CFD automation method for optimal airfoil flow analysis based on intensive training, CFD airfoil flow analysis method and CFD airfoil flow analysis equipment

Abstract

Disclosed is a computational fluid dynamics (CFD) flow analysis method for a blade, the method comprising: generating an analysis automation model that predicts CFD flow analysis results from input analysis conditions, the generating the analysis automation model comprising: inputting a blade shape and flow conditions, determining analysis settings for flow analysis on the blade, and generating an analysis automation model that predicts CFD flow analysis results from the input analysis conditions; performing CFD simulation based on the determined analysis setup parameters to perform flow analysis on the blade, evaluating the result of the performed flow analysis on the blade, and training by reinforcement learning using an artificial neural network to meet predetermined evaluation criteria, inputting any blade shape condition and flow condition, and outputting the result of the flow analysis on the blade. And determining analysis setting parameters by applying the analysis automation model to the input blade shape and flow conditions, predicting a CFD simulation, and outputting the predicted CFD simulation as a result of the flow analysis.

Description

Technical Field

[0001] This disclosure relates to an automated computational fluid dynamics (CFD) method for optimal flow analysis on a blade based on reinforcement learning, a method for analyzing flow around a blade using CFD, and an apparatus for analyzing flow around a blade using CFD. Background Technology

[0002] Flow analysis around blades plays a crucial role in the design of various fluid machinery, such as aircraft wings, ship propellers, aircraft engines, gas turbines, steam turbines, and compressors. This flow analysis is primarily performed through computational fluid dynamics (CFD) and is considered an important step in design optimization. Accurate and efficient flow analysis using CFD requires specialized setup, including generating a computational mesh suitable for the flow conditions and geometry, selecting a turbulence model, and configuring appropriate numerical methods.

[0003] However, the flow around the blades is highly complex, making it difficult to determine the ideal analytical setup a priori. Therefore, appropriate setups must be found through iterative simulations, a time-consuming and expensive process. Furthermore, setups considered suitable for specific conditions may become less effective when flow conditions or geometry change, necessitating repetition of the entire process. Additionally, analytical results can vary significantly based on the engineer's subjective judgment, negatively impacting the reliability of the calculations.

[0004] Therefore, a new method is needed to improve the accuracy and efficiency of flow analysis on blades, while reducing time and cost. Consequently, there is an urgent need for a technical solution to address the various problems encountered in flow analysis and consistently obtain reliable results.

[0005] Public content Technical issues This disclosure attempts to provide an automated computational fluid dynamics (CFD) method for optimal flow analysis on blades based on reinforcement learning, which can cost-effectively obtain highly accurate analysis results by optimizing and automating the entire process of flow analysis around the blade.

[0006] Furthermore, the embodiments aim to provide an automated flow analysis method for blades and a blade flow analysis device that uses CFD through reinforcement learning.

[0007] However, the problems that the embodiments of this disclosure are intended to solve are not limited to those described above, and can be extended in various ways within the scope of the technical concepts included in this disclosure.

[0008] Technical solutions According to an embodiment, a computational fluid dynamics (CFD) flow analysis method for blades includes: generating an automated analysis model that predicts CFD flow analysis results based on input analysis conditions; inputting arbitrary blade shape conditions and flow conditions; determining analysis setup parameters by applying the automated analysis model to the input blade shape conditions and flow conditions; predicting a CFD simulation; and outputting the predicted CFD simulation as a flow analysis result. Generating the automated analysis model includes: inputting blade shape and flow conditions; determining analysis settings for flow analysis on the blade; performing flow analysis on the blade by conducting CFD simulations based on the determined analysis setup parameters; evaluating the results of the performed flow analysis on the blade; and training the model using reinforcement learning of an artificial neural network to meet predetermined evaluation criteria.

[0009] The generation of automated analysis models can include: a generation of analysis setup models that determine analysis setup parameters based on input analysis conditions; and a generation of analysis performance models that predict CFD flow analysis results. The generation of analysis setup models can include: selecting analysis setups for flow analysis on the blade based on input blade shape and flow conditions, and training the analysis setup model using a deep reinforcement learning algorithm leveraging artificial neural networks to ensure the selected analysis setup meets predetermined quality criteria. The generation of analysis performance models can include: performing flow analysis on the blade using CFD simulations, evaluating the results of the performed flow analysis on the blade, and training the analysis performance model using a deep reinforcement learning algorithm leveraging artificial neural networks to meet the evaluation criteria.

[0010] Analysis settings for flow analysis on blades can include generating a computational grid, selecting a turbulence model, and selecting a numerical method.

[0011] The blade shape can be represented by the blade shape parameter BSP and the blade pitch, and the flow conditions can be represented by the isentropic Reynolds number Re at the outlet. is,out and Mach number Ma is,out and the inflow angle θ in Expressed in a dimensionless manner.

[0012] The generated analysis setup model may include defining states s according to the following equation: , Action a is defined according to the following equation: ,as well as The model is generated using gridded parameters obtained by determining actions based on a given state in each episode, the quality of the generated grid is evaluated and a reward is calculated, and the analysis setup model is trained using a single-step actor-critic algorithm.

[0013] Here, BSP represents the blade shape parameters, pitch represents the blade spacing, and N o β represents the total number of nodes in the O-type mesh. le β represents the degree of clustering of nodes at the leading edge. te The clustering degree of the nodes at the trailing edge is represented by Δn1, which represents the height of the first unit in the direction perpendicular to the blade surface. in Indicates the entrance location in the vertical direction, y out Indicates the exit location in the vertical direction, α 中弧线 The x represents the degree of curvature of the periodic boundary following the camber line. o in Indicates the horizontal position of the HO-type interface, x o out δ represents the position of the OH-type interface in the horizontal direction, and δ represents the ratio of the number of nodes in the tangential direction of the O-type mesh to the number of nodes in the normal direction relative to the blade surface.

[0014] The reward can be defined based on the mesh quality that minimizes the numerical error caused by geometric imperfections in the generated mesh, and the mesh quality can be evaluated by calculating the reward using a mesh quality metric that includes the ratio Q of the determinants of the Jacobian matrices. J and element skewness Q S .

[0015] Evaluation criteria used to assess the results of flow analyses performed on blades may include reward functions that evaluate computational accuracy and efficiency in CFD simulations.

[0016] Computational accuracy can be calculated by the error between high-quality data, including experimental data or high-resolution numerical data, and the results of the flow analysis performed on the blade. Computational efficiency can be calculated by measuring the total time required to perform the flow analysis on the blade.

[0017] Generating an analytical performance model may include defining states s according to the following equation: , Action a is defined according to the following equation: ,as well as The CFD simulation is completed by executing actions based on the given state in each round, calculating rewards from the CFD simulation results, and training and analyzing the performance model using the actor-critic algorithm.

[0018] Here, BSP represents the blade shape parameter, pitch represents the blade spacing, and Re... is,outMa represents the isentropic Reynolds number at the exit. is,out θ represents the Mach number at the exit. in Indicates the inflow angle, w represents the weight, and N represents the inflow angle. o β represents the total number of nodes in the O-type mesh. le β represents the degree of clustering of nodes at the leading edge. te Δn1 represents the clustering degree of the nodes at the trailing edge, and Δn1 represents the height of the first unit in the direction perpendicular to the blade surface.

