Multi-physical field prediction method and system of trunk branch PINN architecture

Abstract

The invention relates to a multi-physics field prediction method and system of a trunk branch PINN architecture, and relates to the field of PINN model application, and the method comprises the steps: obtaining a multi-physics field prediction network architecture; obtaining a porous medium steady-state flow and heat transfer model, and configuring a physical law constraint loss function according to a physical constraint equation; collecting a plurality of groups of data, wherein any group of data comprises space coordinate data and labels for identifying physical field attribute record values; and taking the space coordinate data as input, taking a label for identifying a physical field attribute record value as output supervision, performing minimum loss optimization on a physical law constraint loss function through multiple groups of data, and training a multi-physical field prediction network architecture to generate a first multi-physical field prediction model to execute multi-physical field prediction. The technical problem that in the prior art, due to the fact that research on flow and heat transfer characteristics in a porous medium mostly focuses on the flow problem, a feasible scheme for achieving overall physical field prediction of flow and heat transfer lacks exists is solved.

Description

Technical Field

[0001] The present invention relates to the application field of PINN models, and in particular to a multi-physics field prediction method and system for a trunk-branch PINN architecture. Background Art

[0002] Flow and heat transfer in porous media have been widely used in recent years, including aerospace thermal control, fuel cells, and geothermal energy systems. Predicting the physical field information of flow and heat transfer models in porous media is helpful in understanding the flow and heat transfer characteristics in porous media, which can guide subsequent practical applications.

[0003] Currently, research on the flow and heat transfer characteristics in porous media is mostly focused on flow problems. There is no effective research on heat transfer problems, and most of the models are data-driven, resulting in a lack of feasible solutions to achieve overall physical field prediction of flow and heat transfer. Summary of the Invention

[0004] In order to solve the technical problem that the existing technology focuses on flow and heat transfer characteristics in porous media, the research has mainly focused on flow problems, resulting in a lack of feasible solutions for realizing overall physical field prediction of flow and heat transfer. The present invention provides a multi-physical field prediction method and system with a trunk-branch PINN architecture to solve the problem.

[0005] The technical solution of the present invention to solve the above technical problems is as follows: In the first aspect, the present invention provides a multi-physics field prediction method of a trunk-branch PINN architecture, comprising: obtaining a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture comprises a trunk network for extracting global features and a branch network for extracting local features connected in sequence, the trunk network is a fully connected neural network, and the branch network is an output separation fully connected neural network, comprising a number of sub-channels corresponding one-to-one to preset physical field properties; obtaining a porous medium steady-state flow and heat transfer model, and configuring a physical law constraint loss function according to the physical constraint equation of the porous medium steady-state flow and heat transfer model; collecting multiple sets of data of the porous medium steady-state flow and heat transfer model that satisfy the physical constraint equation, wherein any set of the multiple sets of data comprises spatial coordinate data and a label identifying the physical field property record value; using the spatial coordinate data as input and the label identifying the physical field property record value as output supervision, performing minimum loss optimization on the physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture to generate a first multi-physics field prediction model to perform multi-physics field prediction.

[0006] In the second aspect, the present invention provides a multi-physics field prediction system of a trunk-branch PINN architecture, comprising: a network architecture configuration module for obtaining a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture comprises a trunk network for extracting global features and a branch network for extracting local features connected in sequence, the trunk network is a fully connected neural network, and the branch network is an output separation fully connected neural network, including a number of sub-channels corresponding to preset physical field properties; a loss function construction module for obtaining a steady-state flow and heat transfer model of a porous medium, according to the physical constraints of the steady-state flow and heat transfer model of the porous medium Equation, configuring the physical law constraint loss function; a training data acquisition module, used to collect multiple sets of data of the porous medium steady-state flow and heat transfer model that satisfies the physical constraint equation, wherein any set of the multiple sets of data includes spatial coordinate data and labels that identify the physical field attribute record values; a multi-physics field prediction module, used to take the spatial coordinate data as input and the labels that identify the physical field attribute record values ​​as output supervision, and through the multiple sets of data, perform minimum loss optimization on the physical law constraint loss function, train the multi-physics field prediction network architecture to generate a first multi-physics field prediction model to perform multi-physics field prediction.

[0007] The beneficial effect of the present invention is: the local characteristics of flow and heat transfer in porous media are fitted by the branch network of the multi-physics field prediction network architecture, and then the first multi-physics field prediction model is obtained by training the loss function that fits the physical constraint equation. This can ensure that when the model is executed, the characteristics are considered from both global and local dimensions. Compared with the traditional model architecture, the characteristics are considered more comprehensively and the physical laws are satisfied more reliably, thus achieving the technical effect of realizing the prediction of the flow and heat transfer physical fields in porous media. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A schematic flow chart of a multi-physics field prediction method for a trunk-branch PINN architecture provided by the present invention; Figure 2 A schematic diagram of the structure of a multi-physics field prediction system with a trunk-branch PINN architecture provided by the present invention; Figure 3 A schematic diagram of the structure of the multi-physics field prediction model provided by the present invention; Figure 4 Schematic diagram of the porous medium steady-state flow and heat transfer model provided by the present invention; Figure 5 Schematic diagram of the training comparison between TB-net and FNN provided by the present invention; Figure 6 This is a schematic diagram of the heat transfer physical field prediction training effect provided by the present invention. DETAILED DESCRIPTION

