CFD automation method for optimal airfoil flow analysis based on reinforcement training, CFD airfoil flow analysis method, and CFD airfoil flow analysis device

Abstract

The disclosed computational fluid dynamics (CFD) airfoil flow analysis method comprises: an analysis automation model generation step for inputting the shape and flow conditions of a blade, determining analysis settings for airfoil flow analysis, performing the airfoil flow analysis by performing a computational fluid dynamics simulation on the basis of parameters for the determined analysis settings, and generating a model for predicting a CFD flow analysis result according to input analysis conditions by evaluating the results of the performed airfoil flow analysis to train the model through reinforcement training using an artificial neural network so as to satisfy a preset evaluation criterion; an analysis condition input step for inputting a shape condition and flow condition of an arbitrary blade; and an analysis result output step for determining the analysis setting parameters by applying the analysis automation model to the input shape condition and flow condition of the blade and performing CFD simulation to output the predicted CFD flow analysis result.

Description

CFD automation method for optimal analysis of airfoil flow based on reinforcement learning, CFD airfoil flow analysis method, and CFD airfoil flow analysis device

[0001] The present disclosure relates to a CFD automation method for optimal analysis of airfoil flow based on reinforcement learning, a CFD airfoil flow analysis method, and a CFD airfoil flow analysis device.

[0002] Flow analysis around airfoils plays a crucial role in the design of diverse fluid machinery, including aircraft wings, ship propellers, aircraft engines, gas turbines, steam turbines, and compressors. This flow analysis is primarily performed using Computational Fluid Dynamics (CFD) and is considered an essential step for design optimization. Accurate and efficient flow analysis using CFD requires specialized setup, including the creation of a computational grid appropriate for the flow conditions and geometry, selection of a turbulence model, and the setup of appropriate analysis algorithms.

[0003] However, the flow around an airfoil is so complex that it's difficult to determine ideal analysis settings a priori. This necessitates iterative simulations to find the appropriate settings, a time-consuming and costly process. Furthermore, settings deemed appropriate under certain conditions can become ineffective when flow conditions or geometry change, requiring the entire process to be re-implemented. Furthermore, analysis results can vary significantly depending on the engineer's subjective judgment, negatively impacting the reliability of the calculations.

[0004] Therefore, a new approach is needed to improve the accuracy and efficiency of airfoil flow analysis, while reducing time and costs. This necessitates a technological solution that can address various issues arising during the flow analysis process and consistently produce reliable results.

[0005] The present invention aims to provide a computational fluid dynamics (CFD) automation method for optimal analysis of airfoil flow based on reinforcement learning, which can obtain highly accurate analysis results cost-effectively by optimizing and automating the entire process of analyzing the flow around an airfoil.

[0006] In addition, the embodiment aims to provide an automated CFD airfoil flow analysis method and a CFD airfoil flow analysis device through reinforcement learning.

[0007] However, the problems to be solved by the embodiments of the present invention are not limited to the problems described above and can be expanded in various ways within the scope of the technical ideas included in the present invention.

[0008] A computational fluid dynamics (CFD) airfoil flow analysis method according to one embodiment includes: an analysis automation model generation step of inputting a blade shape and flow conditions, determining analysis settings for airfoil flow analysis, performing an airfoil flow analysis through a computational fluid dynamics (CFD) simulation based on the determined analysis setting parameters, evaluating the results of the performed airfoil flow analysis and training an artificial neural network through reinforcement learning to satisfy a preset evaluation standard, thereby generating a model that predicts a CFD flow analysis result according to the input analysis conditions; an analysis condition input step of inputting shape conditions and flow conditions of an arbitrary blade; and an analysis result output step of applying the analysis automation model to the input shape conditions and flow conditions of the blade to determine analysis setting parameters, performing a CFD simulation, and outputting a predicted CFD flow analysis result.

[0009] The above-described analysis automation model generation step may include selecting an analysis setting for the airfoil flow analysis based on the input blade shape and flow conditions, training the selected analysis setting with a deep reinforcement learning algorithm using an artificial neural network to satisfy a preset quality standard, generating an analysis setting model that determines analysis setting parameters according to the input analysis conditions, performing the airfoil flow analysis with the CFD simulation, and evaluating the result of the performed airfoil flow analysis to satisfy the evaluation standard by training with a deep reinforcement learning algorithm using an artificial neural network, thereby generating an analysis performance model that predicts the CFD flow analysis result.

[0010] Analysis settings for the above airfoil flow analysis may include generating a computational grid and selecting a turbulence model and analysis algorithm.

[0011] The shape of the above blade is expressed by the blade shape parameters (Blade Shape Parameters, BSP) and the pitch of the blade, and the flow conditions are the isentropic Reynolds number and Mach number Re at the exit, respectively. is,out Wow Ma is,out , and the inflow angle θ in can be expressed in a dimensionless manner.

[0012] Creating the above interpretation setting model is done by defining the state s according to the following equation,

[0013] ,

[0014] Action a is defined according to the following equation:

[0015] ,

[0016] It may include training using a single-step actor-critic algorithm, generating a grid using grid parameters obtained by determining actions based on a given state for each episode, evaluating the quality of the generated grid and calculating a reward.

[0017] Here, BSP is the blade shape parameter, pitch is the blade spacing, and N is the blade spacing. o is the total number of nodes in the O-type lattice, β le is the degree of node clustering in front, β te is the degree of node clustering at the rear edge, Δn1 represents the height of the first cell in the direction perpendicular to the blade surface, and y in is the vertical inlet position, y out is the exit position in the vertical direction, α camber is the degree of curvature of the periodic boundary along the camber line, x o in is the horizontal HO-type interface position, x o out is the position of the OH-type interface in the horizontal direction, and δ represents the ratio of the number of nodes of the O-type grid in the tangential and vertical directions to the blade surface.

[0018] The above compensation is defined based on the quality of the grid that minimizes numerical errors due to geometrical defects of the generated grid, and the quality of the grid is determined by the determinant ratio Q of the Jacobian matrix. J And cell distortion Q S The grid quality indicator can be used to calculate the defined rewards and evaluate them.

[0019] The evaluation criteria for evaluating the results of the above-described performed airfoil flow analysis may include a compensation function for evaluating the accuracy and efficiency of the calculation of the CFD simulation.

[0020] The accuracy of the above calculation can be calculated through the error between the results of the performed airfoil flow analysis and high-quality data including experimental data or high-resolution numerical data, and the efficiency of the above calculation can be calculated by measuring the total time required to perform the airfoil flow analysis.

[0021] To create the above interpretation execution model, state s is defined according to the following equation,

[0022] ,

[0023] Action a is defined according to the following equation:

[0024] ,

[0025] It may include performing actions based on a given state for each episode to complete a CFD simulation, calculating the reward from the CFD simulation results, and training using an actor-critic algorithm.

[0026] Here, BSP is the blade shape parameter, pitch is the blade spacing, and Re is the blade spacing. is,out is the isentropic Reynolds number at the exit, Ma is,out is the Mach number at the exit, θ in represents the inflow angle, w represents the weight, and N o is the total number of nodes in the O-type lattice, β le is the degree of node clustering in front, β te is the degree of node clustering at the rear edge, and Δn1 represents the height of the first cell in the direction perpendicular to the blade surface.