[0019] Generative analysis performance models can include collecting learning data by using multi-agent reinforcement learning and performing CFD simulations in parallel with multiple agents.

[0020] The computational fluid dynamics (CFD) flow analysis apparatus on a blade according to an embodiment includes: an automated analysis model generator that generates an automated analysis model that predicts CFD flow analysis results based on input analysis conditions; and a flow analyzer on the blade that receives inputs of arbitrary blade shape conditions and flow conditions, applies the automated analysis model to the input blade shape conditions and flow conditions to determine analysis setup parameters, performs CFD simulations to predict the flow field, and outputs the flow field as the CFD flow analysis result. The automated analysis model generator can input blade shape and flow conditions, determine analysis settings for flow analysis on the blade, perform flow analysis on the blade by performing CFD simulations based on the determined analysis setup parameters, evaluate the results of the performed flow analysis on the blade, and train the automated analysis model to meet predetermined evaluation criteria using reinforcement learning of an artificial neural network.

[0021] The automated analysis model generator can include: an analysis setup model that determines analysis setup parameters based on input analysis conditions, and an analysis performance model that predicts CFD flow analysis results. The analysis setup model can select the analysis setup for flow analysis on the blade based on the input blade shape and flow conditions, and can be trained using a deep reinforcement learning algorithm employing artificial neural networks to ensure the selected analysis setup meets predetermined quality criteria. The analysis performance model can perform flow analysis on the blade using CFD simulation, evaluate the results of the performed flow analysis, and is trained using a deep reinforcement learning algorithm employing artificial neural networks to meet evaluation criteria.

[0022] An automated computational fluid dynamics (CFD) method for optimal flow analysis on a blade, according to an embodiment, includes: inputting analysis conditions including blade shape and flow conditions; determining analysis settings for flow analysis on the blade; performing flow analysis on the blade by performing CFD simulation based on the determined analysis settings parameters; evaluating the results of the performed flow analysis on the blade; and training the flow analysis on the blade using reinforcement learning of an artificial neural network so that the results of the flow analysis on the blade meet predetermined evaluation criteria.

[0023] CFD automation methods may also include: selecting an analysis setup for flow analysis on the blade based on the input blade shape and flow conditions, and training the selected analysis setup using a deep reinforcement learning algorithm that utilizes artificial neural networks to ensure that the selected analysis setup meets predetermined quality standards.

[0024] Analysis settings for flow analysis on blades can include generating a computational grid, selecting a turbulence model, and selecting a numerical method.

[0025] Evaluation criteria used to assess the results of flow analyses performed on blades may include reward functions that evaluate computational accuracy and efficiency in CFD simulations.

[0026] Computational accuracy can be calculated by the error between high-quality data, including experimental data or high-resolution numerical data, and the results of the flow analysis performed on the blade. Computational efficiency can be calculated by measuring the total time required to perform the flow analysis on the blade.

[0027] Generative analysis performance models can include collecting learning data by using multi-agent reinforcement learning and performing CFD simulations in parallel with multiple agents.

[0028] Reinforcement learning can include computation using deep reinforcement learning algorithms based on artificial neural network techniques, such as deep Q-networks, proximal policy optimization (PPO), or actor-commentator algorithms.

[0029] According to an embodiment, a computer device includes at least one processor configured to execute computer-readable instructions. The at least one processor may be configured to: generate an automated analysis model that predicts CFD flow analysis results based on input analysis conditions, taking arbitrary blade shape conditions and flow conditions as input, determining analysis setup parameters by applying the automated analysis model to the input blade shape and flow conditions, predicting a CFD simulation, and outputting the predicted CFD simulation as a flow analysis result. Generating the automated analysis model may include: taking blade shape and flow conditions as input, determining analysis settings for flow analysis on the blade, performing a CFD simulation based on the determined analysis setup parameters to perform flow analysis on the blade, evaluating the results of the performed flow analysis on the blade, and training the automated analysis model to meet predetermined evaluation criteria using reinforcement learning of an artificial neural network.

[0030] Beneficial effects According to some embodiments, a reinforcement learning-based CFD automation method for optimal analysis of flow around a blade can be used to automate the entire process of optimizing CFD analysis of flow on a blade.

[0031] This can improve the accuracy and efficiency of analysis, reduce time and cost, increase the reliability of calculation results, and achieve technological innovation in the field of fluid machinery design.

[0032] Furthermore, the automated CFD flow analysis method and CFD flow analysis equipment according to some embodiments can obtain optimal analysis results with improved analysis accuracy and efficiency, while saving time and costs.

[0033] Furthermore, by minimizing the involvement of engineers in flow analysis on blades, the barriers to entry for analysis can be lowered, and progress in related research can be facilitated. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating an automated computational fluid dynamics (CFD) method for optimal flow analysis around a blade based on reinforcement learning, according to an embodiment.

[0035] Figure 2 This is a schematic diagram illustrating an automated CFD method for optimal flow analysis around a blade based on reinforcement learning, according to an embodiment.

[0036] Figure 3 This is a schematic diagram illustrating the conceptual design of the blade parameterization method.

[0037] Figure 4 This is a schematic diagram illustrating the process of generating a computational mesh by setting the required meshing parameters in the mesh generator.

[0038] Figure 5 This is a schematic diagram illustrating the mesh generation algorithm.

[0039] Figure 6 The computational configuration for flow simulation through the blade passage is shown.

[0040] Figure 7 It is a graph showing the time spent per iteration for 1000 randomly generated computational configurations based on the number of units.

[0041] Figure 8 This is a schematic diagram showing the grid used to evaluate grid convergence.

[0042] Figure 9 This is a schematic diagram illustrating the weighted Chebyshev method for determining the Pareto front using two objective functions (normalized error function and normalized cost function).

[0043] Figure 10 This is a schematic diagram of multi-agent reinforcement learning used for automatic mesh generation to achieve optimal CFD simulation.

[0044] Figure 11 This is a schematic diagram illustrating an analytical apparatus for analyzing flow around blades using CFD according to an embodiment.

[0045] Figure 12 This is a flowchart illustrating an analysis method for flow around a blade using CFD according to an embodiment.

[0046] Figure 13 This is a schematic diagram illustrating the reinforcement learning results of a flow analysis method using CFD on a blade according to an embodiment.

[0047] Figure 14 This is a schematic diagram illustrating the optimal flow analysis results on a CFD-using blade automatically obtained after learning is completed in a flow analysis method on a CFD-using blade according to one embodiment.

[0048] Figure 15 This is an example diagram illustrating a neural network model according to an embodiment.

[0049] Figure 16 This is a schematic diagram illustrating an example of a computer device according to an embodiment.

[0050] Inventive Method In the following description, embodiments of the invention will be detailed with reference to the accompanying drawings to enable those skilled in the art to readily implement the invention. In the drawings, for clarity of illustration, parts irrelevant to the description have been omitted, and the same reference numerals are used throughout the specification for the same or similar components. Furthermore, in the drawings, some components may be exaggerated, omitted, or shown schematically, and the dimensions of each component do not perfectly reflect its actual size.