[0009] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0010] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0011] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0012] Example 1: like Figure 1 As shown, an embodiment of the present invention provides a multi-physics field prediction method for a trunk-branch PINN architecture, comprising the steps of: S10: Obtaining a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture includes a backbone network for extracting global features and a branch network for extracting local features, which are connected in sequence, the backbone network is a fully connected neural network, and the branch network is an output separation fully connected neural network, including a plurality of sub-channels corresponding to preset physical field properties one by one; Specifically, if Figure 3 As shown in the figure, the multi-physics field prediction network architecture includes the input layer Inputs, the trunk network Trunk-net, the branch network Branch-nets and the output layer Outputs. The trunk network is a fully connected neural network, and the branch network is an output separated fully connected neural network, which contains several sub-channels, each sub-channel corresponds to a preset physical field property.

[0013] For example, a 4-layer deep backbone network is designed, with 100 neurons in each layer. The corresponding output layer and branch network are 2 layers deep, with 50 neurons in each layer. The first hidden layer in the backbone network uses a sine activation function to avoid the local minimum trap

[55] , and the remaining layers use a hyperbolic tangent activation function. In the scenario of steady-state flow and heat transfer in porous media, the input layer can be the spatial coordinates of the scenario, and the sub-channels of the output layer can correspond to the physical field characteristics of the liquid, such as horizontal and vertical velocity, liquid pressure, kinetic energy enthalpy of the liquid, and temperature of the solid.

[0014] S20: Obtain a steady-state flow and heat transfer model for porous media, and configure a physical law constraint loss function based on the physical constraint equations of the steady-state flow and heat transfer model for porous media; Specifically, if Figure 3 The porous media steady-state flow and heat transfer model refers to a model that requires physical field prediction. Preferably, a porous media steady-state flow and heat transfer model considering the LTNE effect is used. Compared with the traditional local thermal equilibrium (LTE), this model considers more realistic non-equilibrium heat transfer phenomena, that is, considering the heat convection between the solid skeleton and the fluid. For example, Figure 4 The two-dimensional porous physical model with a size of 0.1m×0.02m shown in the figure can be expressed in the general form of the physical constraint equations for incompressible, inviscid flow and heat transfer in porous media: Continuity equation: ; Momentum equation: ,: ; Enthalpy equation: Fluid enthalpy equation: ; Solid enthalpy equation: , and represent the horizontal and vertical velocities of the liquid respectively; is the density of the liquid; Indicates pressure; subscript and represents spatial partial differential; 、 and represent the porosity, absolute permeability and thermal conductivity of the solid skeleton respectively; corresponds to the temperature of the solid, and in addition, represents the specific surface area of ​​the skeleton, is the convective heat transfer coefficient, for the physical properties of the coolant, represents thermal conductivity, represents the specific heat capacity, Represents the temperature, preferably by calling , that is, the kinetic energy enthalpy of the liquid, characterizes the heat transfer parameters of the fluid and simplifies the formula.

[0015] Further, , , , and In order to predict the physical field properties, it is placed in the output layer of NN; at the same time, the spatial coordinates and Will be placed in the input layer.

[0016] Furthermore, since the OOM of the main variables varies greatly, the preset characteristic parameters further make the control equation dimensionless, and we get 、 、 、 、 、 、 ; , , , and Perform dimensionless transformation to obtain , 、 、 、 and Respectively represent the preset characteristic velocity, characteristic length, characteristic pressure, characteristic temperature and characteristic specific heat capacity. Preferably, the preset characteristic velocity, characteristic length, characteristic pressure, characteristic temperature and characteristic specific heat capacity can be dimensionless variables. The OOM should be , the functions updated by the dimensionless formula include: Continuity equation: ; Momentum equation: , ; Enthalpy equation: Fluid enthalpy equation: ; Solid enthalpy equation: ,Furthermore, the dimensionless form of the boundary conditions is as follows: Inlet boundary: 、 、 and ; Export boundary: 、 and , wall boundary: 、 、 and , represents the convective heat transfer coefficient of the coolant, defined as ,in, and are the Prandtl number and Reynolds number of the coolant, is the density of the coolant, corresponding to the heat flux applied to the outlet boundary.

[0017] The components of the TB-net PINN architecture are represented as follows: ,in, and Respectively represent the trainable parameters corresponding to the trunk and branch networks, Representation minimizes the loss function, Characterize data loss, Characterize the PDE loss, Characterize the inlet condition loss, Characterizes the loss of export conditions, Wall condition loss, in detail, , , , , ,in, 、 、 、 Specify the number of collocation points in the computational domain, on the inlet boundary, on the outlet boundary, and on the wall boundary. represents the amount of additional training data and It is a loss term obtained based on the dimensionless governing equations of porous media flow and heat transfer and the dimensionless boundary conditions, representing the constraints of physical laws, and can be written as: 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 and ,in, 、 and Characterize the selected factors of the transfer learning process, and Additional training data related to dimensionless fluid and solid temperatures are pre-set, respectively.

