate the calculation speed and reduce the storage requirements. Moreover, the branched neural network has high flexibility and scalability, and can flexibly design the network structure according to the task requirements and data characteristics. A branched network is a neural network structure designed to process multiple tasks or learn multiple feature representations simultaneously by introducing multiple parallel branches into the network. Its design principle is that the branched neural network creates additional branches in the middle or at the end of the network to process different inputs or tasks. These branches can have different structures, such as having different numbers of layers, parameters, or activation functions. Each branch generates its own output, and then these outputs can be combined through weighted summation, concatenation, or other methods to generate the final output result. To reduce the number of parameters and computational costs, the branched neural network usually shares some parameters, especially when there are shared features between the branches, thereby improving the network's expressive power, generalization ability, and efficiency.

[0005] The development of this structure benefits from the requirements of multi-task learning and domain-specific tasks. Multi-task learning aims to improve the training efficiency and the model's generalization ability by learning multiple related tasks simultaneously through one model. The branched neural network realizes multi-task learning by creating parallel task-specific branches. In domain-specific tasks, it is usually necessary to process different types of inputs. For example, in computer vision, one branch can be responsible for object detection, and another branch can be responsible for image segmentation. The branched neural network provides an effective method to handle these complex tasks and achieve end-to-end learning. Summary of the Invention

[0006] Aiming at the problems of low computational efficiency in multiphase flow numerical simulation and low prediction accuracy of a single neural network, the present invention provides a two-phase flow prediction method based on a branched neural network. This method separately learns the velocities, densities, pressures, or other important physical properties of the two phases and the laws of their phase boundary changes through different branches. By feature coupling, it can share and transfer features, utilize the correlation and complementary feature information between different branches, and use an appropriate fusion strategy to integrate the outputs of each branch to obtain the final two-phase flow prediction result, overcoming the disadvantages of traditional computational fluid dynamics methods, such as a large number of iterations in the calculation process, consuming a large amount of computational resources, and low computational efficiency, as well as solving the problems of poor prediction accuracy and boundary divergence of ordinary neural networks for two-phase flow.

[0007] To achieve the above object, the technical solution provided by the present invention is as follows:

[0008] A two-phase flow prediction method based on a branched neural network, comprising:

[0009] Construct a two-phase flow data acquisition model, and acquire two-phase flow data through this two-phase flow data acquisition model;

[0010] Construct a branch network model, which includes three neural network branches, respectively used for predicting the two-phase flow distribution and physical regulation based on the energy conservation formula; there is an information flow path between the three neural network branches for feature coupling, so that features belonging to different neural network branches can be cross-learned and shared;

[0011] Train the branch network model with the acquired two-phase flow data;

[0012] Perform two-phase flow prediction through the trained branch network model.

[0013] Furthermore, the two-phase flow data acquisition model is a pseudo-potential lattice Boltzmann model. In the standard lattice Boltzmann equation of this pseudo-potential lattice Boltzmann model, the discretization of time and space is related through the discrete velocity of particles, that is, δ x = e αx δ t δ y = e αy δ t δ z = e αz δ t δ x δ y δ z are all corresponding space step lengths, δ t is the discrete unit time step length, e αx e αy e αz are the corresponding discrete velocities. According to this characteristic, the particle motion in physical space is divided into two processes: collision and migration, that is:

[0014] Collision:

[0015] Migration:

[0016] Among them, f α (r, t) is the distribution function, representing the state at position r and time t; represents the distribution function after collision. The collision process will adjust the particle distribution, but the position r and time t remain unchanged; is the equilibrium state distribution function, F α (r, t) is the force term in the discrete velocity space, and τ is the dimensionless relaxation time;

[0017] That is, after a particle collides with other particles at a certain node and passes through a time step, it just moves from this node to its adjacent node; the collision process is completely local, while the migration process is linear;

[0018] F α (r, t) is obtained from the pseudo - force model lattice Boltzmann - BGK equation:

[0019]

[0020] where ω α is the weight coefficient; τ is the relaxation factor; e α is the discrete velocity; u is the velocity; c s is the lattice sound speed, with a value of F is the force exerted on the system, including the gravitational force F b and the pseudo - potential interaction force F m ;

[0021] The calculation formula for the pseudo - potential interaction force F m is as follows:

[0022]

[0023] where G is the interaction strength; ψ(r) is the pseudo - potential at the spatial position r; P EOS is the equation of state, which is divided into the vander Waals equation of state and the Carnahan - Starling equation; ρ is the density, and ρ0 is a constant.