[0051] The accompanying drawings are intended only to facilitate an easy understanding of the embodiments disclosed in this specification, and the technical concepts disclosed in this specification are not limited to the drawings. It should be understood that all modifications, equivalents, or substitutions within the spirit and scope of the invention are included.

[0052] The terms first, second, and similar ordinal numbers can be used to describe various components, but these components are not limited by these terms. These terms are used only for the purpose of distinguishing one component from another.

[0053] Throughout this specification, terms such as “comprise,” “include,” or “have” should be understood to indicate the presence of features, numbers, steps, actions, components, elements, parts, or combinations thereof described in the specification, without excluding the possibility of the presence or addition of one or more other features, numbers, steps, actions, components, elements, parts, or combinations thereof. Therefore, when a part is referred to as “comprise” or “includes” an element, it means that unless otherwise specifically stated, it does not exclude other elements but may include additional elements.

[0054] Furthermore, throughout the specification, when something is described as being "connected," it does not necessarily mean that two or more components are directly connected. It could also mean that two or more components are indirectly connected through other components, physically connected as well as electrically connected, or referred to by different names depending on their location or function, but are actually a single unit.

[0055] Furthermore, terms such as “...unit”, “...device”, and “...module” described in the specification refer to units that perform at least one function or operation, and these units can be implemented as hardware, software, or a combination of hardware and software.

[0056] In this specification, "transmission or provision" may include not only direct transmission or provision, but also indirect transmission or provision via other devices or by using bypass routing.

[0057] The expressions written in the singular in this specification may be interpreted as singular or plural unless expressly stated as "a" or "single".

[0058] Figure 1 This is a flowchart illustrating an automated computational fluid dynamics method for optimal flow analysis on a blade based on reinforcement learning, according to an embodiment. Figure 2 This is a schematic diagram illustrating an automated computational fluid dynamics method for optimal analysis of flow around a blade based on reinforcement learning, according to an embodiment.

[0059] refer to Figure 1 and Figure 2 The automated computational fluid dynamics (CFD) method according to this embodiment includes inputting analysis conditions, determining analysis settings, and evaluating analysis results. Inputting analysis conditions may include a process of inputting blade shape and flow conditions (S110). Determining analysis settings may include a process of determining analysis settings for flow analysis on the blade (S120), and includes generating a mesh, selecting a turbulence model, and selecting a numerical method. Determining analysis settings may involve applying reinforcement learning methods, whereby learning or training can be performed to optimize the analysis settings (S130).

[0060] The evaluation analysis results may include performing flow analysis on the blades via CFD analysis (S140) and evaluating the results of the flow analysis on the blades (S150). When evaluating the analysis results, reinforcement learning methods may be applied to perform learning or training to optimize the analysis results (S160).

[0061] Reinforcement learning is the process of exploring actions that maximize rewards in a given state. For example, a state can be defined as input analysis conditions, an action as analysis setting parameters, and a reward as the accuracy and efficiency of the analysis results. In the case of reinforcement learning, deep reinforcement learning algorithms incorporating artificial neural network techniques can be used to perform the learning. In this embodiment, the actor-critic algorithm can be used, but other algorithms such as deep Q-networks and proximal policy optimization (PPO) can also be used.

[0062] After completing the learning process, the optimal analysis settings can be automatically determined based on newly given arbitrary analysis conditions, thereby automating the entire CFD process for flow analysis on blades. Optimal parameters can be determined non-iteratively without manual intervention, and the optimal mesh resolution can be identified.

[0063] The following describes in more detail, step by step, the automated computational fluid dynamics method for optimal flow analysis on blades.

[0064] First, input the blade shape and flow conditions under the input analysis conditions.

[0065] Blade parameterization methods can be used to generate various types of blades in axial-flow gas turbines, supersonic pulse turbines, axial-flow compressors, etc. Using blade shape parameters represented as BSP, a two-dimensional blade profile with continuous curvature can be generated from a non-uniform rational B-spline curve, as shown in mathematical expression 1.

[0066] [Mathematical Expression 1] .

[0067] Figure 3 This is a schematic diagram illustrating the conceptual design of the blade parameterization method.

[0068] refer to Figure 3 (a) Initially, the middle arc is constructed using the first six elements of the BSP. The chord of length C is determined by the stagger angle ψ at the leading edge. Then, the metal angles θ at the leading and trailing edges are used. le and θ te The curvature is defined by the arc. Here, the tangent ratio at each edge is determined by d. le and d te control.

[0069] like Figure 3 As shown in (b), the remaining elements of the BSP define the upper and lower blade profiles along the mid-arc line. The radii of curvature of the leading and trailing edges are determined by ρ. le and ρ te OK. Parameters ( )and( Define the thickness distribution of the upper and lower portions of the blade, where k indicates the number of parameters for the thickness distribution. In this example, k=6 is chosen because it has been found to be sufficient to accurately represent different ranges of turbine blade profiles while meeting the accuracy standards required for manufacturing blades for axial-flow gas turbines.

[0070] The range of each element of the BSP can be set wide enough to include blades with different shapes and applications from previous literature. In mathematical expression 2, this range is shown below.

[0071] [Mathematical Expression 2] Within these ranges, unrealistic blade shapes that are not simply connected or have two or more extreme points are excluded.

[0072] The blade geometry is captured by the blade spacing (BSP) and the pitch (the distance between the blades), and the flow conditions can be expressed using the isentropic Reynolds number Re at the outlet. is,out and Mach number Ma is,out and the inflow angle θ inIt is expressed in a dimensionless manner. As shown in mathematical expression 3, dimensionlessness can be achieved using the blade's boundary conditions (p... t,in T t,in and p out ) and axial chord length C ax To achieve this.

[0073] [Mathematical Expression 3] , , Here, ρ is,out = p out / (RT is,out This is derived from the ideal gas law, and μ is,out It can be calculated using Sutherland's law, where and This dimensionless approach not only adapts to simulations at different geometric and physical scales, but also incorporates four variables (p... t,in T t,in p out and C ax ) becomes two variables (Re) is,out and Ma is,out This reduces the amount of data required for learning.

[0074] The range of flow conditions is defined broadly enough to cover complex flow characteristics such as impingement and separation, as shown in mathematical expression 4.

[0075] [Mathematical Expression 4] Here, θ in By defining relative to θ le The value of θ is adaptively adjusted based on the blade shape. le It is the metal angle at the leading edge to mitigate unrealistic flow conditions.

[0076] When determining the analysis settings, generate a computational grid suitable for the analysis and select a turbulence model and numerical method.

[0077] Figure 4 This is a schematic diagram illustrating the process of generating a computational mesh by setting the required meshing parameters in the mesh generator.

[0078] refer to Figure 4 A mesh generator trained using deep reinforcement learning can generate structured meshes for two-dimensional blade channels using elliptical mesh generation methods. HOH-type meshes can be generated by specifying meshing parameters, which combine H-type meshes at the inlet and outlet sides with O-type meshes near the blades. Descriptions of the parameters are listed in Table 1.

[0079] Table 1 Figure 5 This is a schematic diagram illustrating the mesh generation algorithm.