[0018] Using physical constraint equations, multiple loss terms are constructed. Preferably, in the embodiment of the present application, 18 loss terms are used as the example above. Then, according to the physical law combination of the loss terms, data loss, PDE loss, inlet condition loss, outlet condition loss, and wall condition loss are constructed. Furthermore, the data loss, PDE loss, inlet condition loss, outlet condition loss, and wall condition loss are added together as a physical law constraint loss function.

[0019] By limiting the loss function based on the physical constraint equation, the model prediction results can be ensured to conform to the physical laws, thereby improving the reliability of the model prediction results.

[0020] S30: collecting multiple sets of data of a porous medium steady-state flow and heat transfer model that satisfies the physical constraint equation, wherein any set of the multiple sets of data includes spatial coordinate data and a label identifying a physical field property record value; Specifically, with the physical constraint equation as the constraint, multiple sets of data of the steady-state flow and heat transfer model of porous media that meet the physical constraint equation are collected through experimental data. Any set of the multiple sets of data includes spatial coordinate data and labels that identify the physical field attribute record values. In the subsequent model training process, the labels that identify the physical field attribute record values ​​will be used as output supervision data, and the model convergence will be constrained together with the physical laws. That is, the loss function is embedded by the labels that identify the physical field attribute record values, which is the same as the traditional data-driven model, to ensure the output accuracy of the model. By embedding the physical laws into the loss function, the model output is guaranteed to conform to the physical laws.

[0021] S40: Using the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, the physical law constraint loss function is optimized for minimum loss through the multiple sets of data, and the multi-physics field prediction network architecture is trained to generate a first multi-physics field prediction model to perform multi-physics field prediction.

[0022] Specifically, if Figure 3As shown, since the implementation of the physical law constraint loss function requires a large number of partial derivative calculations, automatic differentiation AD is provided to solve it. AD is a tool for automatically calculating the differential of the partial derivatives of the output relative to the input. AD is implemented with the help of the chain rule and back propagation. Compared with finite differences or other numerical approximations, it avoids truncation errors and the results may be more accurate. Through automatic differentiation AD, physical law losses, including PDE losses, inlet condition losses, outlet condition losses, and wall condition losses, are calculated. The output data is then evaluated based on the labels that identify the recorded values ​​of the physical field properties to obtain data losses, which are then summed up to obtain the physical law constraint losses. According to the loss value calculated each time, the model parameters are adjusted through the Adam method of stochastic gradient descent and the L-BFGS method based on quasi-Newton until the multi-physics field prediction model corresponding to the minimum loss value is selected as the first multi-physics field prediction model after the preset number of iterations.

[0023] Exemplarily, the mass flux is set to , characteristic parameters 、 and , the loss function of the flow problem includes the loss term 、 、 and , the corresponding computational domain is converted to: , It is worth emphasizing that the performance of the PINNs-like framework is easily affected by the weight parameters corresponding to the loss items. Therefore, weights are assigned through OOM and importance analysis to ensure the performance of TB-net PINN. Preferably, in the basic embodiment, this application adopts conservative subjective weighting. Specifically, the loss items OOM are 、 、 、 、 、 、 and , a common practice is to normalize all loss terms OOM as , that is, the weight parameter 、 、 、 , , , , , the model training user can adjust according to the needs, the training set includes , and 100 is used to enforce boundary conditions on both sides, It is used to enforce the PDE within the computational domain. Specifically, the collocation points are randomly generated using the space-filling Latin Hypercube Sampling (LHS) strategy. In order to infer the entire solution, and in the computational domain, a 4-layer deep backbone network is designed, with 100 neurons in each layer. The branch networks corresponding to the outputs, and are 2 layers deep, with 50 neurons in each layer. The first hidden layer in the backbone network uses a sinusoidal activation function to avoid local minimum traps, and the remaining layers use a hyperbolic tangent activation function. In general, the approximation ability of the NN should adapt to the complexity of the solution space. During training, the Adam and L-BFGS optimizers are used. The Adam optimizer uses the Jacobian matrix containing the first-order derivative of the loss function, while the L-BFGS optimizer uses the Hessian matrix containing the second-order derivative of the loss function. Therefore, the L-BFGS converges faster than Adam. However, for rigid and complex scenarios, L-BFGS is more likely to fall into local minima. Therefore, in this case, the Adam optimizer is first used with a learning rate of 1e-4. Iterates to search for the global minimum and then uses the L-BFGS optimizer until a threshold criterion is reached.

[0024] Further, such as Figure 5 As shown in the figure, (a) corresponds to the case of TB-net and FNN respectively. The mass flux is set to The loss value curve; (b) corresponds to the case where the mass flux is set to The loss value curve; (c) corresponds to the case where the mass flux is set to To further demonstrate the capabilities of TB-net PINN, the training process and prediction accuracy of PINN, FNN, and TB-net are compared. In all three cases, TB-net's training process is significantly more stable and in-depth than FNN. Because FNN has more internal connections than TB-net, this suggests that TB-net consumes less memory and trains faster. Using relative error and maximum relative error as evaluation metrics, we can see that TB-net's loss values ​​are relatively small.

[0025] In summary, the TB-net PINN provided in the embodiment of the present application has the advantages of fast convergence speed, small memory usage, and small output error.