[0027] Generating the above interpretation performance model may include using multi-agent reinforcement learning to have multiple agents perform CFD simulations in parallel to collect training data.

[0028] According to one embodiment, a computational fluid dynamics airfoil flow analysis device includes: an analysis automation model generation unit that inputs a blade shape and flow conditions, determines analysis settings for airfoil flow analysis, performs airfoil flow analysis through a computational fluid dynamics (CFD) simulation based on the determined analysis setting parameters, evaluates the results of the performed airfoil flow analysis, trains an artificial neural network through reinforcement learning to satisfy a preset evaluation standard, and generates a model that predicts a CFD flow analysis result according to the input analysis conditions; and an airfoil flow analysis unit that receives shape conditions and flow conditions of an arbitrary blade, applies the analysis automation model to the input shape conditions and flow conditions of the blade, determines analysis setting parameters, performs a CFD simulation, and predicts and outputs a flow field as a CFD flow analysis result.

[0029] The above-described analysis automation model generation unit may include an analysis setting model that selects an analysis setting for the airfoil flow analysis based on the input blade shape and flow conditions, trains the selected analysis setting using a deep reinforcement learning algorithm using an artificial neural network to satisfy a preset quality standard, and determines an analysis setting parameter according to the input analysis conditions, and an analysis execution model that performs the airfoil flow analysis using the CFD simulation, evaluates the result of the performed airfoil flow analysis, and trains the deep reinforcement learning algorithm using an artificial neural network to satisfy the evaluation standard, and predicts the result of the CFD flow analysis.

[0030] A computational fluid dynamics automation method for optimal analysis of airfoil flow according to one embodiment is a method for automating airfoil flow analysis, comprising: an analysis condition input step for receiving a blade shape and flow conditions, an analysis setting determination step for determining analysis settings for the airfoil flow analysis, an analysis performance step for performing airfoil flow analysis by performing a computational fluid dynamics (CFD) simulation based on the determined analysis setting parameters, an analysis result evaluation step for evaluating the results of the performed airfoil flow analysis, and a reinforcement learning step for training the results of the airfoil flow analysis through reinforcement learning using an artificial neural network so that they satisfy a preset evaluation criterion.

[0031] The above computational fluid dynamics automation method may further include a reinforcement learning step of selecting analysis settings for the airfoil flow analysis based on the input blade shape and flow conditions, and training the selected analysis settings using a deep reinforcement learning algorithm using an artificial neural network to satisfy preset quality standards.

[0032] Analysis settings for the above airfoil flow analysis may include generating a computational grid and selecting a turbulence model and analysis algorithm.

[0033] The evaluation criteria for evaluating the results of the above-described performed airfoil flow analysis may include a compensation function for evaluating the accuracy and efficiency of the calculation of the CFD simulation.

[0034] The accuracy of the above calculation can be calculated through the error between the results of the performed airfoil flow analysis and high-quality data including experimental data or high-resolution numerical data, and the efficiency of the above calculation can be calculated by measuring the total time required to perform the airfoil flow analysis.

[0035] Generating the above interpretation performance model may include using multi-agent reinforcement learning to have multiple agents perform CFD simulations in parallel to collect training data.

[0036] The above reinforcement learning step may include calculating using a deep reinforcement learning algorithm based on artificial neural network technology, including a Deep Q-Network, Proximal Policy Optimization (PPO), or Actor-Critic algorithm.

[0037] According to one embodiment, a computer device includes at least one processor configured to execute computer-readable instructions, wherein the at least one processor inputs a shape and flow conditions of a blade, determines analysis settings for airfoil flow analysis, performs airfoil flow analysis through computational fluid dynamics (CFD) simulation based on the determined analysis setting parameters, evaluates the results of the performed airfoil flow analysis, trains the results through reinforcement learning using an artificial neural network to satisfy preset evaluation criteria, and generates a model for predicting CFD flow analysis results according to the input analysis conditions, inputs shape conditions and flow conditions of an arbitrary blade, and applies the analysis automation model to the shape conditions and flow conditions of the input blade to determine analysis setting parameters, and performs CFD simulation to output a predicted CFD flow analysis result.

[0038] According to the automated computational fluid dynamics method for optimal analysis of airfoil flow based on reinforcement learning according to an embodiment, the entire computational fluid dynamics process for optimization of airfoil flow analysis can be automated.

[0039] This will enable technological innovation in fluid machinery design by improving the accuracy and efficiency of analysis, saving time and cost, and increasing the reliability of calculation results.

[0040] In addition, according to the automated CFD flow analysis method and CFD flow analysis device according to the embodiment, it is possible to derive optimal analysis results with improved analysis accuracy and efficiency and reduced time and cost.

[0041] In addition, by minimizing the intervention of engineers in the analysis of airfoil flow, the entry barrier to analysis can be lowered and the development of related research can be promoted.

[0042] Figure 1 is a flowchart illustrating an automated computational fluid dynamics method for optimal analysis of airfoil flow based on reinforcement learning according to one embodiment.

[0043] FIG. 2 is a schematic diagram illustrating a computational fluid dynamics automation method for optimal analysis of airfoil flow based on reinforcement learning according to one embodiment.

[0044] Figure 3 is a drawing illustrating a conceptual design of a blade parameterization method.

[0045] Figure 4 is a diagram schematically illustrating the process of generating a calculation grid by setting grid parameters required by a grid generator.

[0046] Figure 5 is a schematic diagram illustrating a grid generation algorithm.

[0047] Figure 6 is a computational configuration for simulating flow through a blade passage.

[0048] Figure 7 is a graph showing the time per iteration for 1000 randomly generated calculation configurations according to the number of cells.

[0049] Figure 8 is a diagram showing the grid used to evaluate grid convergence.

[0050] Figure 9 is a schematic diagram illustrating a weighted Chebyshev method for determining a Pareto front using two objective functions (a normalized error function and a normalized cost function).

[0051] Figure 10 is a schematic diagram of multi-agent reinforcement learning for automated mesh generation to achieve optimal computational fluid dynamics (CFD) simulations.

[0052] Fig. 11 is a schematic diagram illustrating a CFD airfoil flow analysis device according to one embodiment.

[0053] Figure 12 is a flowchart illustrating a CFD airfoil flow analysis method according to one embodiment.

[0054] Figure 13 is a diagram showing the results of reinforcement learning of a computational fluid dynamics airfoil flow analysis method according to one embodiment.

[0055] Figure 14 is a diagram showing the optimal CFD airfoil flow analysis result automatically obtained after learning is completed in a computational fluid dynamics airfoil flow analysis method according to one embodiment.

[0056] Figure 15 is an example diagram showing a neural network model according to one embodiment.

[0057] FIG. 16 is a diagram illustrating an example of a computer device according to one embodiment.

[0058] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings so that those skilled in the art can easily practice the present invention. In the drawings, parts irrelevant to the description have been omitted to clearly explain the present invention, and the same reference numerals have been used throughout the specification to refer to the same or similar components. In addition, some components in the attached drawings are exaggerated, omitted, or schematically depicted, and the size of each component does not entirely reflect the actual size.

[0059] The attached drawings are only intended to facilitate understanding of the embodiments disclosed in this specification, and the technical ideas disclosed in this specification are not limited by the attached drawings, and should be understood to include all modifications, equivalents, or substitutes included in the spirit and technical scope of the present invention.