[0024] Furthermore, the neural network branches for two - phase flow distribution prediction respectively adopt a neural network collision model. This neural network collision model adopts a variant of the fully - connected feed - forward neural network, consisting of two hidden layers, with 50 neurons in each layer; the ReLU rectified linear unit is used as the activation function, and there is no bias in the linear layer.

[0025] Furthermore, the neural network mapping Ω N of the neural network collision model satisfies the following conditions:

[0026]

[0027] where are the particle distributions before and after the collision respectively;

[0028] The particle distribution on all grids satisfies:

[0029]

[0030] e - , e + are the lattice velocities before and after the collision respectively.

[0031] Furthermore, the training and learning process of the neural network collision model includes:

[0032] Symmetry input expansion: First, perform all eight symmetry operations of the dihedral group on the input state to generate eight geometrically equivalent input copies; the eight symmetry operations include four rotations and four mirror reflections;

[0033] Network learning: Use the neural network collision model for learning;

[0034] Inverse transformation average output: Reverse the eight outputs to the original direction and then take the average.

[0035] Furthermore, the physical adjustment model for physical adjustment adopts a variant of the fully connected feedforward neural network, consisting of a single hidden layer with 50 neurons in each layer; use ReLU as the activation function to predict the flow field distribution value based on the energy conservation formula;

[0036] The loss function of the physical adjustment model is as follows:

[0037] LOSS N = λ f MSE(f) + λ u MSE(u) + λ du GDL(u)

[0038] where MSE(f) and MSE(u) are the mean square errors of the distribution function and velocity respectively; GDL(u) is the gradient difference loss, and λ f 、λ u 、λ du are the weights corresponding to the errors of each part and will be trained as hyperparameters in the neural network.

[0039] Furthermore, feature coupling is divided into feature sharing and feature transfer. The branch with feature sharing obtains useful information from other branches and utilizes this information in its own calculations, while feature transfer is achieved by copying features from one branch to another or by introducing connection mechanisms including skip connections.

[0040] Furthermore, when performing feature coupling, first couple the data of the two-phase flow distribution prediction, and then replace the data that does not conform to physical phenomena in the data with the data of the physical adjustment model.

[0041] Compared with the prior art, the principles and advantages of this technical solution are as follows:

[0042] This technical solution separately learns the velocities, densities, pressures or other important physical properties of the two phases and the laws of their phase boundary changes through different branches. By feature coupling, features are shared and transmitted, and the correlation and complementary feature information between different branches can be utilized. An appropriate fusion strategy is used to integrate the outputs of each branch to obtain the final two-phase flow prediction result, overcoming the disadvantages of traditional computational fluid dynamics methods, such as a large number of iterations in the calculation process, which consume a large amount of computing resources and have low computing efficiency, as well as solving problems such as poor prediction accuracy and boundary divergence of ordinary neural networks for two-phase flow. Description of the Drawings

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0044] Figure 1 It is a principle flow chart of a two-phase flow prediction method based on a branch neural network of the present invention;

[0045] Figure 2 It is a structural diagram of a neural network collision model;

[0046] Figure 3 It is a three-dimensional geometric structure and size diagram of a cross-shaped microchannel;

[0047] Figure 4 It is a schematic diagram of two-phase flow simulation in a cross-shaped microchannel based on the pseudo-potential lattice Boltzmann model;

[0048] Figure 5 It is a two-phase flow distribution diagram predicted based on a neural network collision model;

[0049] Figure 6 It is a schematic diagram of predicting two-phase flow through a branch network model. Detailed Embodiments

[0050] The present invention will be further described below in conjunction with specific embodiments:

[0051] As Figure 1 shown, a two-phase flow prediction method based on a branch neural network described in this embodiment includes the following steps:

[0052] S1. Construct a two-phase flow data acquisition model and collect two-phase flow data through this two-phase flow data acquisition model;

[0053] In this step, the constructed two-phase flow data acquisition model is the pseudo-potential lattice Boltzmann model. The basic idea of this model is to assume that the interaction between two-phase fluid particles can be represented by a potential function, and this potential function is introduced into the evolution process of the distribution function by adding a force term to the lattice Boltzmann equation.