[0026] Furthermore, a porous medium steady-state flow and heat transfer model is obtained, and a physical law constraint loss function is configured according to the physical constraint equations of the porous medium steady-state flow and heat transfer model. Step S20 includes the following steps: S21: constructing a plurality of physical law loss terms according to the physical constraint equation, wherein the plurality of physical law loss terms include a plurality of flow loss terms and a plurality of heat transfer loss terms; S22: constructing a flow physical law constraint loss function according to the multiple flow loss terms; S23: constructing a heat transfer physical law constraint loss function based on the multiple heat transfer loss terms; S24: Adding the flow physical law constraint loss function and the heat transfer physical law constraint loss function into the physical law constraint loss function.

[0027] Specifically, if Figure 3 As shown, during specific training, several physical law loss terms can be generated based on the physical constraint equations and combined with the automatic differentiation AD tool, which are , in detail, several physical law loss terms have multiple flow loss terms and multiple heat transfer loss terms, and the multiple flow loss terms are 、 、 and , multiple heat transfer loss terms are 、 、 and , only the flow loss term is considered in the above-mentioned data loss, PDE loss, inlet condition loss, outlet condition loss, and wall condition loss, and then the sum is added to obtain the flow physical law constraint loss function; only the heat transfer loss term is considered in the above-mentioned data loss, PDE loss, inlet condition loss, outlet condition loss, and wall condition loss, and then the sum is added to obtain the heat transfer physical law constraint loss function, the flow physical law constraint loss function and the heat transfer physical law constraint loss function are added to the physical law constraint loss function and wait for the subsequent step training call.

[0028] Furthermore, taking the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, the physical law constraint loss function is optimized for minimum loss through the multiple sets of data, and the multi-physics field prediction network architecture is trained to generate a first multi-physics field prediction model. Step S40 includes the following steps: S41: extracting a label identifying a flow physical field property record value and a label identifying a heat transfer physical field property record value from the label identifying the physical field property record value; S42: using the spatial coordinate data as input and the labels identifying the flow physical field attribute record values ​​as output supervision, performing minimum loss optimization on the flow physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture with the heat transfer physical field sub-channel frozen to generate a first-order multi-physics field prediction model; S43: Using the spatial coordinate data as input and the label identifying the heat transfer physical field attribute record value as output supervision, the heat transfer physical law constraint loss function is optimized for minimum loss through the multiple sets of data, the first-order multi-physical field prediction model of the activated heat transfer physical field sub-channel is trained to generate the first multi-physical field prediction model.

[0029] Specifically, the embodiment of the present application divides the training process into two stages based on the staged characteristics of flow and heat transfer in the steady-state flow and heat transfer model of porous media. Therefore, in traditional physical field prediction, it is always hoped to predict all physical field parameters at one time. However, since the physical field of heat transfer in the steady-state flow and heat transfer model of porous media needs to be calculated based on the flow parameters, the traditional method may be able to handle the flow prediction of the first stage, but there will be large errors when facing the heat transfer physical field prediction of the second stage.

[0030] Therefore, in an embodiment of the present application, first, the label identifying the flow physical field property record value and the label identifying the heat transfer physical field property record value are extracted from the label identifying the physical field property record value, and the spatial coordinate data is used as input, and the label identifying the flow physical field property record value is used as output supervision. Through the multiple sets of data, the minimum loss of the flow physical law constraint loss function is optimized, and the multi-physical field prediction network architecture with the heat transfer physical field sub-channel frozen is trained to generate a first-order multi-physical field prediction model; further, the spatial coordinate data is used as input, and the label identifying the heat transfer physical field property record value is used as output supervision, and the multiple sets of data are used to perform minimum loss optimization on the heat transfer physical field constraint loss function, and the first-order multi-physical field prediction model with the heat transfer physical field sub-channel activated is trained to generate the first multi-physical field prediction model and perform two-stage training, thereby obtaining the target model and ensuring the prediction accuracy.

[0031] For example, Figure 6 As shown, based on the aforementioned pressure prediction model, it is assumed that the heat flux applied to the outlet boundary is , the temperature of the injected coolant is , so the characteristic temperature for , the kinetic energy enthalpy is , the rest of the characteristic parameters are the same. Similarly, based on the OOM and the importance of each loss term, weighting is performed to obtain the heat transfer physical law constraint loss function. The branch network related to the heat transfer physical field is set to 4 layers deep, with 100 neurons per layer. The activation function of the first hidden layer is set to sine to avoid local minimum. Then, the backbone network and flow-related branch network of the previously trained model are frozen, and the training work is only used for the branch network related to the heat transfer physical field to obtain the training results. Figure 6It can be seen that (a) the solid temperature predicted by the exact and TB-net (b) Fluid temperature T predicted by the exact and TB-net f and their corresponding absolute errors and relative errors are and , and the corresponding maximum relative errors are and , so optimizing the energy equation loss with the help of pre-trained flow models is an effective way to identify complex heat transfer trends.

[0032] The method further includes step S50, which includes the following steps: S51: when the physical constraint equation of the porous medium steady-state flow and heat transfer model changes, obtaining an updated physical constraint equation configuration and updating a physical law constraint loss function; S52: Collect multiple sets of data of the steady-state flow and heat transfer model of porous media that meets the updated physical constraint equation, perform minimum loss optimization on the updated physical law constraint loss function, train the first multi-physics field prediction model to generate a second multi-physics field prediction model to perform multi-physics field prediction.

