Large eddy simulation method and system based on mechanism-only model

Abstract

The invention provides a large eddy simulation method and system based on a mechanism-only model. The method comprises the following steps: acquiring a current dictionary database of a target turbulent flow field; the current dictionary database comprises multi-dimensional flow field features in a current time step; performing feature screening on the current dictionary database to obtain key flow field features of the target turbulent flow field; inputting the key flow field features into a mechanism-only model, and reconstructing to obtain the sublattice turbulence viscosity of the target turbulence field in the current time step; wherein the mechanism-only model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulence field and a sublattice turbulence viscosity label of the sample turbulence field. According to the method, the simulation precision of small-scale turbulence vortex motion in large vortex simulation is improved, so that the problems of poor performance of a bottom layer model and low simulation calculation precision caused by calculation grid scale limitation and sub-grid stress simplification hypothesis are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of fluid mechanics simulation, and in particular to a large eddy simulation method and system based on a mechanism phenomenological model. Background Art

[0002] At present, large-eddy simulation (LES) can simulate large-scale turbulent motion on a grid with relatively low resolution and can resolve some turbulent structures at a low computational cost. It is an important numerical simulation method in turbulence research. Therefore, how to effectively implement large-eddy simulation is one of the technical problems that the industry urgently needs to solve.

[0003] In order to approximate the simulation of small-scale turbulent eddies below the computational grid scale, related technologies usually establish a series of classic models based on the turbulent large eddy simulation framework according to the traditional subgrid-scale stress (SGS) assumption to achieve large eddy simulation. However, due to the limitations of the computational grid scale and the simplified assumptions about subgrid stress, it is difficult to accurately capture all turbulent details, which leads to poor performance of the underlying model and low simulation calculation accuracy.

[0004] Therefore, there is an urgent need for a large eddy simulation method and system based on a mechanism-based phenomenological model to solve the above technical problems. Summary of the invention

[0005] The present invention provides a large eddy simulation method and system based on a mechanism phenomenological model, which is used to solve the defects of the prior art that a series of classical models based on a turbulent large eddy simulation framework are established according to traditional sub-grid stress to perform large eddy simulation, resulting in poor performance of the underlying model and low simulation calculation accuracy, thereby improving the accuracy of large eddy simulation.

[0006] The present invention provides a large eddy simulation method based on a mechanism phenomenological model, comprising: Acquire a current dictionary library of the target turbulent flow field; the current dictionary library includes multi-dimensional flow field characteristics within the current time step; Performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; Inputting the key flow field characteristics into the mechanism phenomenological model, and reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step; The mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and sub-grid turbulent viscosity labels of the sample turbulent field.

[0007] According to a large eddy simulation method based on a mechanism phenomenological model provided by the present invention, the mechanism phenomenological model is obtained by training based on the following steps: According to the Navier-Stokes equation, the sample turbulence field is simulated and calculated to obtain direct numerical simulation data of the sample turbulence field; Filtering the direct numerical simulation data, and constructing a sample dictionary library of the sample turbulent field according to the filtered direct numerical simulation data; Performing feature screening on the sample dictionary library to obtain key flow field features of the sample turbulent field; Performing sub-grid turbulent viscosity labeling on the sample turbulent field to obtain the sub-grid turbulent viscosity label; Constructing a sample data set according to the key flow field characteristics of the sample turbulent field and the sub-grid turbulent viscosity label; According to the sample data set, obtaining the loss function of the initial neural network model; According to the loss function, the initial neural network model is iteratively trained to obtain the mechanism phenomenological model.

[0008] According to a large eddy simulation method based on a mechanism phenomenological model provided by the present invention, filtering the direct numerical simulation data, and constructing a sample dictionary library of the sample turbulent field according to the filtered direct numerical simulation data, comprising: In a filtering algorithm library, a target filtering algorithm is determined; the filtering algorithm library includes a plurality of filtering algorithms with different filtering scales and different filtering forms; filtering the direct numerical simulation data according to the filtering scale and filtering form corresponding to the target filtering algorithm; The filtered direct numerical simulation data are operated according to a large eddy simulation control model, a Smagorinsky model, and an equivalent equation of a subgrid stress tensor in the large eddy simulation control model; According to the operation result, the sample dictionary library is constructed.

[0009] According to a large eddy simulation method based on a mechanism phenomenological model provided by the present invention, in the filtering algorithm library, determining a target filtering algorithm comprises: According to the simulation accuracy, computing resource consumption and filtered data distribution characteristics of each filtering algorithm in the filtering algorithm library, the performance score corresponding to each filtering algorithm is obtained; In the filtering algorithm library, a filtering algorithm with the highest performance score is determined as the target filtering algorithm.

[0010] According to a large eddy simulation method based on a mechanism phenomenological model provided by the present invention, the feature screening of the sample dictionary library to obtain the key flow field features of the sample turbulent field includes: Based on a sparse Bayesian learning model, simulating the nonlinear relationship between the multidimensional flow field characteristics in the sample dictionary library and the sub-grid turbulent viscosity label; Determining, according to the nonlinear relationship, the correlation parameters between the flow field characteristics of each dimension in the sample dictionary library and the sub-grid turbulent viscosity label; According to the association parameters, the multi-dimensional flow field features in the sample dictionary library are screened to obtain the key flow field features of the sample turbulent field.

[0011] According to a large eddy simulation method based on a mechanism phenomenological model provided by the present invention, the method further comprises: Solving the sub-grid turbulent viscosity of the target turbulent field in the current time step to obtain the next dictionary library; Iteratively reconstructing the sub-grid turbulent viscosity in the next time step according to the next dictionary library until the reconstructed sub-grid turbulent viscosity converges; The flow field simulation is performed according to the dictionary library solved in each time step and the reconstructed sub-grid turbulent viscosity to obtain the simulated flow field of the target turbulent field.

[0012] The present invention also provides a large eddy simulation system based on a mechanism phenomenological model, comprising: An acquisition unit, used to acquire a current dictionary library of a target turbulent flow field; the current dictionary library includes multi-dimensional flow field characteristics within a current time step; A selection unit, used for performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; A reconstruction unit, used for inputting the key flow field characteristics into the mechanism phenomenological model, and reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step; The mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and sub-grid turbulent viscosity labels of the sample turbulent field.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a large eddy simulation method based on a mechanism phenomenological model as described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the large eddy simulation method based on the mechanism phenomenological model as described in any one of the above is implemented.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the large eddy simulation method based on the mechanism phenomenological model as described above is implemented.