[0060] Terms that include ordinal numbers, such as first, second, etc., may be used to describe various components, but the components are not limited by these terms. These terms are used solely to distinguish one component from another.

[0061] Throughout the specification, terms such as "comprises" or "has" should be understood to indicate the presence of a feature, number, step, operation, component, part, or combination thereof described in the specification, but not to exclude in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof. Therefore, when a part is said to "comprise" a certain component, this does not mean that other components are excluded, but rather that other components may be included, unless specifically stated otherwise.

[0062] Additionally, throughout the specification, when we say "connected," this does not only mean that two or more components are directly connected, but also that two or more components are indirectly connected through other components, that they are electrically connected as well as physically connected, or that they are referred to by different names depending on location or function but are one.

[0063] Additionally, terms such as “……part”, “……unit”, and “……module” described in the specification mean a unit that processes at least one function or operation, which may be implemented by hardware, software, or a combination of hardware and software.

[0064] In this specification, “transmitting or providing” may include not only direct transmission or providing, but also indirect transmission or providing via another device or by using a bypass route.

[0065] In this specification, expressions described in the singular may be interpreted as singular or plural, unless explicit expressions such as “one” or “single” are used.

[0066] FIG. 1 is a flowchart illustrating an automated computational fluid dynamics method for optimal analysis of airfoil flow based on reinforcement learning according to one embodiment, and FIG. 2 is a schematic diagram illustrating an automated computational fluid dynamics method for optimal analysis of airfoil flow based on reinforcement learning according to one embodiment.

[0067] Referring to FIGS. 1 and 2, a computational fluid dynamics (CFD) automation method according to the present embodiment includes steps of inputting analysis conditions, determining analysis settings, and evaluating analysis results. The step of inputting analysis conditions may include a process of inputting a blade shape and flow conditions (S110). The step of determining analysis settings includes a process of determining settings for airfoil flow analysis (S120), and may include generating a grid, selecting a turbulence model, and selecting an analysis algorithm. In the step of determining analysis settings, learning or training for optimizing analysis settings may be performed by applying a reinforcement learning approach (S130).

[0068] The step of evaluating the analysis results may include performing an airfoil flow analysis using CFD analysis (S140) and evaluating the results of the airfoil flow analysis (S150). In the analysis result evaluation step, learning or training may be performed to optimize the analysis results by applying a reinforcement learning approach (S160).

[0069] Reinforcement learning is the process of exploring actions that maximize rewards in a given state. For example, a state can be defined as an input interpretation condition, an action as an interpretation setting parameter, and a reward as the accuracy and efficiency of the interpretation result. Reinforcement learning can be performed using deep reinforcement learning algorithms that incorporate artificial neural network technology. In this example, the actor-critic algorithm was used, but other algorithms such as Deep Q-Network and Proximal Policy Optimization (PPO) can also be used.

[0070] After completing the learning process, optimal analysis parameters can be automatically determined based on any newly given analysis conditions, thereby automating the entire CFD process for airfoil flow analysis. Optimal parameters can be determined non-iteratively, without human intervention, and the optimal grid resolution can be identified.

[0071] Below, the automated computational fluid dynamics method for optimal analysis of airfoil flow is described in more detail for each step.

[0072] First, in the interpretation condition input step, the blade shape and flow conditions are input.

[0073] A blade parametrization method can be used to generate various blade types in axial gas turbines, supersonic impulse turbines, axial compressors, etc. Using blade shape parameters represented by BSP, a two-dimensional blade profile with continuous curvature can be generated from non-uniform rational basis spline curves as shown in the following mathematical expression 1.

[0074] [Mathematical Formula 1]

[0075]

[0076] Figure 3 is a drawing illustrating a conceptual design of a blade parameterization method.

[0077] Referring to (a) of Fig. 3, a camber line is initially established using the first six elements of the BSP. A chord line of length C is determined by the stagger angle ψ at the leading edge. The camber line is then determined by the metal angle θ at the leading and trailing edges. le Wow θ te It consists of a curvature defined by where the tangent ratio of each edge is d le Wow d te is controlled by

[0078] The remaining elements of the BSP define the upper and lower profiles of the blade along the camber lines as shown in Fig. 3(b). The radius of curvature of the leading and trailing edges is ρ le and ρ te is determined by the parameter ( )and ( ) defines the thickness distribution of the upper and lower sections of the blade, where k represents the number of parameters of the thickness distribution. In this example, k = 6 is chosen because it has been found to be sufficient to accurately represent various turbine blade profiles while meeting the precision criteria required for blade manufacturing for axial gas turbines.

[0079] The range of each element of the BSP can be set sufficiently wide to include known blades of various shapes and purposes. The range is as shown in the following mathematical expression 2.

[0080] [Equation 2]

[0081]

[0082] Within this range, unrealistic blade shapes that are simply unconnected or have more than two poles are excluded.

[0083] The blade shape is captured by BSP and pitch (spacing of blades), and the flow conditions are the isentropic Reynolds number and Mach number at the outlet, Re, respectively. is,out Wow Ma is,out , and the inflow angle θ in can be expressed in a non-dimensionalized way. Non-dimensionalization is achieved by boundary conditions (p t,in , T t,in , and p out ) and the axial chord length C of the blade ax It can be achieved using the following mathematical formula 3.

[0084] [Equation 3]

[0085] ,

[0086]

[0087] Here, ρ is,out = p out / (RT is,out ) is based on the ideal gas law, and μ is,out can be calculated using Sutherland's law, , and This nondimensionalization not only accommodates simulations of various geometric and physical scales, but also allows for four variables (p t,in , T t,in , p out and C ax ) with two variables (R eis,out and Ma is,out ) can be integrated to reduce the data required for learning.

[0088] The range of flow conditions is defined broadly enough to include complex flow characteristics such as shock and separation, as shown in the following mathematical expression 4.

[0089] [Equation 4]

[0090]

[0091] Here θ in is adaptively adjusted according to the shape of the blade, and the leading edge metal angle θ is adjusted to alleviate unrealistic flow conditions. le is defined as a relative value to .

[0092] In the analysis settings decision step, a computational grid suitable for the analysis is created and a turbulence model and analysis algorithm are selected.

[0093] Figure 4 is a diagram schematically illustrating the process of generating a calculation grid by setting grid parameters required by a grid generator.

[0094] Referring to Figure 4, a grid generator trained using deep reinforcement learning can generate a structured grid for a two-dimensional blade passage using an elliptical grid generation method. An HOH-type grid, combining H-shaped grids at the inlet and outlet sides and O-shaped grids near the blades, can be generated by specifying grid parameters. The parameters are described in Table 1.

[0095] Boundary structure description in Vertical entrance location y out Exit position α in the vertical direction camber Degree of curvature of periodic boundaries along the camber line x o in Horizontal HO-type interface position x o out Horizontal OH-type interface positioning control description N t Number of nodes N of the O-shaped grid in the tangential direction to the blade surface n Number of nodes of the O-shaped grid perpendicular to the blade surface β le Degree of node clustering in front β te Degree of clustering of nodes on the back edgeΔn1 Height of the first cell perpendicular to the blade surface

[0096] Figure 5 is a schematic diagram illustrating a grid generation algorithm.