[0080] Given a blade shape with a pitch, the boundary shape of the mesh can be defined based on the following parameters: y in y out α 中弧线 x o in and x o out Here, y in and y out The inlet and outlet positions in the vertical direction are determined separately. To accommodate mesh generation for blades with different curvatures, the curvature of the periodic boundary is determined by α. 中弧线 Modulation, its range from linear (α) 中弧线 =0) to perfectly match the middle arc (α) 中弧线 =1). The interface positions between the O-type and H-type grids on the inlet and outlet sides are respectively determined by x. o in and x o out Sure.

[0081] Subsequently, the grid resolution is based on N t N n β le β te The number of nodes in the O-type mesh along the tangent and normal directions relative to the blade surface is determined by N. t and N n Specify. N t The nodes can be distributed along the blade surface to ensure increased resolution at the leading and trailing edges. This distribution follows a hyperbolic tangent function, where β... le and β te The clustering strengths towards the leading and trailing edges are controlled separately. N n The nodes are distributed in a direction perpendicular to the blade surface, clustering on the blade surface according to a hyperbolic tangent function to adequately address near-wall physics. Clustering can be defined such that the height of the first element at the blade surface corresponds to Δn1. After the nodes are distributed within the domain, an elliptic mesh generation method can be used to finally generate the mesh, ensuring high orthogonality near the blade surface without slope discontinuities.

[0082] The goal of the first training step is to enable the mesh generator to autonomously select the geometric control portion from the entire set of meshing parameters (see Table 1) for various blade channel configurations. To achieve this goal, the state s can be defined as shown in the following mathematical expression 5.

[0083] [Mathematical Expression 5] , Where BSP represents the blade shape parameters as outlined in mathematical expression 1, pitch refers to the distance between blades, and N o =N t × N n This represents the total number of nodes in the O-type mesh. It includes N. o β le β te The reason for using Δn1 as a state rather than an optimized action is to provide adaptability to the grid resolution under various flow conditions. These parameters can be optimized in subsequent training phases involving CFD simulations.

[0084] Action a includes geometric control meshing parameters, and action a can be defined as follows in mathematical expression 6.

[0085] [Mathematical Expression 6] , The total grid resolution N can be allocated by combining δ. o It is predefined as a state variable along the normal and tangent directions, δ=N t / N n In other words, δ represents the ratio of the number of nodes in the O-type mesh in the tangential direction about the blade surface to the number of nodes in the normal direction.

[0086] The reward can be defined based on the mesh quality that minimizes the numerical error caused by geometric imperfections in the mesh. For this purpose, two mesh quality metrics are considered: the ratio of the determinants of the Jacobian matrices, Q. J and element skewness Q S Measuring Q J and Q S These are essential for achieving a uniform cell distribution and maintaining high orthogonality between adjacent cells. Due to these properties, the metric can be used to generate high-quality quadrilateral meshes. Q J The calculation can be given by the following mathematical expression 7.

[0087] [Mathematical Expression 7] , Among them, J i,j It is each node (x) i,j ,y i,jThe determinant of the Jacobian matrix at position is calculated as shown in the following mathematical expression 8.

[0088] [Mathematical Expression 8] , Where x and y are the coordinates of the node, and i and j are the indices of the node. Q J The value represents the regional consistency between adjacent cells; a higher value indicates a smoother resolution variation. Skewness Q S The following mathematical expression 9 can be used to calculate it.

[0089] [Mathematical Expression 9] , Where, θ min and θ max This represents the minimum and maximum values ​​of the interior angles of the unit. Higher Q S The value indicates that the cell distortion is reduced and the orthogonality with neighboring cells is better. The reward r can be defined by the following mathematical expression 10.

[0090] [Mathematical Expression 10] , Among them, ( )| min and ( )| avg Let represent the minimum and average values ​​of the cells in the O-type mesh, respectively. This focus on the O-type mesh stems from the potential distortion of its cells to conform to the blade profile, unlike the cells in the typically rectangular H-type mesh. To calculate Q... J Additional nodes can be incorporated at the boundaries of the O-type mesh to ensure a smooth transition at mesh interfaces and periodic boundaries. By considering the minimum and average values, reward functions can be designed to effectively evaluate the impact of cells with the lowest quality and provide a comprehensive view of the overall quality distribution of the mesh. The values ​​of these metrics (Q...) J )| min (Q) J )| avg (Q) S )| min and(Q S )| avg By definition, the range is between 0 and 1, thus exhibiting a comparable scale. A squared term is used in the reward function to increase sensitivity to higher values.

[0091] The first step of training uses a single-step actor-critic algorithm. For each round, the state s is randomly assigned. Based on the state, the actor determines the action. The network has exploratory noise ε. The mesh generator then produces a mesh using meshing parameters from a given state and selected actions. The quality of the generated mesh is then evaluated, and a reward r is calculated. The data (s, a, r) ​​is stored in a buffer. Using the data from the buffer, the commentator network is updated to predict the reward more accurately, while the actor network is refined to produce actions that maximize the reward. Afterward, the next round begins. This process is repeated until the network converges. Because it is trained on a randomly given leaf configuration in each round, rather than sequentially on a single configuration before moving to the next round, the convergent network can generate an optimal mesh for configurations within a defined range.

[0092] The actors and commentators are constructed as a fully connected network. All hidden layers use the leaky ReLU activation function, with the exception of the network's output layer, which uses the hyperbolic tangent function to bind action values. The learning rate is 10. -4 The Adam optimizer is used with mini-batch size N b Updating network parameters every 100 iterations is a common practice in actor-critic algorithms. To enhance stability during the learning process, the actor network is updated only after being updated twice by the critic network. The exploration noise ε follows a normal distribution N(0,σ). 2 ) is generated, where the mean is 0 and the standard deviation is σ, where σ is defined as in the following mathematical expression 11.

[0093] [Mathematical Expression 11] In the early rounds, higher σ values ​​are used to collect different datasets. Subsequently, a cosine function is employed to systematically vary the noise level, balancing exploration and exploitation while reducing the impact of specific noise values. Both networks consist of four hidden layers with 512, 256, 256, and 128 neurons, respectively.

[0094] As the number of rounds increases, the network is updated to produce a higher geometric quality mesh for each newly introduced blade channel configuration. Therefore, high-reward data accumulates in the buffer, increasing J. π The value is calculated and eventually converges. To fully train the network for the current optimization problem, approximately 10 [units / times] are needed. 6 Each round. The mesh generator trained in this phase will be used in subsequent training phases for optimal CFD, where CFD simulations can be integrated to optimize resolution-dependent meshing parameters.

[0095] In this embodiment, the selection of the turbulence model includes choosing the type and parameters of the turbulence model. For example, the turbulence model can be selected from k-ε, k-ω, and k-ω SST. To select the turbulence model, the analysis setting determination unit 50 can be optimized using deep reinforcement learning.

[0096] Furthermore, the selection of numerical methods in this embodiment may include the selection of discretization techniques, matrix computation algorithms, time step size, and residual criteria. For example, the discretization technique may be selected from first-order upwind or second-order central difference, and the matrix computation algorithm may be selected from flexible generalized minimum residual, conjugate gradient, or stable biconjugate gradient. For the selection of numerical methods, deep reinforcement learning can be used to optimize the analysis setting of the determining unit (50).