[0016] The large eddy simulation method and system based on the mechanism phenomenological model provided by the present invention obtain a current time step turbulence field dictionary library containing multi-dimensional flow field characteristics, extract key flow field characteristics through feature screening, and then input these characteristics into the mechanism phenomenological model obtained by adaptive training based on direct numerical simulation data and sub-grid turbulent viscosity labels, so as to more accurately capture the turbulence details through the mechanism phenomenological model, thereby adaptively and accurately reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step, thereby improving the simulation accuracy of small-scale turbulent eddies in large eddy simulation, and effectively solving the problems of poor underlying model performance and low simulation calculation accuracy caused by computational grid scale limitations and sub-grid stress simplification assumptions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is one of the flow diagrams of the large eddy simulation method based on the mechanism phenomenological model provided by the present invention.

[0019] Figure 2 This is the second flow chart of the large eddy simulation method based on the mechanism phenomenological model provided by the present invention.

[0020] Figure 3 It is a schematic diagram of a grid in a two-dimensional square column flow scenario provided by the present invention.

[0021] Figure 4 It is a schematic diagram of the principle of key feature screening based on the sparse Bayesian learning algorithm provided by the present invention.

[0022] Figure 5 It is a flow chart of the mechanism-phenomenological model and solver fusion simulation provided by the present invention.

[0023] Figure 6 It is a schematic diagram of the selection of points along the X direction of the center line of the flow field key coordinates provided by the present invention in the calculation domain.

[0024] Figure 7 A line graph showing the horizontal velocity changes of key coordinates within a 28 second range in the simulation results of different models provided by the present invention.

[0025] Figure 8 It is a time-averaged distribution diagram of the fluid velocity value in the X-axis direction under the key coordinates of the flow field in the simulation results of different models provided by the present invention.

[0026] Fig. 9 It is a time-averaged distribution diagram of the fluid velocity value in the Y-axis direction under the key coordinates of the flow field in the simulation results of different models provided by the present invention.

[0027] Fig.10 It is a time-average distribution diagram of fluid pressure values ​​at key coordinates of the flow field in different model simulation results provided by the present invention.

[0028] Fig.11 It is a structural schematic diagram of a large eddy simulation system based on a mechanism phenomenological model provided by the present invention.

[0029] Fig.12 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Industrial software is the core of intelligent manufacturing and the cornerstone for promoting the transformation and upgrading of industry towards digitalization, networking and intelligence. Computational Fluid Dynamics (CFD), as a typical industrial software, involves key industries such as aerospace, shipping, and automobile manufacturing, and is an important tool to ensure equipment design and verification. Turbulence research in CFD is crucial for a deep understanding of the complexity of fluid flow and improving the reliability and efficiency of engineering design. At present, large eddy simulation can simulate large-scale turbulent motion on a grid with relatively low resolution, and can analyze some turbulent structures at a low computational cost. It is an important numerical simulation method in turbulence research. Therefore, how to effectively implement large eddy simulation is one of the technical problems that the industry is currently facing.

[0032] In order to approximate the simulation of small-scale turbulent eddies below the computational grid scale, related technologies usually establish a series of classic models based on the turbulent large eddy simulation framework according to the traditional sub-grid stress assumption to achieve large eddy simulation. However, due to the limitations of the computational grid scale and the simplified assumptions of sub-grid stress, it is difficult to accurately capture all turbulent details, resulting in problems such as poor performance of the underlying model, large relative errors, and excessive dissipation.

[0033] In recent years, artificial intelligence technology has developed by leaps and bounds, providing an important premise for artificial intelligence (AI) to empower fluid mechanics. In this context, how to combine traditional fluid mechanics modeling with artificial intelligence, realize the efficient fusion mechanism and technical upgrade of AI modeling and traditional CFD software, how to use AI empowerment to explore the combination of key characteristic physical quantities, optimize the mechanism phenomenological structure of the turbulent LES equation, and improve the LES equation's ability to characterize turbulence is the key to achieving the numerical simulation accuracy of the flow field control equation. Therefore, empowering the underlying model of CFD software with artificial intelligence and improving the simulation accuracy of the LES turbulence control equation can achieve a comprehensive improvement in the performance of traditional fluid mechanics simulation software, which is of great significance to industries such as shipping, aerospace, etc.

[0034] In this regard, in order to solve the problems of poor underlying model performance and low simulation calculation accuracy in the traditional CFD method in the field of fluid simulation, the embodiment of the present application provides a large eddy simulation method based on a mechanism phenomenological model. The method establishes a mechanism for CFD solver and AI fusion simulation in fluid mechanics, and uses a modeling technology combining sparse Bayesian learning and neural network architecture to train a neural network-based mechanism-phenomenological model. Under the premise of ensuring the efficiency of reasoning, the number and complexity of model parameters are reduced by screening key characteristic physical quantities, thereby realizing turbulent viscosity. The reconstruction and optimization of the model solves the problems of poor underlying model performance and low simulation calculation accuracy in the existing large eddy simulation method of fluid mechanics.

[0035] Figure 1 It is one of the flow diagrams of the large eddy simulation method based on the mechanism phenomenological model provided by the present invention, such as Figure 1 As shown, the method includes step 110 , step 120 and step 130 .

[0036] Step 110, obtaining a current dictionary library of the target turbulent flow field; the current dictionary library includes multi-dimensional flow field features within the current time step.

[0037] Optionally, during the large eddy simulation of the target turbulent field, the sub-grid turbulent viscosity of the target turbulent field in the previous time step can be optimized and solved to obtain the multi-dimensional flow field characteristics in the current time step, thereby obtaining the current dictionary library of the target turbulent field. The so-called current dictionary library contains the multi-dimensional flow field characteristics in the current time step, such as the flow field velocity filter values ​​in different directions. and , flow field pressure filter value , flow field velocity filtering values ​​in different directions and Derivatives of various orders in different directions, flow field velocity filtering values ​​in different directions and Combined derivatives of various orders and flow field pressure filtering values The various order derivatives of , etc., are not specifically limited in this embodiment.