[0097] Given a blade shape with a pitch, the boundary shape of the grid can be defined based on the following parameters: y in , y out , α camber , x o in , x o out . Here y in and y out determine the inlet and outlet locations in the vertical direction, respectively. To accommodate grid generation for blades with different curvatures, the curvature of the periodic boundary is α camber is controlled by , which is a straight line (α camber = 0) to a shape that completely matches the camber line (α) camber = 1) ranges from 0 to 1. The interface positions between the O-grid and H-grid on the inlet and outlet sides are x, respectively. o in and x o out is determined by

[0098] After that, the grid resolution is N t , N n , β le , β te and Δn1 are determined based on . Initially, the number of nodes of the O-shaped grid in the tangential and normal directions to the blade surface is N, respectively. t Wow N n is designated as N along the blade surface. t The nodes can be distributed to increase resolution at the leading and trailing edges. This distribution follows a hyperbolic tangent function, β le and β te Each controls the cluster strength towards the front and rear edges. N in the direction perpendicular to the blade surface. nNodes are distributed and clustered along a hyperbolic tangent function on the blade surface to properly interpret near-wall physics. This cluster can be defined such that the height of the first cell from the blade surface corresponds to Δn1. After the nodes are distributed within the domain, the grid is finally generated using an elliptical grid generation method to ensure high orthogonality near the blade surface and no slope discontinuities.

[0099] The goal of the first phase of training is to enable the grid generator to autonomously select the portion of the geometry that controls it from the full set of grid parameters (see Table 1) for various blade passage configurations. To achieve this, state s can be defined as follows:

[0100] [Equation 5]

[0101]

[0102] BSP represents the blade shape parameter specified in mathematical expression 1, pitch means blade spacing, and N o = N t × N n represents the total number of nodes in the O-shaped lattice. N o , β le , β te , Δn1 is included in the state rather than being left as an optimization action to provide adaptability of the grid resolution under various flow conditions. These parameters can be optimized in subsequent training steps, including CFD simulations.

[0103] Action a, which includes grid parameters that control geometry, can be defined as in the following mathematical expression (6).

[0104] [Equation 6]

[0105]

[0106] Here, δ is the state variable and the total grid resolution N is predetermined. oTo assign δ = N to the normal and tangential directions t / N n can be integrated into . That is, δ represents the ratio of the number of nodes of the O-shaped grid in the tangential and normal directions to the blade surface.

[0107] Compensation can be defined based on grid quality, aiming to minimize numerical errors due to geometric imperfections in the grid. To this end, two grid quality metrics are considered: the determinant ratio Q of the Jacobian matrix; J And cell distortion Q S It is. Q J Wow Q S These metrics are necessary for achieving uniform cell distribution and maintaining high orthogonality between adjacent cells. Because of these properties, these metrics can be used to generate high-quality rectangular grids. Q J The calculation is as follows: Equation 7.

[0108] [Equation 7]

[0109]

[0110] Here J i,j is for each node (x i,j ,y i,j ) can be calculated as the determinant of the Jacobian matrix in the following mathematical expression 8.

[0111] [Equation 8]

[0112]

[0113] 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 area consistency between adjacent cells, with higher values ​​indicating smoother resolution transitions. Skew Q S can be calculated as shown in the following mathematical formula 9.

[0114] [Equation 9]

[0115]

[0116] θmin and θ max represents the minimum and maximum values ​​of the internal angles of the cell. Q S A higher value indicates reduced cell distortion and improved orthogonality with neighboring cells. The compensation r is defined as in the following mathematical expression (10).

[0117] [Equation 10]

[0118]

[0119] Here, ( )| min and ( )| avg represent the minimum and average values ​​of the cells of the O-grid, respectively. The reason for focusing on the O-grid is that, unlike the cells of the H-grid, which generally maintain a rectangular shape, the cells of the O-grid are likely to deform to fit the blade profile. Q J By including additional nodes at the boundaries of the O-grid for computation, smooth transitions at the grid interfaces and periodic boundaries can be ensured. By considering both minimum and average values, the reward function can be designed to effectively evaluate the impact of the lowest quality cells and provide a comprehensive view of the overall quality distribution of the grid. (Q J )| min , (Q J )| avg , (Q S )| min , and (Q S )| avg By definition, the values ​​of metrics range from 0 to 1, indicating a comparable scale. The reason for using a square term in the reward function is to increase sensitivity to high values.

[0120] The first stage of training is performed using a single-step actor-critic algorithm. In each episode, a state s is randomly assigned. Based on this state, the actor learns a = π with a search noise ε. φ(s) determines an action. The grid generator then generates a grid using the grid parameters obtained from the given state and the selected action. The quality of the generated grid is then evaluated and the reward r is calculated. The data (s, a, r) ​​are stored in a buffer. Using the data in the buffer, the critic network is updated to more accurately predict the reward, and the actor network is improved to generate actions that maximize the reward. Then, the next episode begins. This process is repeated until the network converges. Since learning is performed on a randomly assigned blade configuration in each episode, rather than sequentially training on a single configuration and then moving on to the next, the converged network can generate optimal grids for a defined range of configurations.

[0121] The actor and critic are structured as fully connected networks. 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 constrain the action values. 10 -4 Using the Adam optimizer with a learning rate of N for mini-batch size b We update the network parameters by setting to 100, which is a common approach in actor-critic algorithms. To increase the stability of the learning process, the actor network is updated once every two updates of the critic network. The search noise ε is normally distributed N(0, σ) with mean 0 and standard deviation σ. 2 ) is generated, and σ is defined as in the following mathematical expression 11.

[0122] [Equation 11]

[0123]

[0124] In the initial episodes, a higher σ is used to collect diverse data sets. Subsequently, the cosine function is used to systematically vary the noise level to balance exploration and exploitation, and to reduce the influence of specific noise values. Both networks have four hidden layers, each with 512, 256, 256, and 128 neurons, respectively.

[0125] As the number of episodes increases, the network is updated to generate higher geometric quality grids for each newly introduced blade passage configuration. Consequently, high-compensation data is accumulated in the buffer, resulting in J π The value increases and eventually converges. Approximately 106 episodes are required to sufficiently train the network for the current optimization problem. The trained grid generator in this stage is utilized in the subsequent training stage, optimal CFD training, which integrates CFD simulations to optimize resolution-related grid parameters.

[0126] Meanwhile, in this embodiment, selecting a turbulence model includes selecting the type and parameters of the turbulence model. For example, the turbulence model can be selected from among k-epsilon, k-omega, and k-omega SST. For the selection of the turbulence model, the analysis setting decision unit (50) can be optimized through deep reinforcement learning.

[0127] In addition, in this embodiment, the selection of the analysis algorithm may include the selection of a numerical technique, a matrix calculation algorithm, and the selection of a time-step size and a residual criterion. For example, the numerical technique may be selected from First-order upwind or Second-order central differencing, and the matrix calculation algorithm may be selected from Flexible Generalized Minimum Residual, Conjugate Gradient, and Bi-Conjugate Gradient Stabilized. For the selection of the analysis algorithm, the analysis setting decision unit (50) may be optimized through deep reinforcement learning.

[0128]

[0129] In the analysis result evaluation step, the airfoil flow analysis is performed using computational fluid dynamics (CFD) analysis with the determined analysis settings, and then the analysis results are evaluated.