[0054] In fact, the pseudo-potential interaction force also corresponds to the pressure tensor at the macroscopic level. The difference is that the pseudo-potential lattice Boltzmann model does not discretize the pressure tensor or the divergence of the pressure tensor phase by phase, but "from bottom to top" describes the internal interaction forces of the fluid through one simple pseudo-potential interaction force, and then reflects the relatively complex pressure tensor at the macroscopic level to achieve the description of phase separation and surface tension effects.

[0055] In the standard lattice Boltzmann equation, the discretization of time and space is linked by the discrete velocity of particles, that is, δ x = e αx δ t δ y = e αy δ t δ z = e αz δ t δ x δ y δ z are all the corresponding space step sizes, δ t is the discrete unit time step size, e αx e αy e αz are the corresponding discrete velocities. The key to constructing the lattice Boltzmann model according to actual needs is the selection of the discrete velocity model, which will affect whether it is consistent with the target macroscopic equation at the macroscopic level. The DdQm series models are regarded as the basic models of LBM and are widely used. Among them, d is the number of spatial dimensions, m is the number of discrete velocities, and the D2Q9 model and D3Q27 model are common models of LBM in two-dimensional and three-dimensional applications respectively. According to this characteristic, the particle motion in physical space is divided into two processes: collision and migration, that is:

[0056] Collision:

[0057] Migration:

[0058] Among them, taking the two-dimensional model as an example, α = 0, 1,..., 8; f α (r, t) is the distribution function, representing the state at position r and time t; represents the distribution function after collision. The collision process will adjust the particle distribution, but the position r and time t remain unchanged; is the equilibrium step function, F α (r, t) is the force term in the discrete velocity space, and τ is the dimensionless relaxation time;

[0059] That is, after a particle collides with other particles at a certain node and goes through a time step, it just moves from this node to its adjacent node; the collision process is completely local, while the migration process is linear;

[0060] F α (r, t) is obtained from the pseudo - force model lattice Boltzmann - BGK equation:

[0061]

[0062] where ω α is the weight coefficient; τ is the relaxation factor; e α is the discrete velocity; u is the velocity; c s is the lattice sound speed, and its value is F is the force exerted on the system, including the gravity F b and the pseudo - potential interaction force F m ;

[0063] The calculation formula of the pseudo - potential interaction force F m is as follows:

[0064]

[0065] where G is the interaction strength; ψ(r) is the pseudo - potential at the spatial position r; P EOS is the equation of state, which is divided into the vander Waals equation of state and the Carnahan - Starling equation; ρ is the density, and ρ0 is a constant.

[0066] By constructing the pseudo - potential lattice Boltzmann model, the data required for the branch network model can be obtained, including information such as the positions, velocities, densities, and phase interfaces of particles.

[0067] S2. Construct a branch network model, which includes three neural network branches, respectively used for predicting the two - phase flow distribution and physical regulation based on the energy conservation formula; there is an information flow path between the three neural network branches for feature coupling, so that the features belonging to different neural network branches can be cross - learned and shared;

[0068] Among the above, the neural network branch used for predicting the two - phase flow distribution respectively adopts a neural network collision model, which constitutes the core of the LBM. It describes the change of the particle distribution function at the microscopic level, thus determining the macroscopic physical properties of the fluid.

[0069] The neural network mapping Ω of the neural network collision model N Satisfies the following conditions:

[0070]

[0071] Wherein, Are the particle distributions before and after the collision respectively;

[0072] The particle distributions on all grids satisfy:

[0073]

[0074] e - and e + Are the grid velocities before and after the collision respectively.

[0075] Specifically, as Figure 2 shown, this neural network collision model adopts a variant of the fully connected feedforward neural network, consisting of two hidden layers, with 50 neurons in each layer; uses the ReLU rectified linear unit as the activation function, and there is no bias in the linear layer.

[0076] The training and learning process of the neural network collision model includes:

[0077] Symmetry input expansion: First, perform all eight symmetry operations of the dihedral group on the input state, thereby generating eight geometrically equivalent input copies; the eight symmetry operations include four rotations and four mirror reflections;

[0078] Network learning: Use the neural network collision model for learning;

[0079] Inverse transformation average output: Reverse the eight outputs back to the original direction and then take the average.