[0038] Step 120 , performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field.

[0039] Optionally, in order to further reduce the number of model parameters and complexity, the current dictionary library can be feature screened to select key flow field features that are highly correlated with the reconstruction of sub-grid turbulent viscosity from the multi-dimensional flow field features in the current time step, such as filtering the flow field velocity filter values ​​in different directions obtained by screening. and , and the flow field velocity filtering values ​​in different directions and The first-order derivatives in different directions are used as key flow field features to eliminate a large number of redundant features and reduce the complexity of the model.

[0040] The feature screening step here includes: establishing an association relationship between a sample dictionary library of a sample turbulent field and a sub-grid turbulent viscosity label of the sample turbulent field through various machine learning algorithms, such as a sparse Bayesian learning algorithm, so as to select, from the sample dictionary library, sample key flow field features that are highly correlated with the sub-grid turbulent viscosity label based on the association relationship; then, screening flow field features that match the sample key flow field features in the current dictionary library as key flow field features of the target turbulent field.

[0041] Step 130, input the key flow field characteristics into the mechanism phenomenological model, and reconstruct the sub-grid turbulent viscosity of the target turbulent field in the current time step; wherein the mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of the sample turbulent field and the sub-grid turbulent viscosity label of the sample turbulent field.

[0042] Optionally, after obtaining the key flow field characteristics, the key flow field characteristics can be input into a mechanism phenomenological model pre-trained based on the initial neural network model, so as to perform flow field simulation and solution reasoning by the mechanism phenomenological model, and reconstruct the sub-grid turbulent viscosity of the output target turbulent field in the current time step, thereby realizing AI-enabled CFD to accelerate and guide large eddy simulation of complex flow fields.

[0043] The mechanism phenomenological model here is used to reconstruct the sub-grid turbulent viscosity based on the key flow field characteristics. It is pre-trained by the initial neural network model. The specific training steps include: Firstly, direct numerical simulation data of sample turbulence fields are collected and filtered to construct a data set for reconstructing LES turbulence equations. Based on the physical characteristics of fluid data and integrating basic conservation law constraints, the direct numerical simulation data are filtered to construct a complete function library and obtain a sample dictionary library of sample turbulence fields. Feature screening is performed on the sample dictionary library to obtain sample key flow field characteristics of sample turbulence fields. Based on the sample key flow field characteristics of sample turbulence fields and the sub-grid turbulence viscosity labels of sample turbulence fields, a sample data set is constructed.

[0044] In addition, an initial neural network model is constructed; the initial neural network model here can be constructed based on one or more neural network models; for example, the initial neural network model can be based on a shallow fully connected neural network model, and the shallow fully connected neural network model can be constructed by an input layer, three hidden layers and an output layer; wherein each hidden layer can contain 20 neurons, the input layer contains 6 neurons, and the output layer contains 1 neuron.

[0045] In addition, configure the hyperparameters of model training, including but not limited to the optimizer, loss function, number of iterations, learning rate and other hyperparameters of network training; the configuration of hyperparameters can be configured according to actual training requirements, such as setting the optimizer of network training to Adaptive Moment Estimation (Adam) optimizer, setting the number of iterations to 1000, and setting the learning rate to 0.0001.

[0046] Next, the sample key flow field features of the sample turbulent field in the sample data set are used as samples, and the sub-grid turbulent viscosity labels of the sample turbulent field are used as labels to iteratively train the initial neural network model to model the sub-grid turbulent viscosity in the large eddy simulation equation, capture the relationship between the features, and fit the key flow field feature input and sub-grid turbulent viscosity output. The accurate mapping relationship is obtained The mechanism-phenomenological model is used to optimize and reconstruct the turbulence subgrid model.

[0047] The method provided in this embodiment obtains a turbulent field dictionary library of the current time step containing multi-dimensional flow field features, extracts key flow field features through feature screening, and then inputs these features into a mechanism phenomenological model obtained by adaptive training based on direct numerical simulation data and sub-grid turbulent viscosity labels, so as to more accurately capture the turbulent details through the mechanism phenomenological model, thereby adaptively and accurately reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step, thereby improving the simulation accuracy of small-scale turbulent eddies in large eddy simulations, and effectively solving the problems of poor underlying model performance and low simulation calculation accuracy caused by computational grid scale limitations and sub-grid stress simplification assumptions.

[0048] In some embodiments, the mechanism phenomenological model is trained based on the following steps: According to the Navier-Stokes equation, the sample turbulence field is simulated and calculated to obtain direct numerical simulation data of the sample turbulence field; Filtering the direct numerical simulation data, and constructing a sample dictionary library of the sample turbulent field according to the filtered direct numerical simulation data; Performing feature screening on the sample dictionary library to obtain key flow field features of the sample turbulent field; Performing sub-grid turbulent viscosity labeling on the sample turbulent field to obtain the sub-grid turbulent viscosity label; Constructing a sample data set according to the key flow field characteristics of the sample turbulent field and the sub-grid turbulent viscosity label; According to the sample data set, obtaining the loss function of the initial neural network model; According to the loss function, the initial neural network model is iteratively trained to obtain the mechanism phenomenological model.

[0049] Figure 2 2 is a flow chart of the large eddy simulation method based on the mechanism phenomenological model provided by the present invention; Figure 2 As shown, the mechanism phenomenological model training includes data acquisition step, data processing step, and sub-grid modeling step.

[0050] Step 210, data acquisition; specifically, the turbulence field (flow around a two-dimensional square cylinder) in a typical incompressible flow turbulence scenario in fluid mechanics is selected as a sample turbulence field, and direct numerical simulation (DNS) spatiotemporal data of the sample turbulence field is collected to obtain high-precision original data of corresponding grid coordinates and physical field information, that is, direct numerical simulation data of the sample turbulence field.