[0130] For blade passage simulation, the open-source software SU2 can be used to solve the two-dimensional steady-state compressible Reynolds Averaged Navier Stokes (RANS) equations. These equations are expressed as follows: Equation 12.

[0131] [Equation 12]

[0132]

[0133] F c Wow F v represent convective flow and viscous flow, respectively, and are expressed as in the following mathematical equation 13.

[0134] [Equation 13]

[0135]

[0136] where ρ is the fluid density, v is the velocity vector, E is the total energy per unit mass, is the unit matrix, p is the static pressure, κ is the thermal conductivity, T is the temperature, and is the viscous stress tensor defined as in the following mathematical expression 14.

[0137] [Equation 14]

[0138]

[0139] Here, μ is the viscosity.

[0140] For turbulent flows, based on the Boussinesq hypothesis, the viscosity μ is the dynamic viscosity μ d and turbulent viscosity μ t Divided by μ = μ d + μ t is expressed as . Similarly, thermal conductivity is κ = κ d + κ t is expressed as , where κ is the viscosity μ and constant pressure specific heat c p , and using the Prandtl number Pr, κ = (c p μ) / Pr is calculated as the dynamic viscosity μ d is calculated using Sutherland's law. Turbulent viscosity μ t is obtained from a turbulence model. In this embodiment, the shear stress transport-k-ω (SST-k-ω) turbulence model can be used. This adopts the k-ω model near the wall and switches to the k-model in the free stream, and can perform accurate predictions in both the near-wall region and the core region by using a blending function.

[0141] To complete Equation 12, the ideal gas equation of state using the specific heat ratio γ and the gas constant R is applied. The pressure is determined by p = (γ - 1)ρ[E - 0.5(v · v)], the temperature is given by T = p / (ρR), and the specific heat at constant pressure is c p= (γR) / (γ-1). The governing equations are discretized using the finite volume method (FVM). The convective flux is discretized using the central Jameson-Schmidt-Turkel (JST) technique. The gradients of the flow variables required to evaluate the viscosity flux are calculated using the weighted least squares method.

[0142] Figure 6 is a computational configuration for simulating flow through a blade passage.

[0143] Referring to Figure 6, the total pressure p at the inlet t,in and total temperature T t,in The inflow angle θ in is imposed with. At the exit, the static pressure p out This applies. Periodic conditions are assigned to the upper and lower boundaries, which are spaced apart by the pitch distance. No-slip and insulating wall boundary conditions are assigned to the blade surface.

[0144] When evaluating analysis results, both computational accuracy and computational efficiency can be considered. Both computational accuracy and efficiency can be applied simultaneously to objective functions through the concept of Pareto optimality in multi-objective optimization. Another approach is to adjust the weights between the two objective functions. Computational accuracy can be calculated by comparing the error between high-quality data, such as experimental data or high-resolution numerical data, and the current analysis results. Computational efficiency can be calculated by measuring the total time required to perform the analysis.

[0145] The reward function for evaluating the interpretation results can be designed to simultaneously consider both the accuracy and efficiency of the simulation. To this end, two objective functions are used in a multi-objective approach: the cost and error of the simulation. First, r c The cost function, denoted as , is designed to reflect the total time required for one simulation and can be calculated as shown in the following mathematical expression 15.

[0146] [Equation 15]

[0147]

[0148] Here N total represents the total number of cells. This formula takes O(N) time per iteration. total ) and this can be confirmed through simulations for 1000 random flow configurations as shown in Fig. 7. This linear scaling is due to the sparse matrix operations used in the current iterative solver, which can utilize the flexible generalized minimum residual (FGMRES) method enhanced with an incomplete lower upper (ILU) preprocessor.

[0149] An error function can be designed to evaluate grid convergence by calculating the difference between simulation results obtained using two successively refined grids (a base grid and a fine grid). The base grid initially generated by the DRL network can be further refined to create a fine grid by doubling the resolution along each axis, as shown in Fig. 8(a). Cubic interpolation can be used to preserve the blade curvature, particularly at the leading and trailing edges, as shown in Fig. 8(b). The error function can be defined as the root-mean-square (RMS) error calculated at the node locations of the base grid, which remain constant during the refinement process, as shown in Equation 16 below.

[0150] [Equation 16]

[0151]

[0152] Here, the error of Ma can be chosen to capture the flow characteristics essential for blade passage analysis, such as boundary layers, wakes, and shock waves. Its nondimensional nature also facilitates scaling across a variety of flow configurations. The error for the entire flow field is evaluated as the first term, while the error along the blade surface is included as the second term to reflect the near-wall flow physics. For the second term, the isentropic Mach number Ma is is p in mathematical formula 3 out Instead, it is calculated by substituting the pressure along the blade surface.

[0153] Finally, using the weighted Chebyshev method, which guarantees that all Pareto optimal solutions can be identified for both convex and non-convex problems, the reward function r can be defined as follows:

[0154] [Equation 17]

[0155]

[0156] Here, w represents the weight between the two objective functions, and can be scaled as shown in the following mathematical expression (18).

[0157] [Equation 18]

[0158]

[0159] Log scaling can be used to mitigate the effects of outliers during the learning process. Factor 10 2 to r e Multiply by r c By dividing by 107, we can achieve better learning performance by matching the scale between the two objective functions.

[0160] A schematic illustration of the weighted Chebyshev method for determining the Pareto front is shown in Figure 9. For a given weight w, rectangles with fixed aspect ratios and variable sizes are constructed on the plane of the two objective functions f and g. As indicated by the minus sign in Equation (17), the learning process aims to find the rectangle with the minimum size, which leads to the Pareto optimal solution. By varying the value of w from 0 to 1, all Pareto front solutions can be obtained.

[0161] To mitigate the high computational costs associated with data acquisition through computational fluid dynamics (CFD) simulations, data replication methods can be applied to increase both the amount and variety of data. In a single simulation, r e Wow r c Once obtained, a reward function can be calculated by applying various weights to these values. For example, 100 data sets can be replicated by randomly changing the weight w between 0 and 1 for each episode.

[0162] 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 as in the following mathematical expression 19.

[0163] [Equation 19]

[0164]

[0165] The state s is formulated as follows, Equation 20, to reflect the calculation configuration including blade geometry and flow conditions.

[0166] [Equation 20]

[0167]

[0168] Here, the variable w is included as a state variable with a range from 0 to 1, rather than a fixed value, to determine the weight between two objectives: cost and simulation error. This approach allows users to obtain an optimal grid that balances accuracy and efficiency according to their preferences.

[0169] Multi-agent reinforcement learning can be used to accelerate the data acquisition required for training grid generators for optimal CFD. As shown in Figure 10, multiple agents can perform CFD simulations in parallel to collect data.

[0170] Blade shape, indicated by BSP and pitch, Re is,out , Ma is,out , θ in A state consisting of a flow condition represented by and a weight w is randomly assigned to each agent. Each agent takes an action a = π φ (s) and the search noise ε are used to generate a grid, and then a CFD simulation is performed. When agent i completes the CFD simulation, the error function r is derived from the results. e,i and cost function r c,i is calculated, and the reward r i is calculated. After that, the number of data (s) is calculated using the data reproduction method. i ,a i ,r i )1,...,(s i ,a i ,r i ) 100 The data is amplified and stored in a buffer. The stored data is used to update the actor and critic networks. Then, a new task is assigned to the corresponding agent, and the next episode begins. Tasks are dynamically reallocated as soon as the agent completes the task, so the data acquisition rate increases proportionally to the number of agents. This process is repeated until the network converges.