[0033] Furthermore, a porous medium steady-state flow and heat transfer model is obtained, and a physical law constraint loss function is configured according to the physical constraint equations of the porous medium steady-state flow and heat transfer model. Step S20 includes the following steps: S21: performing order of magnitude weight analysis on the plurality of physical law loss terms to obtain order of magnitude weight distribution; S22: Sending the order of magnitude weight distribution to the training management terminal to obtain the loss item weight distribution; S23: According to the weight distribution of the loss terms, weight configuration is performed on the multiple physical law loss terms, and the physical law constraint loss function is constructed by combining the multiple physical law loss terms.

[0034] Specifically, this step provides a process for configuring the weights of each loss term, performing order of magnitude weight analysis on the several physical law loss terms, and obtaining order of magnitude weight distribution, that is, performing order of magnitude analysis on the several physical law loss terms, obtaining the order of magnitude parameters of each term, and then first normalizing the loss term to O(10 0 ) is the order of magnitude weight distribution of each loss item.

[0035] Furthermore, the order of magnitude weight distribution is sent to the training management end, and the weight is adjusted by the management end to obtain the subjectively assigned loss item weight distribution to configure the weights of the several physical law loss items. Combined with the several physical law loss items, the physical law constraint loss function is constructed.

[0036] Furthermore, the method further includes step S40, wherein the execution steps include: When the number of training times exceeds the preset number, a physical field characteristic value deviation record sequence, a first loss item deviation record sequence, and an Nth loss item deviation record sequence are obtained; Normalizing the physical field characteristic value deviation record sequence to obtain a reference sequence; Normalizing the first loss item deviation record sequence to the Nth loss item deviation record sequence to obtain a first alignment sequence and a second alignment sequence to the Nth alignment sequence; Constructing a grey correlation matrix based on the reference sequence, the first comparison sequence, the second comparison sequence, and the Nth comparison sequence to perform grey correlation analysis to obtain first correlations to Nth correlations; Adding the first association degree up to the Nth association degree to obtain an association degree sum value; Traversing the first association degree until the Nth association degree is compared with the sum of the association degrees to obtain an objective weight distribution of importance; The mean weight is calculated based on the objective weight distribution of importance and the weight distribution of loss item to obtain the updated weight distribution of loss item, and the weight of the physical law constraint loss function is updated to continue training. At the same time, the counter starts counting from 0. When the number of training times exceeds the preset number and still has not converged, the weight is updated again.

[0037] Specifically, in a possible embodiment, in order to improve the output accuracy, the present application proposes that after the number of training times exceeds a preset number of times, which is 200 times by default, the physical field characteristic value deviation record sequence, the first loss item deviation record sequence, and the Nth loss item deviation record sequence of each training are extracted; the physical field characteristic value deviation record sequence is then normalized to obtain a reference sequence; each sequence from the first loss item deviation record sequence to the Nth loss item deviation record sequence is normalized to obtain a first comparison sequence, a second comparison sequence, and the Nth comparison sequence; further, the reference sequence is used as the first column of data, and the first comparison sequence, the second comparison sequence, and the Nth comparison sequence are respectively set as the second column of data and the N+1th column of data to obtain the gray correlation coefficient. Degree matrix, and then perform grey correlation analysis to obtain the correlation parameter of each loss item, that is, the first correlation degree up to the Nth correlation degree; then add the first correlation degree up to the Nth correlation degree to obtain the correlation degree sum value; traverse the first correlation degree until the Nth correlation degree and compare it with the correlation degree sum value to obtain the importance objective weight distribution; perform mean weight calculation based on the importance objective weight distribution and the loss item weight distribution to obtain the updated loss item weight distribution, update the weight of the physical law constraint loss function and continue to train, and at the same time, the counter starts counting from 0. When the number of training times exceeds the preset number and still has not converged, the weight is updated again, that is, as long as it has not converged, the weight is updated every time the preset number is met, and the counter starts counting from 0.

[0038] The correlation between loss terms and output deviations is analyzed through grey correlation, which is then used as an objective weighting standard to dynamically increase the weights of more important loss terms to ensure that more attention is paid to loss terms that have a greater impact on physical field predictions, which is conducive to improving output accuracy.

[0039] Furthermore, step S40 is included, and the execution steps further include: When the training fails to converge after exceeding a population optimization number threshold, multiple sets of model record parameters of the multi-physics field prediction network architecture are obtained, wherein any set of model record parameters includes a neuron record weight set and a neuron record bias set, and the neuron record weight set and the neuron record bias set have a one-to-one correspondence; Using a physical law constraint loss function as a fitness function, sorting the multiple groups of model record parameters according to loss values ​​from large to small to obtain a model record parameter sequence; Extracting a first number of tail model record parameters of the model record parameter sequence as adjustment targets, performing Euclidean distance reduction adjustment on a second number of head model record parameters of the model record parameter sequence, and obtaining updated model parameters; Training is performed according to the updated model parameters.

[0040] Specifically, in a possible embodiment, in order to improve the convergence efficiency of the model, the present application proposes that when the training exceeds the population optimization times threshold and still has not converged, the default is 500 times, multiple sets of model recording parameters of the multi-physics field prediction network architecture are obtained, wherein any set of model recording parameters includes a neuron recording weight set and a neuron recording bias set, and the neuron recording weight set and the neuron recording bias set correspond one to one.