[0051] The steps for obtaining direct numerical simulation data are as follows: In the two-dimensional square cylinder flow scene in the pipeline, the DNS model is used to solve the flow field. First, the grid is divided to obtain the grid coordinates of the sample turbulent field; for example, Figure 3 As shown, the X direction can be divided into 2016 grids and the Y direction can be divided into 900 grids, so the number of grids in the sample turbulent field is 1814400.

[0052] In addition, the operating data of the sample turbulent field is collected; the operating data includes but is not limited to the density of the fluid. , characteristic velocity of the fluid , characteristic length of the fluid , the dynamic viscosity of the fluid and the Reynolds number ; The calculation formula of Reynolds number is: ; in, represents the density of the fluid, represents the characteristic velocity of the fluid, represents a characteristic length, such as the width or height of a square column, Represents the dynamic viscosity of the fluid.

[0053] Secondly, in the Naviix software, the operating data is configured to form the corresponding sample operating conditions. For example, the calculation domain size is set to 3m×6.72m, the side length of the square column is 0.2m, the grid scale is (0.01 / 3)m, the pipe inlet velocity is 1m / s, the fluid density is 1, the viscosity is 6.6667e-5, and the time step is 0.001s. That is, in the Naviix software, the following settings are made: , , , , to obtain the corresponding working condition data. Therefore, substituting into the calculation formula of Reynolds number, in this two-dimensional square cylinder flow scenario, the Reynolds number The value of is 3000. Naviix will calculate from 0 seconds to 20 seconds to ensure stable turbulence, output a flow field result every 0.4 seconds, the total calculation time is 48 seconds, and a total of 71 time points are output.

[0054] Finally, based on the Navier-Stokes equations, simulation calculations were performed using the Naviix software to obtain the direct numerical simulation (DNS) data of the flow field at 71 time points in the two-dimensional flow around a square cylinder, including but not limited to grid coordinates, flow field velocity values ​​and flow field pressure values ​​in each direction, and the derivative terms of each order corresponding to the flow field velocity values ​​and flow field pressure values ​​in each direction.

[0055] Among them, for DNS direct numerical simulation, in two-dimensional incompressible flow, since the density is a constant, the calculation formula of the NS (Navier-Stokes) equation is as follows: ; ; in, is the Navier-Stokes continuity equation, which is used to describe the incompressibility or mass conservation of fluid flow, that is, the volume of the fluid remains unchanged during the flow; for Direction (i.e. The flow field velocity value on the axis); for The first derivative of the direction. is the Navier-Stokes momentum equation, for Flow field velocity value in direction versus time The first derivative of , that is, the rate of change of flow velocity with time, It is the nonlinear convection term of the flow field velocity value, which is used to represent the change of flow velocity caused by the movement (convection) of fluid micro-element. for Direction (i.e. The velocity value of the flow field on the axis, for The first derivative of the direction. is the flow field pressure value, is the pressure gradient term, which represents the change in flow velocity due to the change in pressure in space; is the sticky term, where is the dynamic viscosity coefficient of the fluid, For The coordinates in the direction are taken as the second partial differential (i.e., the second-order derivative). The viscosity term describes the change in flow velocity due to the internal viscosity force (internal friction) of the fluid, which causes shear stress and energy dissipation in the fluid during flow.

[0056] Step 220, data processing, specifically, filtering the direct numerical simulation data to construct a data set for LES turbulence equation reconstruction, providing a data basis for equation optimization. And around the physical characteristics of the flow data of the filtered direct numerical simulation data, a complete function library is constructed by integrating the basic conservation law constraints; then, a sample dictionary library is constructed using the physical characteristics of the filtered direct numerical simulation data and its complete functions.

[0057] The filtering algorithm library here may include filtering algorithms based on convolutional neural networks, uniform filtering, Gaussian filtering, etc. with various filtering scales, and this embodiment does not specifically limit this.

[0058] The filtering here may be implemented by randomly selecting a filtering algorithm from a plurality of filtering algorithms, or may be implemented by evaluating the performance of a plurality of filtering algorithms to select a filtering algorithm with the best performance.

[0059] Exemplarily, the sample dictionary library in step 220 may be implemented based on the following steps: In a filtering algorithm library, a target filtering algorithm is determined; the filtering algorithm library includes a plurality of filtering algorithms with different filtering scales and different filtering forms; filtering the direct numerical simulation data according to the filtering scale and filtering form corresponding to the target filtering algorithm; The filtered direct numerical simulation data are operated according to a large eddy simulation control model, a Smagorinsky model, and an equivalent equation of a subgrid stress tensor in the large eddy simulation control model; According to the operation result, the sample dictionary library is constructed.

[0060] Optionally, the construction of the sample dictionary library in step 220 specifically includes steps 221 , 222 and 223 .

[0061] Step 221 : Select a filtering algorithm with the best performance from a filtering algorithm library including a plurality of filtering algorithms with different filtering scales and different filtering forms as a target filtering algorithm.

[0062] Here, the screening of the filtering algorithms may be achieved by performing a performance test and comparison on each filtering algorithm; or may be achieved by performing a scene matching on each filtering algorithm, which is not specifically limited in this embodiment.

[0063] Exemplarily, the filtering algorithm screening in step 221 can be implemented based on the following steps: According to the simulation accuracy, computing resource consumption and filtered data distribution characteristics of each filtering algorithm in the filtering algorithm library, the performance score corresponding to each filtering algorithm is obtained; In the filtering algorithm library, a filtering algorithm with the highest performance score is determined as the target filtering algorithm.

[0064] Optionally, each filtering algorithm in the filtering algorithm library is tested to determine the simulation accuracy of the sub-grid turbulent viscosity based on the DNS data filtered by each filtering algorithm; in addition, the computing resource consumption required for each filtering algorithm to filter the DNS data is determined; in addition, the data distribution characteristics of the direct numerical simulation data after filtering by each filtering algorithm are obtained.

[0065] Subsequently, the performance scores of each filtering algorithm are calculated by combining the simulation accuracy, computing resource consumption and data distribution characteristics of the filtered direct numerical simulation data corresponding to each filtering algorithm, so as to determine the filtering algorithm with the highest performance score in the filtering algorithm library as the target filtering algorithm.