[0171] To train the network sufficiently, a total of 5 × 10 5 A conference RANS simulation is required. Training involves 256 agents over eight weeks, each running on a single CPU core of an Intel(R) Xeon(R) E5-2650 v2 processor.

[0172] By applying the computational fluid dynamics automation method shown in the above drawing 1, a CFD airfoil flow analysis device can be implemented, and CFD airfoil flow analysis can be performed using this.

[0173] Fig. 11 is a schematic diagram illustrating a CFD airfoil flow analysis device according to one embodiment, and Fig. 12 is a flowchart illustrating a CFD airfoil flow analysis method according to one embodiment.

[0174] Referring to FIG. 11, a CFD airfoil flow analysis device (100) according to one embodiment may include an analysis automation model generation unit (130) and an airfoil flow analysis unit (150). Referring to FIG. 12, a CFD airfoil flow analysis method according to one embodiment may include an analysis automation model generation step (S10), an analysis condition input step (S20), and an analysis result output step (S30). The analysis automation model generation unit (130) may perform the analysis automation model generation step, and the airfoil flow analysis unit (150) may perform the analysis condition input step (S20) and the analysis result output step (S30).

[0175] The interpretation automation model generation unit (130) can generate an interpretation setting model (132) that determines interpretation setting parameters according to the input interpretation conditions, and can generate an interpretation execution model (136) that predicts CFD flow analysis results according to the input interpretation conditions.

[0176] The analysis setting model (132) selects analysis settings for airfoil flow analysis based on the input analysis conditions, i.e., the blade shape and flow conditions, and can be trained using a deep reinforcement learning algorithm using an artificial neural network so that the selected analysis settings satisfy preset quality standards. The analysis setting model (132) can receive the blade shape and flow conditions as inputs and determine analysis setting parameters for airfoil flow analysis.

[0177] Analysis settings for airfoil flow analysis may include generating a computational grid, selecting a turbulence model, and selecting an analysis algorithm. Generating the computational grid may include generating a structured grid for a two-dimensional blade passage using an elliptical grid generation method. Selecting the turbulence model may include selecting a type and parameters of the turbulence model. The type of the turbulence model may be selected from k-epsilon, k-omega, and k-omega SST. Selecting the analysis algorithm may include selecting a numerical technique, selecting a matrix calculation algorithm, and selecting a time-step size and a residual criterion.

[0178] The shape of the above blade can be expressed by the blade shape parameters (Blade Shape Parameters, BSP) and the pitch of the blade according to mathematical equation 1. The flow conditions are the isentropic Reynolds number and Mach number Re at the outlet, respectively. is,out Wow Ma is,out , and the inflow angle θ in It can include the Reynolds number and the Mach number Re is,out Wow Ma is,out , and the inflow angle θ in can be expressed in a non-dimensional manner according to mathematical formula 3.

[0179] Creating an interpretation setting model may include defining states s and actions a, determining actions based on given states for each episode, generating a grid using the obtained grid parameters, evaluating the quality of the generated grid and calculating a reward, and training using a single-step actor-critic algorithm. Here, state s may be defined according to Equation 5, and action a may be defined according to Equation 6.

[0180] The above compensation can be defined based on the quality of the grid that minimizes numerical errors due to geometrical defects of the generated grid. Specifically, the quality of the grid is determined by the determinant ratio Q of the Jacobian matrix. J And cell distortion Q S The grid quality index can be evaluated by calculating the defined compensation r, where the compensation r can be defined according to mathematical expression 10.

[0181] The analysis execution model (136) performs airfoil flow analysis by CFD simulation based on the above-determined analysis setting parameters, and can be trained using a deep reinforcement learning algorithm using an artificial neural network to evaluate the results of the performed airfoil flow analysis and satisfy the preset evaluation criteria. The analysis execution model (136) can predict the results of the CFD flow analysis according to the input analysis conditions based on the above-determined analysis settings.

[0182] The evaluation criteria for evaluating the results of the airfoil flow analysis performed above may include a compensation function for evaluating the accuracy and efficiency of the calculation of the CFD simulation. The accuracy of the calculation may be calculated through the error between high-quality data including experimental data or high-resolution numerical data and the results of the airfoil flow analysis performed above, and the efficiency of the calculation may be calculated by measuring the total time required to perform the airfoil flow analysis.

[0183] In order to evaluate the accuracy of the calculation, an error function can be calculated and evaluated, and the error function can be defined according to mathematical formula 16. In order to evaluate the efficiency of the calculation, a cost function can be calculated and evaluated, and the cost function can be defined according to mathematical formula 15. In order to consider both the accuracy and efficiency of the calculation, a reward function r defined according to mathematical formula 17 can be calculated and used.

[0184] Creating the above interpretation performance model may include defining states s and actions a, performing actions based on the given states for each episode to complete a CFD simulation, calculating the reward from the CFD simulation results, and training using an actor-critic algorithm. Here, state s may be defined according to Equation 20, and action a may be defined according to Equation 19.

[0185] Creating the above-described interpretation performance model may involve using multi-agent reinforcement learning to have multiple agents perform CFD simulations in parallel to collect training data. This can accelerate the data acquisition speed required to train a grid generator for optimal CFD.

[0186] The airfoil flow analysis unit (150) can receive shape conditions and flow conditions of an arbitrary blade, apply the analysis automation model to the input shape conditions and flow conditions of the blade, determine analysis setting parameters, perform CFD simulation, and predict and output CFD flow analysis results. A flow field can be predicted and output by predicting the CFD flow analysis results.

[0187] As described above, once training is completed in the analysis automation model generation unit (130), the mesh generator in the airfoil flow analysis unit (150) can generate a mesh that provides a converged solution within a desired computational cost in a single simulation for any blade shape and flow conditions. That is, the optimal analysis settings are automatically determined, thereby eliminating the need for an iterative CFD process generally required for grid convergence. In addition, the airfoil flow analysis unit (150) can predict optimal CFD airfoil flow analysis results with improved accuracy and efficiency based on the automatically determined optimal analysis settings.

[0188] Figure 13 is a diagram showing the results of reinforcement learning of a computational fluid dynamics airfoil flow analysis method according to one embodiment.

[0189] Reynolds number (Re) 5.90×10 for blade shape for high pressure turbine 5 , Mach number (Ma) 0.84, angle of attack (α) 0.0 degrees, the flow field root mean square error (RMSE) was minimized while considering the efficiency of calculation. As a result, the root mean square error (RMSE) of the flow variables at the wall surface presented in the upper part was 3.9×10 -3 9.87×10 in -4 It was observed that the error at the trailing edge was significantly reduced as shown in the center, and the root mean square error (RMSE) of the entire flow field was 1.007×10 -2 3.734×10 in -3 It can be confirmed that the number of steps has been reduced. In fact, it has been confirmed that the physical phenomena of the wake part are better captured after learning as presented in the lower part.

[0190] Figure 14 is a diagram showing the optimal CFD airfoil flow analysis result automatically obtained after learning is completed in a computational fluid dynamics airfoil flow analysis method according to one embodiment.