[0041] Construct the Euclidean distance evaluation formula for model parameters: ,in, The model parameters that characterize any two groups, Characterizes the number of internal nodes of the model, The z-th node model weight value representing the first model parameter, The z-th node model bias representing the first model parameter, The z-th node model weight value representing the second model parameter, The z-th node model bias representing the second model parameter, The important parameters that characterize the weights, The important parameters that characterize the bias are configured subjectively by humans. The Euclidean distance evaluation formula of the model parameters is used to evaluate the physical law constraint loss function as the fitness function. The multiple groups of model record parameters are sorted from large to small according to the loss value to obtain a model record parameter sequence. The first number of tail model record parameters of the model record parameter sequence are extracted as adjustment targets, and the Euclidean distance reduction adjustment is performed on the second number of head model record parameters of the model record parameter sequence to obtain updated model parameters. Training is performed based on the updated model parameters. Combined with the swarm optimization algorithm, the probability of obtaining the optimal solution is increased, so as to help improve training efficiency.

[0042] The parameters of the first number of tail model records with better performance are used as the guiding target, and the Euclidean distance reduction adjustment is performed on the parameters of the second number of head model records with worse performance to ensure that the traversal space is large enough. At the same time, the probability of an optimal solution existing near the parameters of the first number of tail model records is higher, thereby increasing the probability of obtaining the optimal solution, thereby taking into account both globality and convergence efficiency.

[0043] The embodiment of the present invention provides a multi-physics field prediction method for a trunk-branch PINN architecture, which has at least the following technical effects: The local characteristics of flow and heat transfer in porous media are fitted through the branch network of the multi-physics field prediction network architecture, and then the first multi-physics field prediction model is obtained by training with the loss function that fits the physical constraint equation. This ensures that when the model is executed, the characteristics are considered from both global and local dimensions. Compared with the traditional model architecture, the characteristics are considered more comprehensively and the physical laws are satisfied more reliably, thus achieving the technical effect of realizing the prediction of the flow and heat transfer physical fields in porous media.

[0044] The correlation between loss terms and output deviations is analyzed through grey correlation, which is then used as an objective weighting standard to dynamically increase the weights of more important loss terms to ensure that more attention is paid to loss terms that have a greater impact on physical field predictions, which is conducive to improving output accuracy.

[0045] The parameters of the first number of tail model records with better performance are used as the guiding target, and the Euclidean distance reduction adjustment is performed on the parameters of the second number of head model records with worse performance to ensure that the traversal space is large enough. At the same time, the probability of an optimal solution existing near the parameters of the first number of tail model records is higher, thereby increasing the probability of obtaining the optimal solution, thereby taking into account both globality and convergence efficiency.

[0046] Example 2: like Figure 2 As shown, based on the same inventive concept as the multi-physics field prediction method of the trunk-branch PINN architecture provided in the first embodiment, the embodiment of the present invention further provides a multi-physics field prediction system of the trunk-branch PINN architecture, including: A network architecture configuration module is used to obtain a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture includes a backbone network for extracting global features and a branch network for extracting local features, which are connected in sequence. The backbone network is a fully connected neural network, and the branch network is an output separation fully connected neural network, including a plurality of sub-channels corresponding to preset physical field properties. The loss function construction module is used to obtain the steady-state flow and heat transfer model of porous media. According to the physical constraint equations of the steady-state flow and heat transfer model of porous media, the physical law constraint loss function is configured; a training data acquisition module, configured to acquire multiple sets of data of a porous medium steady-state flow and heat transfer model satisfying the physical constraint equations, wherein any set of the multiple sets of data includes spatial coordinate data and labels identifying recorded values ​​of physical field properties; The multi-physics field prediction module is used to take the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, and through the multiple sets of data, perform minimum loss optimization on the physical law constraint loss function, train the multi-physics field prediction network architecture to generate a first multi-physics field prediction model to perform multi-physics field prediction.

[0047] Furthermore, a steady-state flow and heat transfer model of porous media is obtained. According to the physical constraint equations of the steady-state flow and heat transfer model of porous media, a physical law constraint loss function is configured, including: constructing a plurality of physical law loss terms according to the physical constraint equation, wherein the plurality of physical law loss terms include a plurality of flow loss terms and a plurality of heat transfer loss terms; constructing a flow physics law constraint loss function based on the multiple flow loss terms; Constructing a heat transfer physical law constraint loss function based on the multiple heat transfer loss terms; The flow physics law constraint loss function and the heat transfer physics law constraint loss function are added to the physics law constraint loss function.

[0048] Furthermore, taking the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, performing minimum loss optimization on the physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture to generate a first multi-physics field prediction model, including: Extracting a label identifying a flow physical field property record value and a label identifying a heat transfer physical field property record value from the label identifying the physical field property record value; Taking the spatial coordinate data as input and the labels identifying the recorded values ​​of the flow physical field attributes as output supervision, the flow physical law constraint loss function is optimized for minimum loss through the multiple sets of data, and the multi-physics field prediction network architecture with the heat transfer physical field sub-channel frozen is trained to generate a first-order multi-physics field prediction model; Taking the spatial coordinate data as input and the label identifying the heat transfer physical field property record value as output supervision, the heat transfer physical law constraint loss function is optimized for minimum loss through the multiple sets of data, the first-order multi-physical field prediction model of the activated heat transfer physical field sub-channel is trained to generate the first multi-physical field prediction model.