[0066] The performance score here can be calculated for each corresponding score for simulation accuracy, computing resource consumption and data distribution characteristics, and finally, the scores are summarized to obtain the performance score corresponding to each filtering algorithm; or the simulation accuracy, computing resource consumption and data distribution characteristics can be input into a pre-trained scoring model, so that the scoring model can predict and output the performance score corresponding to each filtering algorithm. This embodiment does not make specific limitations on this.

[0067] The method provided in the embodiment of the present application carefully evaluates the performance of each filtering algorithm in the filtering algorithm library in terms of simulation accuracy, computing resource consumption, and data distribution characteristics after filtering, and determines the filtering algorithm with the highest performance score as the target filtering algorithm. This method can significantly improve the accuracy and efficiency of large eddy simulation, thereby ensuring the reliability of simulation results while optimizing the use of computing resources to achieve efficient and economical turbulent field simulation.

[0068] Step 222, filtering the direct numerical simulation data according to the filtering scale and filtering form corresponding to the selected target filtering algorithm, removing small-scale turbulent structures from the fluid characteristics, so as to perform large eddy simulation on a large scale.

[0069] Exemplarily, when the target filtering algorithm is filtering based on a convolutional neural network, filtering based on a convolutional neural network can be performed on the grid flow field data (or direct numerical simulation data) at each time point, and the specific steps include: The cloud map information of DNS data is analyzed, and a complete rectangular sub-area containing turbulence is selected behind the square column for filtering, and the filtering range and filtering scale are set, such as setting the rectangular range interval in the X direction to [3,6], the rectangular range interval in the Y direction to [0.5, 2.5], and the filtering scale to 4, 8, 16, and 32 respectively. By adjusting the convolution kernel and step size of the neural network, efficient parallel filtering is achieved, and the precise coordinates corresponding to several grid points at different times, the filtered flow field velocity values ​​and filtered flow field pressure values ​​in each direction, and their corresponding derivatives of each order are obtained.

[0070] Step 223, according to the large eddy simulation control model, the Smagorinsky model, and the equivalent equation of the sub-grid stress tensor in the large eddy simulation control model, the direct numerical simulation data filtered in step 222 is operated to construct a sample dictionary library of the sample turbulent field.

[0071] Optionally, the filtered flow field velocity values ​​in each direction are calculated to obtain , The above data are important components of the classic LES subgrid model and serve as sample data sets for reconstruction of the LES equation.

[0072] Among them, the theoretical LES control model after filtering (that is, the large eddy simulation control model) is similar to the NS equation. The physical quantities in the equation can represent the mean values ​​(or filtered values) of different regions: ; ; in, for Direction (i.e. Filtered value of flow field velocity on the axis); for Direction (i.e. The velocity filter value of the flow field on the axis, is the flow field pressure filtering value, is the subgrid stress tensor, which is the result of the interaction between the nonlinear term of the filtered velocity field and the velocity field itself; , which is used to quantify the contribution of small-scale turbulence. Among them, for The velocity value of the flow field in the direction The filtered value after combining the flow field velocity values ​​in the direction; for The flow field velocity filtering value in the direction The combined value of the flow field velocity filter values ​​in the direction.

[0073] Subgrid stress tensor in the theoretical LES equation Need to know the real physical quantity This makes the theoretical equation unsolvable. In order to make the equation closed and solvable, it is necessary to use the equivalent equation of the subgrid stress tensor in the large eddy simulation control model, and calculate the calculable variables. Make an approximation and replace.

[0074] Exemplarily, the specific calculation formula of the equivalent equation is as follows: ; ; in, is the sub-grid stress tensor; is the filtered strain rate tensor; is the subgrid turbulent viscosity; for About space coordinates The first derivative of ; for About space coordinates The first derivative of .

[0075] In the above equivalent equation, The calculation depends on the deformation rate of the filtered velocity field ,and is defined by the spatial partial derivatives of the velocity field. In this way, the subgrid model can provide an approximate subgrid stress tensor, which allows the LES equations to be closed and solved numerically.

[0076] In addition, the Smagorinsky model can further explain the sub-grid turbulent viscosity. To model: ; ; in, is the subgrid turbulent viscosity; is the fluid density; It is a constant term in the Smagorinsky model, and can be taken as 0.1; is the filter length, which is usually related to the selected grid size for filtering; is the filtered strain rate tensor, is the invariant of the strain rate tensor.

[0077] It is usually assumed Need to be determined dynamically, so the As the dynamic eddy viscosity coefficient, it can be expressed as , so the calculation formula of the Smagorinsky-Lilly model can be modified as follows: .

[0078] In summary, according to the large eddy simulation control model, the Smagorinsky model, and the equivalent equation of the sub-grid stress tensor in the large eddy simulation control model, additional nonlinear combinations can be introduced into the filtered direct numerical simulation data, thereby integrating the basic conservation law constraints of the NS equation to build a complete function library. Through the addition, subtraction, multiplication and division combination operations between the function libraries, the sample dictionary library is constructed using physical quantities and their complete functions, that is, .

[0079] The method provided in the embodiment of the present application pre-processes the direct numerical simulation data by carefully selecting an algorithm with a specific filtering scale and form in the filtering algorithm library, and then uses the large eddy simulation control model and the Smagorinsky model, combined with the equivalent equation of the sub-grid stress tensor, to conduct in-depth analysis and calculation of the filtered data. Finally, an accurate and efficient sample dictionary library that comprehensively reflects the flow field characteristics can be constructed, which provides reliable data support and technical foundation for large eddy simulation, thereby significantly improving the accuracy and computational efficiency of the simulation.

[0080] Step 230 , sub-grid modeling, specifically includes step 231 , step 232 , step 233 and step 234 .

[0081] Step 231, feature screening is performed on the sample dictionary library to remove weakly related items in the dictionary library, and to identify and screen the items related to the sub-grid turbulent viscosity. Key flow characteristics of the associated sample turbulence field.

[0082] The feature screening here can be implemented based on various screening algorithms, such as a sparse Bayesian learning algorithm, a principal component analysis algorithm, etc., which is not specifically limited in this embodiment.