[0191] The left side shows the analysis results under flow conditions where shock waves occur in the shape of a high-pressure turbine blade. Typically, shock waves cause rapid changes in the gradient of the flow field, leading to high nonlinearity. Therefore, even experienced engineers must perform iterative analysis processes to obtain accurate analysis results. However, a trained neural network can predict the location of shock wave occurrence in advance, and with a single analysis, the root mean square error (RMSE) at the wall is 1.482×10 -3 , root mean square error (RMSE) in the entire flow field is 4.909×10 -3 Accurate results were obtained. The right side shows the analysis results for the shape of a low-pressure turbine blade. Because the curvature of this blade is much greater than that of a high-pressure turbine blade, generating a computational grid is very difficult. Nevertheless, the trained neural network not only automatically generates the grid but also achieves a root mean square error (RMSE) of 1.339×10 at the wall without repetitive analysis. -3 , root mean square error (RMSE) in the entire flow field is 3.289×10 -3 A high level of accuracy was achieved.

[0192] FIG. 15 is an exemplary diagram showing a neural network model (800) according to one embodiment. The neural network model (800) is an example of a machine learning model, and in machine learning technology and cognitive science, is a statistical learning algorithm implemented based on the structure of a biological neural network or a structure that executes the algorithm.

[0193] According to one embodiment, the neural network model (800) may represent a machine learning model having problem-solving capabilities by learning that nodes, which are artificial neurons that form a network by combining synapses like in a biological neural network, repeatedly adjust the weights of synapses so that the error between the correct output corresponding to a specific input and the inferred output is reduced. For example, the neural network model (800) may include any probability model, neural network model, etc. used in artificial intelligence learning methods such as machine learning and deep learning.

[0194] The neural network model (800) is implemented as a multilayer perceptron (MLP) composed of multiple layers of nodes and connections therebetween. The artificial neural network model (800) according to one embodiment may be implemented using one of various artificial neural network model structures including MLP. As illustrated in FIG. 15, the artificial neural network model (800) is composed of an input layer (820) into which the shape of the blade and the 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) which receive signals from the input layer (820), extract characteristics, and transmit them to the output layer (840) (where n is a positive integer). Here, the output layer (840) receives signals from the hidden layers (830_1 to 830_n) and outputs them to the outside.

[0195] When the machine learning model is learned, the analysis automation model generation unit (130) of the computational fluid dynamics airfoil flow analysis device (100) according to the embodiment can store the learned analysis setting model (132) and analysis execution model (136). Here, the learned analysis setting model (132) and analysis execution model (136) can be stored in the memory of the airfoil flow analysis unit (150). Alternatively, the learned analysis setting model (132) and analysis execution model (136) can be stored in the memory of a server connected to the airfoil flow analysis unit (150) via a wired or wireless network.

[0196] The wing shape flow analysis unit (150) applies the input blade shape conditions and flow conditions to the prediction model generated by the analysis automation model generation unit (130) and outputs the predicted CFD flow analysis results.

[0197] FIG. 16 is a diagram illustrating an example of a computer device according to one embodiment. The analysis automation model generation unit (130) and the airfoil flow analysis unit (150) described with reference to FIG. 11 can be implemented by the computer device (900) illustrated in FIG. 16.

[0198] The computer device (900) may include a memory (910), a processor (920), a communication interface (930), and an input / output interface (940). The memory (910) may be a computer-readable storage medium, and may include a random access memory (RAM), a read only memory (ROM), and a permanent mass storage device such as a disk drive. In addition, an operating system and at least one program code may be stored in the memory (910). These software components may be loaded into the memory (910) from a computer-readable storage medium separate from the memory (910). This separate computer-readable storage medium may include a computer-readable storage medium such as a hard disk, a flash memory, an optical disk, an external hard disk, etc. In addition, these software components may be loaded into the memory (910) via the communication interface (930).

[0199] The processor (920) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (920) via memory (910) or a communication interface (930).

[0200] The communication interface (930) may provide a function for the computer device (900) to communicate with other devices via a network (1000).

[0201] The input / output interface (940) may be a means for interfacing with an input / output device (950). For example, the input device may include a device such as a microphone, a keyboard, or a mouse, and the output device may include a device such as a display or a speaker.

[0202] The embodiments described above may be implemented in the form of a computer program that can be executed through various components on a computer, and such a computer program may be recorded on a computer-readable medium. At this time, the medium may include a magnetic medium such as a hard disk, a floppy disk, and a magnetic tape, an optical recording medium such as a CD-ROM and a DVD, a magneto-optical medium such as a floptical disk, and a hardware device specifically configured to store and execute program instructions, such as a ROM, a RAM, a flash memory, etc.

[0203] Unless the steps constituting a method according to one embodiment are explicitly described in a specific order or contrary to the description, the steps may be performed in any appropriate order. The present invention is not necessarily limited to the order in which the steps are described. The use of all examples or exemplary terms (e.g., "etc.") herein is merely intended to illustrate the present invention in detail and does not limit the scope of the present invention. Furthermore, those skilled in the art will appreciate that various modifications, combinations, and variations can be made within the scope of the claims or their equivalents.

[0204] 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 a person having ordinary skill in the art to which the present invention pertains also fall within the scope of the present invention.

Claims

1. An analysis automation model generation step in which the shape and flow conditions of a blade are input, analysis settings for airfoil flow analysis are determined, and airfoil flow analysis is performed through computational fluid dynamics (CFD) simulation based on the determined analysis setting parameters, and the results of the performed airfoil flow analysis are evaluated to satisfy a preset evaluation standard, and a model is generated that predicts the results of the CFD flow analysis according to the input analysis conditions by training through reinforcement learning using an artificial neural network; Analysis condition input step for inputting shape conditions and flow conditions of any blade; and An analysis result output step that determines analysis setting parameters by applying the above analysis automation model to the shape conditions and flow conditions of the blade input above, performs CFD simulation, and outputs predicted CFD flow analysis results. A computational fluid dynamics airfoil flow analysis method including .

2. In paragraph 1, The above interpretation automation model creation step is: Based on the shape and flow conditions of the input blade, an analysis setting for the airfoil flow analysis is selected, and the selected analysis setting is trained using a deep reinforcement learning algorithm using an artificial neural network so that it satisfies a preset quality standard, thereby generating an analysis setting model that determines analysis setting parameters according to the input analysis conditions. The airfoil flow analysis is performed using the above CFD simulation, and the results of the performed airfoil flow analysis are evaluated to satisfy the above evaluation criteria, and a deep reinforcement learning algorithm using an artificial neural network is trained to create an analysis performance model that predicts the CFD flow analysis results. A computational fluid dynamics airfoil flow analysis method including the following.

3. In paragraph 2, The analysis setting for the above airfoil flow analysis is a computational fluid dynamics airfoil flow analysis method including generating a computational grid and selecting a turbulence model and analysis algorithm.

4. In paragraph 3, The shape of the above blade is expressed by the blade shape parameters (Blade Shape Parameters, BSP) and the pitch of the blade. A computational fluid dynamics airfoil flow analysis method in which the above flow conditions are expressed in a non-dimensional manner by the isentropic Reynolds number, Mach number, and inlet angle at the outlet, respectively.