[0080] In the above, the physical adjustment model for physical adjustment adopts a variant of the fully connected feedforward neural network, consisting of one hidden layer, with 50 neurons in each layer; uses ReLU as the activation function to predict the flow field distribution value based on the energy conservation formula;

[0081] The loss function of the physical adjustment model is as follows:

[0082] LOSS N = λ f MSE(f)+λ u MSE(u)+λ du GDL(u)

[0083] Wherein, MSE(f) and MSE(u) are the mean square errors of the distribution function and velocity respectively; GDL(u) is the gradient difference loss, λ f and λ u and λdu is the weight corresponding to each part of the error, which will be used as a hyperparameter for training in the neural network.

[0084] Among the above, feature coupling is divided into feature sharing and feature transfer. The branch with feature sharing obtains useful information from other branches and utilizes this information in its own calculations, while feature transfer is achieved by copying features from one branch to another or by introducing a connection mechanism including skip connections.

[0085] When performing feature coupling, first couple the data of the two-phase flow distribution prediction, and then replace the data that does not conform to physical phenomena in the data with the data of the physical adjustment model.

[0086] S3. Train the branch network model with the two-phase flow data obtained by acquisition;

[0087] S4. Perform two-phase flow prediction with the trained branch network model.

[0088] To prove the effectiveness of the method described in the present invention, the following experiments are carried out:

[0089] Taking the cross-shaped microchannel as an example, the cross-sections of the horizontal and vertical sections of the cross-shaped microchannel are both rectangular. The lengths of the inlet microchannels of the discrete phase and the continuous phase are taken as 120 μm. To facilitate the observation and statistics of the microdroplet generation situation, the length of the fluid outlet microchannel is taken as 500 μm. Use W B the width of the horizontal pipe, use W D represent the width of the vertical pipe in the continuous phase inflow section, use W H represent the width of the microchannel outlet. After calculation and verification, when the lengths of the pipes in the continuous phase and discrete phase inflow sections exceed 3 times the pipe width, it will not affect the simulation results. The size of the droplet is represented by the maximum length L of the droplet on the central symmetry plane of the pipe, and a geometric model as shown in Figure 3 is established.

[0090] The pseudo-potential lattice Boltzmann model is used to simulate the immiscible two-phase flow. The pseudo-potential lattice Boltzmann model uses two sets of distribution functions to describe the immiscible two phases respectively. Among them, the lattice Boltzmann equation of each phase with the multiple relaxation time collision operator can be expressed as:

[0091]

[0092] In the formula, σ is the number of fluid phases, generally 2; represents the distribution function of a certain phase; represents the equilibrium distribution function of a certain phase; e i is the discrete velocity; I is the unit matrix; I is the relaxation matrix; S σ is the coefficient matrix of a certain phase; is the external force term. The multi-relaxation lattice Boltzmann equation has advantages in terms of numerical stability, parameter selection, etc., which is conducive to data preparation and processing.

[0093] Through the pseudo-potential lattice Boltzmann model, the distribution function of the cross-shaped microchannel model can be obtained, as Figure 4 shown. From this, the data required for predicting the branch network model can be obtained, and two-phase flow data can be obtained through data processing.

[0094] Finally, two-phase flow prediction is performed through the trained branch network model, as Figure 6 shown. The prediction process includes:

[0095] First, the two-phase flow distribution is predicted through the neural network collision model, and the predicted two-phase flow distribution is as Figure 5 shown.

[0096] Then, physical adjustment based on the energy conservation formula is performed through the physical adjustment model;

[0097] Finally, the distribution prediction map of two immiscible fluids in the cross-shaped microchannel is obtained through feature coupling.

[0098] From the above experiments, it can be seen that the method described in the present invention is effective.

[0099] The present invention separately learns the velocities, densities, pressures or other important physical properties of the two phases and the law of phase boundary changes through different branches. By feature coupling, features can be shared and transmitted. The correlation and complementary feature information between different branches can be utilized, and an appropriate fusion strategy can be used to integrate the outputs of each branch to obtain the final two-phase flow prediction result, overcoming the disadvantages of traditional computational fluid dynamics methods, such as a large number of iterations in the calculation process, which consume a large amount of computing resources and have low computing efficiency, and solving the problems of poor prediction accuracy and boundary divergence of ordinary neural networks for two-phase flow.

[0100] The above-described embodiments are only the preferred embodiments of the present invention, and do not limit the scope of implementation of the present invention. Therefore, any changes made according to the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A two-phase flow prediction method based on branch neural network, characterized in that: include: Constructing a two-phase flow data acquisition model, and collecting two-phase flow data through the two-phase flow data acquisition model; A branch network model is constructed, which includes three neural network branches, which are respectively used for prediction of two-phase flow distribution and physical regulation based on the energy conservation formula; an information flow path is provided between the three neural network branches for feature coupling, so that features belonging to different neural network branches are cross-learned and shared; The branch network model is trained by using the collected two-phase flow data; Two-phase flow prediction is performed using the trained branch network model.