[0049] Furthermore, a second multi-physics field prediction model construction module is included, and the execution steps include: When the physical constraint equation of the porous medium steady-state flow and heat transfer model changes, obtaining an updated physical constraint equation configuration and updating a physical law constraint loss function; Collect multiple sets of data of the porous medium steady-state flow and heat transfer model that satisfies the updated physical constraint equation, perform minimum loss optimization on the updated physical law constraint loss function, train the first multi-physics field prediction model to generate a second multi-physics field prediction model to perform multi-physics field prediction.

[0050] Furthermore, a steady-state flow and heat transfer model of porous media is obtained. According to the physical constraint equations of the steady-state flow and heat transfer model of porous media, a physical law constraint loss function is configured, including: Performing an order of magnitude weight analysis on the plurality of physical law loss terms to obtain an order of magnitude weight distribution; Sending the order of magnitude weight distribution to the training management end to obtain the loss item weight distribution; According to the weight distribution of the loss terms, the weights of the multiple physical law loss terms are configured, and the physical law constraint loss function is constructed by combining the multiple physical law loss terms.

[0051] Furthermore, a loss function weight configuration module is included, and the execution steps include: When the number of training times exceeds the preset number, a physical field characteristic value deviation record sequence, a first loss item deviation record sequence, and an Nth loss item deviation record sequence are obtained; Normalizing the physical field characteristic value deviation record sequence to obtain a reference sequence; Normalizing the first loss item deviation record sequence to the Nth loss item deviation record sequence to obtain a first alignment sequence and a second alignment sequence to the Nth alignment sequence; Constructing a grey correlation matrix based on the reference sequence, the first comparison sequence, the second comparison sequence, and the Nth comparison sequence to perform grey correlation analysis to obtain first correlations to Nth correlations; Adding the first association degree up to the Nth association degree to obtain an association degree sum value; Traversing the first association degree until the Nth association degree is compared with the sum of the association degrees to obtain an objective weight distribution of importance; The mean weight is calculated based on the objective weight distribution of importance and the weight distribution of loss item to obtain the updated weight distribution of loss item, and the weight of the physical law constraint loss function is updated to continue training. At the same time, the counter starts counting from 0. When the number of training times exceeds the preset number and still has not converged, the weight is updated again.

[0052] Furthermore, taking the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, performing minimum loss optimization on the physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture to generate a first multi-physics field prediction model, including: When the training fails to converge after exceeding a population optimization number threshold, multiple sets of model record parameters of the multi-physics field prediction network architecture are obtained, wherein any set of model record parameters includes a neuron record weight set and a neuron record bias set, and the neuron record weight set and the neuron record bias set have a one-to-one correspondence; Using a physical law constraint loss function as a fitness function, sorting the multiple groups of model record parameters according to loss values ​​from large to small to obtain a model record parameter sequence; Extracting a first number of tail model record parameters of the model record parameter sequence as adjustment targets, performing Euclidean distance reduction adjustment on a second number of head model record parameters of the model record parameter sequence, and obtaining updated model parameters; Training is performed according to the updated model parameters.

[0053] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0054] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0055] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0056] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0057] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0058] Although preferred embodiments of the present invention have been described, additional changes and modifications to these embodiments may occur to those skilled in the art once the basic inventive concepts become known.

[0059] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A multi-physics field prediction method based on a trunk-branch PINN architecture, characterized in that: include: Obtaining a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture includes a backbone network for extracting global features and a branch network for extracting local features, which are connected in sequence, the backbone network is a fully connected neural network, and the branch network is an output separation fully connected neural network, including a plurality of sub-channels corresponding one-to-one to preset physical field properties; Obtain a steady-state flow and heat transfer model for porous media, and configure a physical law constraint loss function based on the physical constraint equations of the steady-state flow and heat transfer model for porous media; Collecting multiple sets of data of a porous medium steady-state flow and heat transfer model that satisfies the physical constraint equations, wherein any set of the multiple sets of data includes spatial coordinate data and labels identifying recorded values ​​of physical field properties; Taking the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, the physical law constraint loss function is optimized for minimum loss through the multiple sets of data, and the multi-physics field prediction network architecture is trained to generate a first multi-physics field prediction model to perform multi-physics field prediction.

2. The method according to claim 1, wherein Obtain a steady-state flow and heat transfer model for porous media. Based on the physical constraint equations of the steady-state flow and heat transfer model for porous media, configure a physical law constraint loss function, including: constructing a plurality of physical law loss terms according to the physical constraint equation, wherein the plurality of physical law loss terms include a plurality of flow loss terms and a plurality of heat transfer loss terms; constructing a flow physics law constraint loss function based on the multiple flow loss terms; Constructing a heat transfer physical law constraint loss function based on the multiple heat transfer loss terms; The flow physics law constraint loss function and the heat transfer physics law constraint loss function are added to the physics law constraint loss function.