5. In paragraph 3, Creating the above interpretation setting model is done by defining the state s according to the following equation, Action a is defined according to the following equation: A computational fluid dynamics airfoil flow analysis method, comprising: generating a grid using grid parameters obtained by determining actions based on a given state for each episode; evaluating the quality of the generated grid and calculating a reward; and training using a single-step actor-critic algorithm. Here, BSP is the blade shape parameter, pitch is the blade spacing, and N is the blade spacing. o is the total number of nodes in the O-type lattice, β le is the degree of node clustering in front, β te is the degree of node clustering in the rear, Δn1 represents the height of the first cell in the direction perpendicular to the blade surface, y in is the vertical inlet position, y out is the exit position in the vertical direction, α camber is the degree of curvature of the periodic boundary along the camber line, x o in is the horizontal HO-type interface position, x o out is the position of the OH-type interface in the horizontal direction, and δ represents the ratio of the number of nodes of the O-type grid in the tangential and vertical directions to the blade surface.

6. In paragraph 5, The above compensation is defined based on the quality of the grid that minimizes numerical errors due to geometrical defects of the generated grid, The quality of the above grid is determined by the determinant ratio Q of the Jacobian matrix. J And cell distortion Q S A computational fluid dynamics (CFD) airfoil flow analysis method, which is evaluated by calculating the defined compensation using the grid quality index.

7. In paragraph 2, A computational fluid dynamics airfoil flow analysis method, wherein the evaluation criteria for evaluating the results of the above-described airfoil flow analysis include a compensation function for evaluating the accuracy and efficiency of the calculation of the CFD simulation.

8. In paragraph 7, A computational fluid dynamics airfoil flow analysis method, wherein the accuracy of the above calculation is calculated through an error between high-quality data including experimental data or high-resolution numerical data and the results of the performed airfoil flow analysis, and the efficiency of the above calculation is calculated by measuring the total time required to perform the airfoil flow analysis.

9. In paragraph 8, To create the above interpretation execution model, state s is defined according to the following equation, Action a is defined according to the following equation: A computational fluid dynamics airfoil flow analysis method, comprising: performing an action based on a given state for each episode to complete a CFD simulation, calculating the reward from the CFD simulation results, and training using an actor-critic algorithm. Here, BSP is the blade shape parameter, pitch is the blade spacing, and Re is the blade spacing. is,out is the isentropic Reynolds number at the exit, Ma is,out is the Mach number at the exit, θ in represents the inflow angle, w represents the weight, and N o is the total number of nodes in the O-type lattice, β le is the degree of node clustering in front, β te is the degree of node clustering at the rear edge, and Δn1 represents the height of the first cell in the direction perpendicular to the blade surface.

10. In paragraph 2, A computational fluid dynamics airfoil flow analysis method, wherein generating the above analysis performance model includes collecting learning data by having multiple agents perform CFD simulations in parallel using multi-agent reinforcement learning.

11. An analysis automation model generation unit that inputs the shape and flow conditions of a blade, determines analysis settings for airfoil flow analysis, performs airfoil flow analysis through Computational Fluid Dynamics (CFD) simulation based on the determined analysis setting parameters, evaluates the results of the performed airfoil flow analysis, and trains an artificial neural network through reinforcement learning to satisfy a preset evaluation standard, thereby generating a model that predicts the CFD flow analysis results according to the input analysis conditions; and An airfoil flow analysis unit that receives shape conditions and flow conditions of an arbitrary blade, applies the analysis automation model to the shape conditions and flow conditions of the input blade, determines analysis setting parameters, performs CFD simulation, and predicts and outputs a flow field as a CFD flow analysis result. A computational fluid dynamics airfoil flow analysis device including a .

12. In paragraph 11, The above interpretation automation model generation unit is, An analysis setting model that selects an analysis setting for the airfoil flow analysis based on the shape and flow conditions of the input blade, and trains the selected analysis setting using a deep reinforcement learning algorithm using an artificial neural network so that the selected analysis setting satisfies a preset quality standard, thereby determining an analysis setting parameter according to the input analysis conditions; and An analysis model that performs the airfoil flow analysis using the above CFD simulation, evaluates the results of the performed airfoil flow analysis, and trains with a deep reinforcement learning algorithm using an artificial neural network to satisfy the above evaluation criteria, thereby predicting the results of the CFD flow analysis. A computational fluid dynamics airfoil flow analysis device including:

13. In a method for automating the analysis of airfoil flow, Analysis condition input step where the shape and flow conditions of the blade are input; An analysis setting determination step for determining analysis settings for the above airfoil flow analysis; An analysis execution step of performing airfoil flow analysis by performing computational fluid dynamics (CFD) simulation based on the above-determined analysis setting parameters; An analysis result evaluation step for evaluating the results of the above-described airfoil flow analysis; and A reinforcement learning step that trains the results of the above airfoil flow analysis to satisfy pre-set evaluation criteria through reinforcement learning using an artificial neural network. An automated computational fluid dynamics method for optimal analysis of airfoil flow including .

14. In paragraph 13, A reinforcement learning step in which analysis settings for the airfoil flow analysis are selected based on the shape and flow conditions of the blade input above, and a deep reinforcement learning algorithm using an artificial neural network is used to train the selected analysis settings so that they satisfy preset quality standards. An automated computational fluid dynamics method for optimal analysis of airfoil flow including more.

15. In paragraph 14, An automated computational fluid dynamics method for optimal analysis of airfoil flow, comprising: generating a computational grid, selecting a turbulence model and an analysis algorithm; 16. In paragraph 13, A computational fluid dynamics automated method for optimal analysis of airfoil flow, wherein the evaluation criteria for evaluating the results of the performed airfoil flow analysis include a compensation function for evaluating the accuracy and efficiency of the calculation of the CFD simulation.

17. In paragraph 16, A computational fluid dynamics automated method for optimal analysis of airfoil flow, wherein the accuracy of the above calculation is calculated through an error between high-quality data including experimental data or high-resolution numerical data and the result of the performed airfoil flow analysis, and the efficiency of the above calculation is calculated by measuring the total time taken to perform the airfoil flow analysis.

18. In paragraph 13, A computational fluid dynamics automated method for optimal analysis of airfoil flow, wherein generating the above analysis performance model includes collecting learning data by having multiple agents perform CFD simulations in parallel using multi-agent reinforcement learning.

19. In paragraph 13, A computational fluid dynamics automation method for optimal analysis of airfoil flow, comprising calculating using a deep reinforcement learning algorithm based on artificial neural network technology including a deep Q-Network, a proximal policy optimization (PPO), or an actor-critic algorithm, wherein the above reinforcement learning step is performed.

20. Comprising at least one processor implemented to execute computer-readable instructions, By at least one processor, Inputting the blade shape and flow conditions, determining the analysis settings for airfoil flow analysis, performing airfoil flow analysis through Computational Fluid Dynamics (CFD) simulation based on the determined analysis setting parameters, evaluating the results of the performed airfoil flow analysis, and training through reinforcement learning using an artificial neural network to satisfy a preset evaluation standard, thereby generating a model that predicts the CFD flow analysis results according to the input analysis conditions. Enter the shape conditions and flow conditions of any blade, and A computer device that applies the above-mentioned analysis automation model to the shape conditions and flow conditions of the input blade to determine analysis setting parameters, performs CFD simulation, and outputs predicted CFD flow analysis results.

Images