2. A two-phase flow prediction method based on branched neural network according to claim 1, characterized in that: The two-phase flow data acquisition model is a pseudo-potential lattice Boltzmann model. In the standard lattice Boltzmann equation of the pseudo-potential lattice Boltzmann model, the discreteness of time and space is linked together through the discrete velocity of particles, that is, δ x =e αx δ t , δ y =e αy δ t , δ z =e αz δ t , δ x , δ y , δ z are the corresponding spatial steps, δ t is the discrete unit time step, e αx 、e αy 、e αz is the corresponding discrete speed. According to this characteristic, the particle motion is divided into two processes in physical space: collision and migration, namely: collision: migrate: Among them, f α (r, t) is the distribution function, which represents the state at position r and time t; represents the distribution function after the collision. The collision process will adjust the particle distribution, but the position r and time t remain unchanged; is the equilibrium step function, F α (r, t) is the force term in the discrete velocity space, and τ is the relaxation time with dimension 1; That is, after a particle collides with other particles at a certain node, it moves from that node to its adjacent node after one time step; the collision process is completely local, while the migration process is linear; F α (r, t) is obtained from the pseudo-force model lattice Boltzmann-BGK equation: Among them, ω α is the weight coefficient; τ is the relaxation factor; e α is the discrete speed; u is the speed; c s is the grid sound speed, and its value is F is the force on the system, including gravity F b and the pseudo-potential interaction force F m ; Pseudopotential interaction force F m The calculation formula is as follows: Where G is the interaction strength; ψ(r) is the pseudopotential at spatial position r; P EOS is the state equation, which is divided into the van der Waals state equation and the Carnahan-Starling equation; ρ is the density, and ρ0 is a constant.

3. The two-phase flow prediction method based on branched neural network according to claim 1 is characterized in that: The neural network branches used for two-phase flow distribution prediction adopt a neural network collision model, which adopts a variant of a fully connected feedforward neural network, consisting of two hidden layers with 50 neurons in each layer; ReLU rectified linear unit is used as the activation function, and there is no bias in the linear layer.

4. A two-phase flow prediction method based on branched neural network according to claim 3, characterized in that: The neural network map Ω of the neural network collision model N The following conditions must be met: in, are the particle distributions before and after the collision, respectively; The particle distribution on the entire grid satisfies: Distribution Conservation Conservation of Momentum e - 、e + are the grid velocities before and after the collision, respectively.

5. A two-phase flow prediction method based on branched neural network according to claim 3 or 4, characterized in that: The training and learning process of the neural network collision model includes: Symmetry input expansion: First, all eight symmetry operations of the dihedral group are performed on the input state to generate eight geometrically equivalent input copies; the eight symmetry operations include four rotations and four mirror reflections; Network learning: learning using a neural network collision model; Inverse transform average output: Reverse the eight outputs to their original directions and then take the average value.

6. The two-phase flow prediction method based on branched neural network according to claim 1 is characterized in that: The physical regulation model used for physical regulation adopts a variant of a fully connected feedforward neural network, consisting of one hidden layer with 50 neurons in each layer; ReLU is used as the activation function to predict the flow field distribution value based on the capacity conservation formula; The loss function of the physical regulation model is as follows: LOSS N =λ f MSE(f)+λ u MSE(u)+λ du GDL(u) Among them, MSE(f) and MSE(u) are the mean square errors of the distribution function and velocity respectively; GDL(u) is the gradient loss, λ f , u , du The weights corresponding to the errors of each part will be trained as hyperparameters in the neural network.

7. A two-phase flow prediction method based on branched neural network according to claim 6, characterized in that: Feature coupling is divided into feature sharing and feature transfer. Branches with feature sharing obtain useful information from other branches and utilize this information in their own calculations, while feature transfer is achieved by copying features from one branch to another, or by introducing connection mechanisms including skip connections.

8. The two-phase flow prediction method based on branched neural network according to claim 7 is characterized in that: When performing feature coupling, the data of two-phase flow distribution prediction are coupled first, and then the data that does not conform to the physical phenomenon is replaced with the data of the physical regulation model.

Images