3. The method according to claim 2, wherein Taking the spatial coordinate data as input and the labels identifying the recorded values ​​of the physical field attributes as output supervision, performing minimum loss optimization on the physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture to generate a first multi-physics field prediction model, including: Extracting a label identifying a flow physical field property record value and a label identifying a heat transfer physical field property record value from the label identifying the physical field property record value; Taking the spatial coordinate data as input and the labels identifying the recorded values ​​of the flow physical field attributes as output supervision, the flow physical law constraint loss function is optimized for minimum loss through the multiple sets of data, and the multi-physics field prediction network architecture with the heat transfer physical field sub-channel frozen is trained to generate a first-order multi-physics field prediction model; Taking the spatial coordinate data as input and the label identifying the heat transfer physical field property record value as output supervision, the heat transfer physical law constraint loss function is optimized for minimum loss through the multiple sets of data, the first-order multi-physical field prediction model of the activated heat transfer physical field sub-channel is trained to generate the first multi-physical field prediction model.

4. The method according to claim 1, wherein Also includes: When the physical constraint equation of the porous medium steady-state flow and heat transfer model changes, obtaining an updated physical constraint equation configuration and updating a physical law constraint loss function; Collect multiple sets of data of the porous medium steady-state flow and heat transfer model that satisfies the updated physical constraint equation, perform minimum loss optimization on the updated physical law constraint loss function, train the first multi-physics field prediction model to generate a second multi-physics field prediction model to perform multi-physics field prediction.

5. The method according to claim 2, wherein Obtain a steady-state flow and heat transfer model for porous media. Based on the physical constraint equations of the steady-state flow and heat transfer model for porous media, configure a physical law constraint loss function, including: Performing an order of magnitude weight analysis on the plurality of physical law loss terms to obtain an order of magnitude weight distribution; Sending the order of magnitude weight distribution to the training management end to obtain the loss item weight distribution; According to the weight distribution of the loss terms, the weights of the multiple physical law loss terms are configured, and the physical law constraint loss function is constructed by combining the multiple physical law loss terms.

6. The method according to claim 5, wherein Also includes: When the number of training times exceeds the preset number, a physical field characteristic value deviation record sequence, a first loss item deviation record sequence, and an Nth loss item deviation record sequence are obtained; Normalizing the physical field characteristic value deviation record sequence to obtain a reference sequence; Normalizing the first loss item deviation record sequence to the Nth loss item deviation record sequence to obtain a first alignment sequence and a second alignment sequence to the Nth alignment sequence; Constructing a grey correlation matrix based on the reference sequence, the first comparison sequence, the second comparison sequence, and the Nth comparison sequence to perform grey correlation analysis to obtain first correlations to Nth correlations; Adding the first association degree up to the Nth association degree to obtain an association degree sum value; Traversing the first association degree until the Nth association degree is compared with the sum of the association degrees to obtain an objective weight distribution of importance; The mean weight is calculated based on the objective weight distribution of importance and the weight distribution of loss item to obtain the updated weight distribution of loss item, and the weight of the physical law constraint loss function is updated to continue training. At the same time, the counter starts counting from 0. When the number of training times exceeds the preset number and still has not converged, the weight is updated again.

7. The method according to claim 1, wherein Taking the spatial coordinate data as input and the labels identifying the recorded values ​​of the physical field attributes as output supervision, performing minimum loss optimization on the physical law constraint loss function through the multiple sets of data, and training the multi-physics field prediction network architecture to generate a first multi-physics field prediction model, including: When the training fails to converge after exceeding a population optimization number threshold, multiple sets of model record parameters of the multi-physics field prediction network architecture are obtained, wherein any set of model record parameters includes a neuron record weight set and a neuron record bias set, and the neuron record weight set and the neuron record bias set have a one-to-one correspondence; Using a physical law constraint loss function as a fitness function, sorting the multiple groups of model record parameters according to loss values ​​from large to small to obtain a model record parameter sequence; Extracting a first number of tail model record parameters of the model record parameter sequence as adjustment targets, performing Euclidean distance reduction adjustment on a second number of head model record parameters of the model record parameter sequence, and obtaining updated model parameters; Training is performed according to the updated model parameters.

8. A multi-physics field prediction system with a trunk-branch PINN architecture, characterized in that: Used to implement the method according to any one of claims 1 to 7, comprising: A network architecture configuration module is used to obtain a multi-physics field prediction network architecture, wherein the multi-physics field prediction network architecture includes a backbone network for extracting global features and a branch network for extracting local features, which are connected in sequence. The backbone network is a fully connected neural network, and the branch network is an output separation fully connected neural network, including a plurality of sub-channels corresponding to preset physical field properties. The loss function construction module is used to obtain the steady-state flow and heat transfer model of porous media. According to the physical constraint equations of the steady-state flow and heat transfer model of porous media, the physical law constraint loss function is configured; a training data acquisition module, configured to acquire multiple sets of data of a porous medium steady-state flow and heat transfer model satisfying the physical constraint equations, wherein any set of the multiple sets of data includes spatial coordinate data and labels identifying recorded values ​​of physical field properties; The multi-physics field prediction module is used to take the spatial coordinate data as input and the label identifying the physical field attribute record value as output supervision, and through the multiple sets of data, perform minimum loss optimization on the physical law constraint loss function, train the multi-physics field prediction network architecture to generate a first multi-physics field prediction model to perform multi-physics field prediction.

Images