[0097] The evaluation analysis results involve performing a flow analysis on the blade using computational fluid dynamics with the defined analysis settings, and then evaluating the results.

[0098] The open-source software SU2 can be used to solve the two-dimensional steady-state compressible Reynolds-averaged Navier-Stokes (RANS) equations for blade passage simulation. These equations are expressed in mathematical expression 12 as follows.

[0099] [Mathematical Expression 12] , Among them, F c and F v These represent convective flux and viscous flux, respectively, as shown in the mathematical expression 13 below.

[0100] [Mathematical Expression 13] , Where ρ is the fluid density, v is the velocity vector, and E is the total energy per unit mass. It is an identity matrix, where p is the hydrostatic pressure, κ is the thermal conductivity, and T is the temperature. It is the viscous stress tensor defined by the following mathematical expression 14.

[0101] [Mathematical Expression 14] , Where μ represents viscosity.

[0102] For turbulent flow, based on the Boussinesq assumption, μ is divided into dynamic viscosity μ d and turbulent viscosity μ t , represented as μ=μ d +μ t Similarly, thermal conductivity is expressed as κ = κ d +κ t κ is composed of viscosity μ and specific heat capacity at constant pressure c. p And the Prandtl number Pr is calculated as κ = (c p μ) / Pr. Dynamic viscosity μ d It was calculated using Sutherland's law. Turbulent viscosity μt This is derived from a turbulence model. In this embodiment, a shear stress transport-k-ω (SST-k-ω) turbulence model can be employed. This model uses a k-ω model near the wall and transitions to a k-ε model in free flow, allowing for accurate predictions in both the near-wall and core regions using a mixing function.

[0103] To close mathematical expression 12, an ideal gas equation of state with a specific heat ratio γ and a gas constant R is adopted. Pressure is determined by p = (γ - 1)ρ[E - 0.5(v·v)], temperature by T = p / (ρR), and the isobaric specific heat capacity is c. p =(γR) / (γ-1). The governing equations are discretized using the finite volume method (FVM). The flux is discretized using the central Jameson-Schmidt-Turkel (JST) scheme. The gradient of the flow variables required to evaluate the viscous flux is calculated using the weighted least squares method.

[0104] Figure 6 The computational configuration for flow simulation through the blade passage is shown.

[0105] refer to Figure 6 At the inlet, the total pressure p t,in and total temperature T t,in With the inflow angle θ in Apply static pressure p out Apply at the outlet. Specify periodic conditions for the upper and lower boundaries, separated by pitch distance. Specify no-slip and adiabatic wall boundary conditions on the blade surfaces.

[0106] When evaluating the results of an analysis, computational accuracy and efficiency can be considered. In multi-objective optimization, the concept of Pareto optimality allows both computational accuracy and efficiency to be applied as objective functions simultaneously, and this can be achieved by adjusting the weights between the two objective functions. Computational accuracy can be calculated by comparing the results of the analysis with high-quality data, such as experimental data or high-resolution numerical data, to the current analysis results. Computational efficiency can be calculated by measuring the total time spent performing the analysis.

[0107] A reward function can be designed to evaluate the analysis results while simultaneously considering the accuracy and efficiency of the simulation. To this end, two objective functions are employed in the multi-objective method: the cost and error of the simulation. First, the cost function (denoted as r...) c It is designed to reflect the total time required for one simulation and can be calculated as shown in the mathematical expression 15 below.

[0108] [Mathematical Expression 15] , Where, N总 This represents the total number of units. The formula is based on a time complexity of O(N) for each iteration. 总 The observation can be made by simulating 1000 random flow configurations (e.g.) Figure 7 (As shown). This linear scaling is attributed to sparse matrix operations in the current iterative solver, which utilizes the Flexible Generalized Minimal Residual (FGMRES) method enhanced with an Incomplete Up-Down (ILU) preprocessor.

[0109] The error function can be designed to evaluate mesh convergence by calculating the difference between simulation results obtained using two successively refined meshes (a base mesh and a fine mesh). The base mesh initially generated by the DRL network can be further refined to create a fine mesh by doubling the resolution along each axis, as shown below. Figure 8 As shown on the left. Bicubic interpolation can be used to preserve the blade curvature, especially at the leading and trailing edges, such as... Figure 8 As shown on the right. The error function can be defined by the root mean square (RMS) error calculated at the node locations of the base mesh, which remains unchanged during the refinement process, as shown in the following mathematical expression 16.

[0110] [Mathematical Expression 16] The error Ma can be selected to capture the flow characteristics necessary for blade channel analysis, such as boundary layer, wake, and shock wave. Furthermore, its dimensionless nature facilitates scaling across different flow configurations. The error across the entire flow field is evaluated by the first term, while the error along the blade surface is incorporated as the second term to reflect near-wall flow physics. For the second term, the isentropic Mach number Ma... is This is achieved by replacing p in mathematical expression 3 with the pressure along the blade surface. out It is used for calculation.

[0111] Finally, using the weighted Chebyshev method that guarantees the identification of all Pareto optimal solutions to convex and non-convex problems, the reward function r can be defined as shown in the following mathematical expression 17.

[0112] [Mathematical Expression 17] , Here, w represents the weight between the two objective functions, and can be scaled as shown in the following mathematical expression 18.

[0113] [Mathematical Expression 18] Logarithmic scaling can be used to mitigate the impact of outliers during the learning process. By making r e Multiply by 10 2and make r c Divide by 10 7 This allows for matching the scale between two objective functions to achieve better learning performance.

[0114] exist Figure 9 This paper illustrates a schematic explanation of the weighted Chebyshev method used to determine the Pareto front. For a given weight w, in two objective functions r... * e and r * c A rectangle with a fixed aspect ratio and variable dimensions is constructed on the plane. As indicated by the minus sign in mathematical expression 17, the learning process aims to find the rectangle with the smallest dimensions, which produces the Pareto optimal solution. All Pareto front solutions can be obtained by changing the value of w from 0 to 1.

[0115] To mitigate the high computational costs associated with acquiring data through CFD simulations, methods for data reproduction can be applied to increase the quantity and diversity of data. Once r is obtained from a single simulation... e and r c Various weights can be applied to these values ​​to calculate the reward function. For example, 100 datasets can be reproduced by randomly varying the weights w between 0 and 1 for each round.

[0116] To completely eliminate human intervention in the mesh generation process, the remaining resolution control parameters are optimized through CFD simulation and defined as action a according to the following mathematical expression 19.

[0117] [Mathematical Expression 19] .

[0118] State s is formulated to reflect the computational configuration, including blade geometry and flow conditions, as shown in the mathematical expression 20 below.

[0119] [Mathematical Expression 20] , In this approach, the variable w is combined into a state variable ranging between 0 and 1, rather than a fixed value, to determine the weights between two objectives: the cost and error of the simulation. This method allows users to obtain an optimal grid that balances accuracy and efficiency according to their preferences.

[0120] Multi-agent reinforcement learning can be used to accelerate the data acquisition speed required to train a mesh generator for optimal CFD. For example... Figure 10 As shown, multiple agents can perform CFD simulations in parallel to collect data.