[0083] Exemplarily, the key flow field features of the sample turbulent field can be obtained based on the following steps: Based on a sparse Bayesian learning model, simulating the nonlinear relationship between the multidimensional flow field characteristics in the sample dictionary library and the sub-grid turbulent viscosity label; Determining, according to the nonlinear relationship, the correlation parameters between the flow field characteristics of each dimension in the sample dictionary library and the sub-grid turbulent viscosity label; According to the association parameters, the multi-dimensional flow field features in the sample dictionary library are screened to obtain the key flow field features of the sample turbulent field.

[0084] Among them, the sparse Bayesian learning algorithm is used to screen feature items. By introducing sparsity priors to balance the numerical fitting ability and model complexity, the posterior distribution of parameters is updated according to the Bayesian theorem, thereby realizing automatic screening of features, eliminating weakly related items in the function library, and identifying and screening sub-grid turbulent viscosity. Key flow characteristics of the associated sample turbulence field.

[0085] Among them, the calculation formula for the nonlinear relationship between the multidimensional flow field characteristics in the sample dictionary library formed by the simulation of the sparse Bayesian learning algorithm and the sub-grid turbulent viscosity label is as follows: ; in, It is the label of sub-grid turbulent viscosity, which can be obtained by calculating DNS simulation data; is a sample dictionary library, which contains multi-dimensional flow field features; is the model parameter, also known as the association parameter between the flow field characteristics of each dimension in the sample dictionary library and the sub-grid turbulent viscosity label, which is used for key feature selection; is unknown random noise.

[0086] Figure 4 The schematic diagram of the key feature screening based on the sparse Bayesian learning algorithm provided by the present invention. Figure 4 As shown in the figure, after the modeling is completed based on the sparse Bayesian learning algorithm, most of the model parameters will tend to 0, and accordingly, the sparsity of the model parameters can be controlled. For example, the dictionary function (or flow field feature) corresponding to the model parameter being 0 will be eliminated and will no longer be used for the subsequent sub-grid turbulent viscosity modeling, and only the model parameter The dictionary function corresponding to values ​​greater than 0 is used as the final key flow field feature to participate in the subsequent sub-grid turbulent viscosity modeling. This feature screening method can be used to identify and screen key features, eliminate a large number of redundant features, and reduce the complexity of the model. For example, the key simple features finally extracted are reduced to 6 items: .

[0087] The method provided in the embodiment of the present application, by adopting a sparse Bayesian learning algorithm, realizes the accurate modeling of the nonlinear relationship between the multidimensional flow field characteristics in the sample turbulent field and the sub-grid turbulent viscosity label, and automatically screens out the key flow field characteristics that are closely related to the sub-grid turbulent viscosity. This method not only improves the prediction accuracy of the model, but also effectively reduces the complexity of the model by eliminating redundant features, thereby providing a highly efficient and accurate feature selection mechanism for large eddy simulation.

[0088] Step 232 , labeling the sample turbulent field with sub-grid turbulent viscosity to obtain a sub-grid turbulent viscosity label.

[0089] Optionally, the sub-grid model equations satisfied by the flow around the two-dimensional square cylinder are substituted into the momentum equations in the X and Y directions, labeled with the sub-grid turbulent viscosity; wherein the calculation formulas of the momentum equations in the X and Y directions are as follows: ; ; ; ; ; ; in, , , , are calculated by the above formulas, and , , , DNS data can also be obtained by Formula calculation.

[0090] Step 233, constructing a sample data set based on the key flow field features of the sample turbulent field and the sub-grid turbulent viscosity labels.

[0091] Step 234, based on the sample data set constructed in step 233, model the sub-grid turbulent viscosity in the large eddy simulation equation to capture the relationship between the characteristics and fit the key physical quantity input and subgrid turbulent viscosity output The accurate mapping relationship is obtained The mechanism-phenomenological model is used to optimize and reconstruct the turbulence subgrid model.

[0092] Optionally, the key flow field features of the sample turbulent field in the sample data set are input into the initial neural network model, and the sample sub-grid turbulent viscosity of the output sample turbulent field is reconstructed by the initial neural network model to obtain a loss function based on the deviation between the reconstructed sample sub-grid turbulent viscosity and the sub-grid turbulent viscosity label.

[0093] Exemplarily, the loss function is calculated as follows: ; in, is the loss function; is the reconstructed sample subgrid turbulent viscosity; is the square function of the L2 norm; is the number of equation constraints in the calculation formula of the momentum equation in the X and Y directions, that is, 4. The loss function is obtained by calculating the deviation between the sub-grid turbulent viscosity label calculated according to the above four equation constraints and the reconstructed sample sub-grid turbulent viscosity, such as the mean square error (MSE). Based on the loss function, the initial neural network model is iteratively trained to obtain The mechanism-phenomenological model is used to optimize and reconstruct the turbulence subgrid model.

[0094] In summary, the method provided in this embodiment, through data preprocessing, key physical quantity screening, network structure design, model training and model reasoning, and through the integration of sparse Bayesian learning and neural network architecture, establishes a simulation solution framework of AI modeling fusion CFD, constructs the mechanism-phenomenological model of the turbulent viscosity term in the LES equation, can realize the reconstruction and optimization of the LES sub-grid model, and then realizes the accurate prediction of the sub-grid turbulent viscosity in large eddy simulation, which significantly improves the computational efficiency and simulation accuracy of large eddy simulation, while reducing the complexity of the model, and can effectively adapt to the efficient numerical simulation of complex turbulent fields.

[0095] In some implementations, the method further comprises: Solving the sub-grid turbulent viscosity of the target turbulent field in the current time step to obtain the next dictionary library; Iteratively reconstructing the sub-grid turbulent viscosity in the next time step according to the next dictionary library until the reconstructed sub-grid turbulent viscosity converges; The flow field simulation is performed according to the dictionary library solved in each time step and the reconstructed sub-grid turbulent viscosity to obtain the simulated flow field of the target turbulent field.