[0121] States are randomly assigned to individual agents, each state consisting of the blade geometry represented by the BSP and pitch, and the state represented by Re. is,out Ma is,out θ in The flow conditions are represented by weights w. Each agent is based on action. Generate a mesh, explore noise ε, and then continue the CFD simulation. When agent i completes the CFD simulation, calculate the error function r based on the results. e,i and cost function r c,i And calculate the reward r i The amount of data was subsequently amplified to (s) using data reproduction methods. i ,a i ,r i )1,...,(s i ,a i ,r i ) 100 The data is then stored in a buffer. This stored data is then used to update the actor and critic network. Subsequently, at the start of the next round, new tasks are assigned to the corresponding agents. Because tasks are dynamically reassigned immediately after the agents complete them, the data acquisition rate increases proportionally to the number of agents. This process is repeated until the network converges.

[0122] A total of 5 × 10 is needed. 5 The network was fully trained using RANS simulations. Training was conducted over 8 weeks using 256 agents, each running on a single CPU core of an Intel(R) Xeon(R) E5-2650 v2 processor.

[0123] Analysis equipment using CFD to study the flow around the blades can be applied... Figure 1 The computational fluid dynamics automation method shown is used to achieve this, and the device can be used to perform flow analysis on blades using CFD.

[0124] Figure 11 This is a schematic diagram illustrating an analysis apparatus for analyzing flow around blades using CFD according to an embodiment, and Figure 12 This is a flowchart illustrating an analysis method for flow around a blade using CFD according to an embodiment.

[0125] refer to Figure 11 The analysis device 100 for flow around a blade using CFD, according to an embodiment, may include an automated analysis model generator 130 and a flow analyzer 150 on the blade. (See reference...) Figure 12The method for analyzing flow around a blade using CFD according to an embodiment may include an automated analysis model generation S10, analysis condition input S20, and analysis result output S30. The automated analysis model generator 130 performs the automated analysis model generation step, and the flow analyzer 150 on the blade may perform the analysis condition input step S20 and the analysis result output step S30.

[0126] The analysis automation model generator 130 can generate an analysis setting model 132 that determines the analysis setting parameters based on the input analysis conditions, and an analysis performance model 136 that predicts the CFD flow analysis results based on the input analysis conditions.

[0127] The analysis setup model 132 can select an analysis setup for flow analysis on the blade based on input analysis conditions, namely blade shape and flow conditions. It can be trained using a deep reinforcement learning algorithm employing an artificial neural network to ensure that the selected analysis setup meets predetermined quality standards. The analysis setup model 132 can determine the analysis setup parameters for flow analysis on the blade by receiving blade shape and flow conditions as input.

[0128] Analysis setup for flow analysis on blades can include generating a computational grid, selecting a turbulence model, and selecting a numerical method. Generating the computational grid may involve generating a structured grid for the two-dimensional blade passage using an elliptic mesh generation method. Selecting a turbulence model may include choosing the type and parameters of the turbulence model. The type of turbulence model can be selected from k-ε, k-ω, or k-ω SST. Selecting a numerical method may include selecting a numerical method, choosing a matrix calculation algorithm, and setting the time step size and residual criteria.

[0129] The blade shape can be represented by the blade shape parameter (BSP) according to mathematical expression 1 and the blade pitch. Flow conditions may include the isentropic Reynolds number Re at the outlet. is,out and Mach number Ma is,out and the entrance angle θ in Reynolds number Re is,out and Mach number Ma is,out and entrance angle θ in It can be expressed in a dimensionless manner according to mathematical expression 3.

[0130] The generative analysis setup model may include defining states s and actions a, generating a grid using gridded parameters obtained by determining actions based on the given states in each round, evaluating the quality of the generated grid and calculating rewards, and training using a single-step actor-critic algorithm. Here, states s can be defined according to mathematical expression 5, and actions a can be defined according to mathematical expression 6.

[0131] The reward can be defined based on the mesh quality that minimizes numerical errors caused by geometric imperfections in the generated mesh. Specifically, it can be calculated by using the ratio Q of the determinants of the Jacobian matrix. J and element skewness Q S The reward r is defined by a mesh quality metric to evaluate mesh quality. Here, the reward r can be defined according to mathematical expression 10.

[0132] The analytical performance model 136 can perform CFD simulations for flow analysis on blades based on determined analytical setup parameters, and can be trained using a deep reinforcement learning algorithm utilizing artificial neural networks to meet predetermined evaluation criteria by assessing the results of the performed flow analysis on the blades. The analytical performance model 136 can predict CFD flow analysis results based on the determined analytical setup and input analytical conditions.

[0133] Evaluation criteria for assessing the results of flow analyses on blades can include reward functions that evaluate the accuracy and efficiency of computations in CFD simulations. Computational accuracy can be determined by measuring the error between the results of the performed flow analysis on the blade and high-quality data, including experimental or high-resolution numerical data. Computational efficiency can be determined by measuring the total time spent performing the flow analysis on the blade.

[0134] The accuracy of the computation can be evaluated by calculating and assessing the error function defined according to mathematical expression 16. To assess the efficiency of the computation, the cost function defined according to mathematical expression 15 can be calculated and assessed. The reward function r defined according to mathematical expression 17 can be calculated and used to consider both the accuracy and efficiency of the computation.

[0135] The generation of the performance model can include: defining a state s and an action a, performing actions based on the given state to complete each round of CFD simulation, calculating rewards based on the CFD simulation results, and training using an actor-critic algorithm. Here, the state s can be defined according to mathematical expression 20, and the action a can be defined according to mathematical expression 19.

[0136] Generative analysis performance models can include the use of multi-agent reinforcement learning, where multiple agents perform CFD simulations in parallel to collect training data. This can accelerate the data acquisition speed required to train a mesh generator to achieve optimal CFD performance.

[0137] The flow analyzer 150 on the blade receives arbitrary blade shape conditions and flow conditions, applies an automatic analysis model to the input blade shape conditions and flow conditions to determine the analysis setting parameters, performs CFD simulation, and can predict and output CFD flow analysis results. Through the prediction of the CFD flow analysis results, the flow field can be predicted and output.

[0138] As described above, once training is complete in the analysis automation model generator 130, the mesh generator in the flow analyzer 150 on the blade can generate a mesh that provides a convergent solution within the desired computational cost for any blade shape and flow conditions in a single simulation. In other words, by automatically determining the optimal analysis settings, the repetitive CFD process typically required for mesh convergence can be eliminated. Furthermore, the flow analyzer 150 on the blade can use CFD to predict optimal flow analysis results on the blade with improved accuracy and efficiency based on the automatically determined optimal analysis settings.

[0139] Figure 13 This is a schematic diagram illustrating the reinforcement learning results of a computational fluid dynamics flow analysis method on a blade according to an embodiment.