[0096] Optionally, after obtaining the mechanism phenomenological model in step 230, a fusion model may be constructed. The reconstruction model (or mechanism-phenomenological model) and the simulation system of the Naviix solver communicate using file transfer. The Naviix solver iteratively calls the mechanism-phenomenological model in each time step and reads the new turbulent viscosity. The next iteration is performed until the calculation converges to solve the turbulent flow scene around the two-dimensional square cylinder and obtain the simulated flow field of the turbulent sub-grid reconstruction model. The steps for obtaining the simulated flow field specifically include: The mechanism phenomenological model formed by the modeling in step 230 is deployed, and a file reading and writing interface is developed as an external program. Figure 5 As shown, the Naviix solver outputs field variables and their partial derivatives at each time step. , and call an external program to read the above data and calculate the sub-grid turbulent viscosity of LES Calculations are performed, and then the external program outputs the space The solver reads the new And solve it to get the next dictionary library to iteratively solve the sub-grid turbulent viscosity in the next time step until the calculation converges.

[0097] By building a fusion Reconstruct the simulation system of the model and Naviix solver, use file transfer for communication, and Naviix solver iteratively calls the mechanism-phenomenological model in each time step to read the new turbulent viscosity The next iterative solution is performed until the calculation converges, and the turbulent flow scene around the two-dimensional square cylinder is solved to obtain the simulated flow field of the turbulent sub-grid reconstruction model.

[0098] The method provided in this embodiment, in terms of system integration, is based on file transfer for communication, does not require complex network programming or specific communication protocols, and achieves decoupling of the solver from the external program. In addition, by combining the Naviix simulation software to calculate and verify the reconstructed sub-grid model (or mechanism phenomenological model), a more general and accurate sub-grid model equation and simulation software system are obtained, thereby realizing the use of the mechanism-phenomenological model to optimize the traditional large eddy simulation sub-grid model for turbulence simulation solution, greatly enhancing the solution capability of turbulence simulation, and achieving a more refined and efficient turbulence simulation effect.

[0099] In addition, in order to verify the effectiveness of the method provided in the embodiment of the present application, the trained large eddy simulation equation reconstruction model (that is, the mechanism-phenomenological model) will be used to reconstruct the flow field under the same working conditions under the same parameter configuration conditions. By analyzing and comparing the physical field data simulated by different underlying models, the effectiveness of the method provided in the embodiment of the present application will be verified.

[0100] During the verification process, for the same two-dimensional square cylinder flow scene, by using the same grid division and boundary conditions, different simulation solvers and other underlying turbulence model simulations were used for comparative verification. Specifically, the Naviix solver was used to simulate the flow field of the sub-gridless model and the Smagorinsky-Lilly model, and the Fluent solver was used to simulate the flow field of the Smagorinsky-Lilly model. By comparing the dynamic cloud map of the flow field solution results of each turbulence model and the time-averaged value of the key coordinates of the flow field, the feasibility verification of the LES reconstruction of the sub-grid model was completed. Among them, the key coordinates of the flow field selected 10 points along the X direction of the midline of the calculation domain, which are locations where the turbulence phenomenon is more complicated. The position distribution is as follows Figure 6 shown.

[0101] Through verification, it can be seen that based on the mechanism-phenomenological model and other underlying equations of large eddy simulation, the dynamic cloud diagram of different simulation results under the same scene shows that the turbulence phenomenon is more obvious in Naviix than in Fluent. The range of longitudinal velocity and pressure of the method provided in the embodiment of the present application (or Naviix+mechanism-phenomenological model, or NaviixLes) is closer to the calculation results of Naviix+SL and Fluent.

[0102] Figure 7The line graph shows the horizontal velocity changes of the key coordinates of the flow field in the 28-second range in the simulation results of different models used. Figure 7 It can be seen that in terms of the speed fluctuation amplitude in the X direction, the NaviixLes provided in the embodiment of the present application is relatively similar to the DNS filtering, and the fluctuation amplitude of the Fluent calculation result is relatively small.

[0103] Figure 8 , Fig. 9 and Fig.10 The following table shows the distribution of the time-averaged values ​​of the 10 key coordinates of the flow field in the simulation results of different models used, namely the fluid velocity value in the X-axis direction (or X-Velocity), the fluid velocity value in the Y-axis direction (or Y-Velocity), and the flow field pressure value. Figure 8-10 It can be seen that the physical quantities of NaviixLes provided in the embodiment of the present application are closer to the DNS filtering in terms of the time statistical mean compared with Naviix+SL (or Naviix).

[0104] It can be seen from the above that by comparing with the original model method and the simulated flow field, the effectiveness of the method provided in the embodiment of the present application is effectively verified.

[0105] The large eddy simulation system based on the mechanism phenomenological model provided by the present invention is described below. The large eddy simulation system based on the mechanism phenomenological model described below and the large eddy simulation method based on the mechanism phenomenological model described above can be referenced to each other.

[0106] Fig.11 Schematic diagram of the structure of the large eddy simulation system based on the mechanism phenomenological model provided by the present invention; Fig.11 As shown, the system includes: The acquisition unit 1110 is used to acquire the current dictionary library of the target turbulent flow field; the current dictionary library includes multi-dimensional flow field features within the current time step; The selection unit 1120 is used to perform feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; The reconstruction unit 1130 is used to input the key flow field characteristics into the mechanism phenomenological model, and reconstruct the sub-grid turbulent viscosity of the target turbulent field in the current time step; The mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and sub-grid turbulent viscosity labels of the sample turbulent field.

[0107] The system provided in the embodiment of the present application obtains a turbulent field dictionary library of the current time step containing multi-dimensional flow field characteristics, extracts key flow field characteristics through feature screening, and then inputs these characteristics into a mechanism phenomenological model obtained through adaptive training based on direct numerical simulation data and sub-grid turbulent viscosity labels, so as to more accurately capture the turbulent details through the mechanism phenomenological model, thereby adaptively and accurately reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step, thereby improving the simulation accuracy of small-scale turbulent eddies in large eddy simulations, and effectively solving the problems of poor underlying model performance and low simulation calculation accuracy caused by computational grid scale limitations and sub-grid stress simplification assumptions.