[0140] The blade shape of the high-pressure turbine is trained to minimize the root mean square error (RMSE) of the flow field, while considering a Reynolds number (Re) of 5.90 × 10⁻⁶. 5 The computational efficiency was evaluated under flow conditions of Mach number (Ma) 0.84 and angle of attack (α) 0.0 degrees. As a result, the root mean square error (RMSE) of the wall flow variables, shown in the upper part of the figure, decreased from 3.9 × 10⁻⁶. -3 Reduced to 9.87×10 -4 Furthermore, as shown in the center of the figure, the error at the trailing edge decreases significantly, resulting in a reduction in the root mean square error (RMSE) of the entire flow field from 1.007 × 10⁻⁶. -2 Reduced to 3.734 × 10 -3 In fact, as shown in the lower part of the figure, it is confirmed that the physical phenomena in the wake region were captured better after training.

[0141] Figure 14 This is a schematic diagram illustrating the optimal flow analysis results on a blade using CFD, automatically obtained after learning is completed in a computational fluid dynamics flow analysis method on a blade according to an embodiment.

[0142] The left side of the figure shows the analysis results of the high-pressure turbine blade shape under flow conditions with shock waves. Typically, due to the rapid changes in the flow field gradient caused by shock waves, nonlinearity is increased, requiring even experienced engineers to perform repetitive analyses to obtain accurate results. However, a trained neural network can predict the location of the shock wave in advance, and it produces accurate results with only one analysis, where the root mean square error (RMSE) at the wall is 1.482 × 10⁻⁶. -3 The root mean square error in the entire flow field is 4.909 × 10⁻⁶. -3 The right side of the figure shows the analysis results for the low-pressure turbine blade shape. Compared to the high-pressure turbine blade, this blade has a greater curvature, making computational mesh generation very challenging. However, the trained neural network not only automatically generates the mesh but also achieves high accuracy without iterative analysis, with an RMSE of 1.339 × 10⁻⁶ at the wall. -3 The RMSE in the entire flow field is 3.289 × 10⁻⁶. -3 .

[0143] Figure 15 This is an example diagram illustrating a neural network model 800 according to an embodiment. As an example of a machine learning model, the neural network model 800 is a statistical learning algorithm implemented based on the structure of biological neural networks in machine learning techniques and cognitive science, or a structure that executes the algorithm.

[0144] According to an embodiment, neural network model 800 can represent a machine learning model that learns to reduce the error between the correct output and the inferred output corresponding to a specific input, thereby enabling it to solve problems. This is achieved through artificial neurons called nodes, which form a network through synaptic connections similar to those in biological neural networks, with the weights of the synapses being repeatedly adjusted. For example, neural network model 800 can include any probabilistic model or neural network model used in artificial intelligence learning methods such as machine learning and deep learning.

[0145] The neural network model 800 is implemented as a multilayer perceptron (MLP) comprising multiple layers of nodes and connections between them. According to embodiments, the neural network model 800 can be implemented using one of various artificial neural network model architectures that include an MLP. For example... Figure 15As shown, the neural network model 800 may include: an input layer 820, where blade shape and flow conditions are input as input data; an output layer 840, which outputs analysis setting parameters and CFD flow analysis results as output data; and n hidden layers 830_1 to 830_n located between the input layer 820 and the output layer 840, where n is a positive integer. These hidden layers receive signals from the input layer 820, extract features, and transmit them to the output layer 840. Here, the output layer 840 receives signals from the hidden layers 8301 to 830_n and outputs signals to the outside.

[0146] Once the machine learning model is trained, the analysis automation model generator 130 of the computational fluid dynamics flow analysis device 100 on the blade, according to the embodiment, can store the trained analysis setup model 132 and analysis performance model 136. Here, the trained analysis setup model 132 and analysis performance model 136 can be stored in the memory of the flow analyzer 150 on the blade. Alternatively, the trained analysis setup model 132 and analysis performance model 136 can be stored in the memory of a server connected to the flow analyzer 150 on the blade via a wired or wireless network.

[0147] The flow analyzer 150 on the blade outputs predicted CFD flow analysis results by applying a predictive model generated by the analysis automation model generator 130 to the input blade shape conditions and flow conditions.

[0148] Figure 16 An example of a computer device according to an embodiment is shown. Reference Figure 11 The described automated analysis model generator 130 and the flow analyzer 150 on the blade can be generated by... Figure 16 The computer device 900 shown is implemented.

[0149] Computer device 900 may include memory 910, processor 920, communication interface 930, and input / output interface 940. Memory 910, as a computer-readable recording medium, may include random access memory (RAM), read-only memory (ROM), and permanent mass storage devices such as disk drives. Additionally, memory 910 may store an operating system and at least one program code. These software components may be loaded into memory 910 from a separate computer-readable recording medium. This separate computer-readable recording medium may include computer-readable recording media such as hard disks, flash memory, optical disks, and external hard disk drives. These software components may also be loaded into memory 910 via communication interface 930.

[0150] Processor 920 can be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions can be provided to processor 920 via memory 910 or communication interface 930.

[0151] The communication interface 930 can provide the computer device 900 with the ability to communicate with other devices via the network 1000.

[0152] Input / output interface 940 can be a device for interface connection with input / output device 950. For example, input devices may include devices such as microphones, keyboards, or mice, while output devices may include devices such as displays and speakers.

[0153] The above embodiments can be implemented in the form of a computer program, which can be executed by various components of a computer, and such a computer program can be recorded on a computer-readable medium. Here, the medium may include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floppy disks; and hardware devices specifically configured to store and execute program instructions, such as ROMs, RAMs, flash memory, etc.

[0154] Unless there is an explicit statement of the order of steps in the method according to the embodiments or conflicting information is provided, these steps may be performed in an appropriate order. The invention is not necessarily limited to the order in which these steps are described. Throughout this invention, all examples or illustrative terms (e.g., etc.) are used merely to explain the invention in detail and not to limit its scope. Furthermore, those skilled in the art will recognize that various modifications, combinations, and alterations can be made within the scope of the patent claims or their equivalents.

[0155] Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by those skilled in the art to which this invention pertains also fall within the scope of the present invention.

Claims

1. A computational fluid dynamics (CFD) flow analysis method for blades, the method comprising: An automated analysis model is generated, which predicts CFD flow analysis results based on input analysis conditions. The generation of the automated analysis model includes: inputting blade shape and flow conditions, determining analysis settings for flow analysis on the blade, performing flow analysis on the blade by performing computational fluid dynamics CFD simulation based on the determined analysis settings parameters, evaluating the results of the performed flow analysis on the blade, and training the model using reinforcement learning of an artificial neural network to meet predetermined evaluation criteria. Input arbitrary blade shape conditions and flow conditions; and The analysis automation model is applied to the input blade shape conditions and flow conditions to determine the analysis settings parameters, predict the CFD simulation, and output the predicted CFD simulation as the flow analysis result.

2. The method according to claim 1, wherein, Generating the automated analysis model includes: A model for generating an analysis setup is provided. This model determines analysis setup parameters based on the input analysis conditions. Generating the model includes: selecting an analysis setup for flow analysis on the blade based on the input blade shape and flow conditions; and training the analysis setup model using a deep reinforcement learning algorithm employing an artificial neural network, such that the selected analysis setup meets predetermined quality standards. Generate an analytical performance model that predicts the results of CFD flow analysis. Generating the analytical performance model includes: performing flow analysis on the blade using the CFD simulation, evaluating the results of the performed flow analysis on the blade, and training the analytical performance model using a deep reinforcement learning algorithm utilizing an artificial neural network to meet the evaluation criteria.

3. The method according to claim 2, wherein, The analysis setup for flow analysis on blades includes generating a computational grid, selecting a turbulence model, and selecting a numerical method.

4. The method according to claim 3, wherein, The blade shape is represented by the blade shape parameter BSP and the blade pitch, and the flow conditions are represented in a dimensionless manner by the isentropic Reynolds number and Mach number at the outlet and the inflow angle.

5. The method according to claim 3, wherein, Generating the analysis setup model includes: Define state s as follows: ， Define action 'a' as the following equation. ,as well as A grid is generated using gridded parameters obtained by determining actions based on a given state in each round. The quality of the generated grid is evaluated and a reward is calculated. The analytical setup model is then trained using a single-step actor-critic algorithm. Here, BSP represents the blade shape parameters, pitch represents the blade spacing, and N o β represents the total number of nodes in the O-type mesh. le β represents the degree of clustering of nodes at the leading edge. te This indicates the clustering degree of the nodes at the trailing edge, and Δn1 represents the height of the first unit in the direction perpendicular to the blade surface. y in Indicates the entrance location in the vertical direction, y out Indicates the exit location in the vertical direction, α 中弧线 Indicates the degree of curvature of the periodic boundary following the mid-curve, x o in Indicates the horizontal position of the HO-type interface, x o out δ represents the position of the OH-type interface in the horizontal direction, and δ represents the ratio of the number of nodes in the tangential direction of the O-type mesh relative to the blade surface to the number of nodes in the normal direction.

6. The method according to claim 5, wherein, The reward is defined based on the mesh quality that minimizes numerical errors caused by geometric defects in the generated mesh, and The grid quality is evaluated by calculating a reward defined using a grid quality metric, which includes the ratio of the determinants of the Jacobian matrix, Q. J and element skewness Q S .

7. The method according to claim 2, wherein, The evaluation criteria used to assess the results of the flow analysis performed on the blades include a reward function that evaluates the computational accuracy and computational efficiency in the CFD simulation.

8. The method according to claim 7, wherein, The calculation accuracy is determined by the error between high-quality data, including experimental data or high-resolution numerical data, and the results of the flow analysis performed on the blade. The computational efficiency is calculated by measuring the total time required to perform flow analysis on the blade.

9. The method according to claim 8, wherein, Generating the analytical performance model includes: Define state s as follows: ， Define action 'a' as the following equation. ,as well as Actions are executed based on the given state in each round to complete the CFD simulation, the reward is calculated from the CFD simulation results, and the analytical performance model is trained using an actor-critic algorithm. Here, BSP represents the blade shape parameter, pitch represents the blade spacing, and Re... is,out Ma represents the isentropic Reynolds number at the exit. is,out θ represents the Mach number at the exit. in Indicates the inflow angle, w represents the weight, and N represents the inflow angle. o β represents the total number of nodes in the O-type mesh. le β represents the degree of clustering of nodes at the leading edge. te Δn1 represents the clustering degree of the nodes at the trailing edge, and Δn1 represents the height of the first unit in the direction perpendicular to the blade surface.

10. The method according to claim 2, wherein, Generating the analytical performance model involves collecting learning data by using multi-agent reinforcement learning and performing CFD simulations in parallel with multiple agents.

11. A computational fluid dynamics (CFD) flow analysis apparatus for blades, the apparatus comprising: An automated analysis model generator generates an automated analysis model that predicts computational fluid dynamics flow analysis results based on input analysis conditions. The automated analysis model generator takes blade shape and flow conditions as input, determines the analysis settings for flow analysis on the blade, performs flow analysis on the blade by performing computational fluid dynamics simulation based on the determined analysis settings parameters, evaluates the results of the performed flow analysis on the blade, and trains the automated analysis model to meet predetermined evaluation criteria by using reinforcement learning of an artificial neural network. and The flow analyzer on the blade receives input of arbitrary blade shape conditions and flow conditions, applies the automated analysis model to the input blade shape conditions and flow conditions to determine the analysis setting parameters, performs CFD simulation to predict the flow field, and outputs the flow field as the CFD flow analysis result.

12. The device according to claim 11, wherein, The automated analysis model generator includes: An analysis setup model is configured to determine analysis setup parameters based on the input analysis conditions. This model selects analysis settings for flow analysis on the blade based on the input blade shape and flow conditions, and is trained using a deep reinforcement learning algorithm employing artificial neural networks to ensure that the selected analysis settings meet predetermined quality standards. An analytical performance model is used to predict the results of CFD flow analysis. The analytical performance model performs flow analysis on the blade using the CFD simulation, evaluates the results of the performed flow analysis on the blade, and is trained using a deep reinforcement learning algorithm with an artificial neural network to meet the evaluation criteria.

13. An automated computational fluid dynamics (CFD) method for optimal flow analysis on blades, the method comprising: The input includes analysis conditions such as blade shape and flow conditions; Determine the analysis settings for flow analysis on the blades; Flow analysis on the blades is performed by computational fluid dynamics simulation based on the determined analysis settings parameters; Evaluate the results of the flow analysis performed on the blades; and By using reinforcement learning to train an artificial neural network, the results of flow analysis on the blades are made to meet predetermined evaluation criteria.

14. The method of claim 13, further comprising: The analysis setup for flow analysis on the blade is selected based on the input blade shape and flow conditions, and trained using a deep reinforcement learning algorithm with artificial neural networks to ensure that the selected analysis setup meets predetermined quality standards.

15. The method according to claim 14, wherein, The analysis setup for flow analysis on blades includes generating a computational grid, selecting a turbulence model, and selecting a numerical method.

16. The method according to claim 13, wherein, The evaluation criteria used to assess the results of the flow analysis performed on the blades include a reward function that evaluates computational accuracy and computational efficiency in the CFD simulation.

17. The method according to claim 16, wherein, The calculation accuracy is determined by the error between high-quality data, including experimental data or high-resolution numerical data, and the results of the flow analysis performed on the blade. The computational efficiency is calculated by measuring the total time required to perform flow analysis on the blade.

18. The method according to claim 13, wherein, Generative analysis of performance models involves collecting learning data by using multi-agent reinforcement learning and performing CFD simulations in parallel with multiple agents.

19. The method according to claim 13, wherein, The reinforcement learning includes computation using deep reinforcement learning algorithms based on artificial neural network technology, including deep Q-networks, proximal policy optimization (PPO), or actor-commentator algorithms.

20. A computer device comprising at least one processor configured to execute computer-readable instructions, wherein, The at least one processor is configured to: An automated analysis model is generated to predict CFD flow analysis results based on input analysis conditions. The generation of the automated analysis model includes: inputting blade shape and flow conditions, determining analysis settings for flow analysis on the blade, performing flow analysis on the blade by performing computational fluid dynamics simulation based on the determined analysis settings parameters, evaluating the results of the performed flow analysis on the blade, and training the automated analysis model to meet predetermined evaluation criteria by using reinforcement learning of artificial neural networks. Input arbitrary blade shape conditions and flow conditions; and The analysis automation model is applied to the input blade shape and flow conditions to determine the analysis settings parameters, predict the CFD simulation, and output the predicted CFD simulation as the flow analysis result.

Images