[0108] Fig.12 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.12 As shown, the electronic device may include: a processor 1210, a communication interface 1220, a memory 1230 and a communication bus 1240, wherein the processor 1210, the communication interface 1220 and the memory 1230 communicate with each other through the communication bus 1240. The processor 1210 may call the logic instructions in the memory 1230 to execute the large eddy simulation method based on the mechanism phenomenological model provided by the above methods, the method comprising: obtaining the current dictionary library of the target turbulent field; the current dictionary library includes the multi-dimensional flow field features within the current time step; feature screening is performed on the current dictionary library to obtain the key flow field features of the target turbulent field; the key flow field features are input into the mechanism phenomenological model, and the sub-grid turbulent viscosity of the target turbulent field within the current time step is reconstructed; wherein the mechanism phenomenological model is obtained by training the initial neural network model based on the direct numerical simulation data of the sample turbulent field and the sub-grid turbulent viscosity label of the sample turbulent field.

[0109] In addition, the logic instructions in the above-mentioned memory 1230 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0110] On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the large eddy simulation method based on the mechanism phenomenological model provided by the above-mentioned methods, and the method includes: obtaining a current dictionary library of the target turbulent field; the current dictionary library includes multidimensional flow field characteristics within the current time step; performing feature screening on the current dictionary library to obtain key flow field characteristics of the target turbulent field; inputting the key flow field characteristics into the mechanism phenomenological model, and reconstructing the sub-grid turbulent viscosity of the target turbulent field within the current time step; wherein the mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and the sub-grid turbulent viscosity label of the sample turbulent field.

[0111] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the large eddy simulation method based on the mechanism phenomenological model provided by the above-mentioned methods, the method comprising: obtaining a current dictionary library of the target turbulent field; the current dictionary library comprises multidimensional flow field features within a current time step; performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; inputting the key flow field features into the mechanism phenomenological model to reconstruct the sub-grid turbulent viscosity of the target turbulent field within the current time step; wherein the mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and the sub-grid turbulent viscosity label of the sample turbulent field.

[0112] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0113] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A large eddy simulation method based on a mechanism phenomenological model, characterized in that: include: Acquire a current dictionary library of the target turbulent flow field; the current dictionary library includes multi-dimensional flow field characteristics within the current time step; Performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; Inputting the key flow field characteristics into the mechanism phenomenological model, and reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step; The mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and sub-grid turbulent viscosity labels of the sample turbulent field.

2. The large eddy simulation method based on the mechanism phenomenological model according to claim 1, characterized in that: The mechanism phenomenological model is trained based on the following steps: According to the Navier-Stokes equation, the sample turbulence field is simulated and calculated to obtain direct numerical simulation data of the sample turbulence field; Filtering the direct numerical simulation data, and constructing a sample dictionary library of the sample turbulent field according to the filtered direct numerical simulation data; Performing feature screening on the sample dictionary library to obtain key flow field features of the sample turbulent field; Performing sub-grid turbulent viscosity labeling on the sample turbulent field to obtain the sub-grid turbulent viscosity label; Constructing a sample data set according to the key flow field characteristics of the sample turbulent field and the sub-grid turbulent viscosity label; According to the sample data set, obtaining the loss function of the initial neural network model; According to the loss function, the initial neural network model is iteratively trained to obtain the mechanism phenomenological model.

3. The large eddy simulation method based on the mechanism phenomenological model according to claim 2, characterized in that: The filtering of the direct numerical simulation data and constructing a sample dictionary library of the sample turbulent field according to the filtered direct numerical simulation data include: In a filtering algorithm library, a target filtering algorithm is determined; the filtering algorithm library includes a plurality of filtering algorithms with different filtering scales and different filtering forms; filtering the direct numerical simulation data according to the filtering scale and filtering form corresponding to the target filtering algorithm; The filtered direct numerical simulation data are operated according to a large eddy simulation control model, a Smagorinsky model, and an equivalent equation of a subgrid stress tensor in the large eddy simulation control model; According to the operation result, the sample dictionary library is constructed.

4. The large eddy simulation method based on the mechanism phenomenological model according to claim 3, characterized in that: Determining the target filtering algorithm in the filtering algorithm library includes: According to the simulation accuracy, computing resource consumption and filtered data distribution characteristics of each filtering algorithm in the filtering algorithm library, the performance score corresponding to each filtering algorithm is obtained; In the filtering algorithm library, a filtering algorithm with the highest performance score is determined as the target filtering algorithm.

5. The large eddy simulation method based on the mechanism phenomenological model according to claim 2, characterized in that: The feature screening of the sample dictionary library to obtain key flow field features of the sample turbulent field includes: Based on a sparse Bayesian learning model, simulating the nonlinear relationship between the multidimensional flow field characteristics in the sample dictionary library and the sub-grid turbulent viscosity label; Determining, according to the nonlinear relationship, the correlation parameters between the flow field characteristics of each dimension in the sample dictionary library and the sub-grid turbulent viscosity label; According to the association parameters, the multi-dimensional flow field features in the sample dictionary library are screened to obtain the key flow field features of the sample turbulent field.

6. The large eddy simulation method based on the mechanism phenomenological model according to any one of claims 1 to 5, characterized in that: The method further comprises: Solving the sub-grid turbulent viscosity of the target turbulent field in the current time step to obtain the next dictionary library; Iteratively reconstructing the sub-grid turbulent viscosity in the next time step according to the next dictionary library until the reconstructed sub-grid turbulent viscosity converges; The flow field simulation is performed according to the dictionary library solved in each time step and the reconstructed sub-grid turbulent viscosity to obtain the simulated flow field of the target turbulent field.

7. A large eddy simulation system based on a mechanism phenomenological model, characterized in that: include: An acquisition unit, used to acquire a current dictionary library of a target turbulent flow field; the current dictionary library includes multi-dimensional flow field characteristics within a current time step; A selection unit, used for performing feature screening on the current dictionary library to obtain key flow field features of the target turbulent field; A reconstruction unit, used for inputting the key flow field characteristics into the mechanism phenomenological model, and reconstructing the sub-grid turbulent viscosity of the target turbulent field in the current time step; The mechanism phenomenological model is obtained by training an initial neural network model based on direct numerical simulation data of a sample turbulent field and sub-grid turbulent viscosity labels of the sample turbulent field.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the large eddy simulation method based on the mechanism phenomenological model as claimed in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the large eddy simulation method based on the mechanism phenomenological model as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the large eddy simulation method based on the mechanism phenomenological model as claimed in any one of claims 1 to 6 is implemented.

Images