Impeller optimization design method and system based on flow field simulation

Abstract

The invention discloses an impeller optimization design method and system based on flow field simulation, and belongs to the technical field of computer data processing and artificial intelligence, and the method comprises the steps: obtaining initial geometric parameters, carrying out flow field simulation, extracting key flow structure features in a flow field, generating an intervention strategy and a virtual control field through a neural controller, and carrying out the simulation of the flow field. And finally, an optimized geometric parameter of the impeller is generated by using a reverse mapping network, and performance verification is carried out. According to the method, the neural controller trained by the flow field simulation and reinforcement learning method and the reverse mapping network trained based on the supervised learning method are adopted, the complex impeller optimization process can be converted into an intelligent iteration problem, the design efficiency and the impeller performance are improved, and the design period is shortened.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and artificial intelligence technology, and in particular to a method and system for impeller optimization design based on flow field simulation. Background Technology

[0002] Impellers are core energy exchange components in fluid machinery such as pumps, compressors, and turbines. The quality of their three-dimensional geometry directly determines the performance, efficiency, and reliability of the entire machine. With the development of computer technology, impeller design has become heavily reliant on computer-aided engineering, especially numerical simulations using computational fluid dynamics. This process is essentially complex scientific calculation and data processing; the computer solves the governing equations of fluid motion based on the input impeller geometry data and outputs flow field data describing the fluid's motion state, thereby evaluating the design performance.

[0003] In existing technologies, computer-based impeller optimization design methods typically follow an iterative data processing flow. The designer first defines a set of impeller geometric parameters as initial input data. The computer then executes a flow field simulation program to process this input and generate performance result data. Next, the designer or an automated optimization algorithm adjusts the initial geometric parameter data based on the performance result data and initiates a new round of simulation calculations. Even with the introduction of artificial intelligence algorithms, their role is largely limited to performing a more efficient search within the geometric parameter space to find a set of geometric parameters that produces better performance results more quickly; the basic data processing flow remains unchanged.

[0004] However, the data processing flow employed in existing technologies has inherent limitations. First, the flow is computationally inefficient because every small adjustment to the input geometric parameters must trigger a complete and time-consuming flow field simulation calculation, resulting in a lengthy optimization cycle. Second, this optimization method is an indirect control approach; the algorithm cannot directly intervene in and correct undesirable flow structures appearing in the flow field data during the simulation calculation process. It can only indirectly affect the flow field by modifying the surrounding geometric data, making it difficult to solve problems accurately and efficiently. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an impeller optimization design method and system based on flow field simulation. It employs a novel data processing architecture that first directly intervenes in and optimizes the flow field data within the simulation data domain, and then reverse maps the optimized flow field data to generate geometric parameters. This approach can improve design efficiency and fundamentally realize intelligent design driven by the target flow field.

[0006] The above objectives can be achieved through the following approach: A method for impeller optimization design based on flow field simulation includes: acquiring initial impeller geometric parameters and performing flow field simulation on the impeller to obtain initial flow field data reflecting the distribution of velocity, pressure, turbulence, and vortices in the flow field; analyzing the initial flow field data using a model for identifying flow field structures to extract feature vectors of key flow structures in the flow field; inputting the feature vectors into a preset neural controller to generate an intervention strategy and obtaining a virtual control field based on the intervention strategy; adding the virtual control field as a source term to the flow field simulation process for iterative calculation to obtain an optimized flow field state; establishing an inverse mapping network for the mapping relationship between the optimized flow field state and the virtual control field to the geometric shape, and generating final impeller geometric parameters based on the optimized flow field state and the virtual control field; constructing a three-dimensional impeller model based on the final impeller geometric parameters, and verifying the performance of the three-dimensional impeller model.

[0007] Optionally, the method further includes: acquiring historical flow field data containing flow field data and flow field features, and forming a historical flow field simulation dataset with the initial impeller geometry parameters; training a model and inverse mapping network for identifying flow field structures using a supervised learning method based on the historical flow field simulation dataset; and training a neural controller for generating intervention strategies using a reinforcement learning method based on the historical flow field simulation dataset.

[0008] Optionally, the step of analyzing the initial flow field data using a model for identifying the flow field structure and extracting feature vectors of key flow structures in the flow field includes: processing the initial flow field data into flow field cloud map data; extracting features from the flow field cloud map data using a convolutional neural network to obtain digital features characterizing eddies and flow separation; and combining the digital features to generate feature vectors of key flow structures.

[0009] Optionally, inputting the feature vector into a preset neural controller to generate an intervention strategy and obtaining a virtual control field based on the intervention strategy includes: calculating the optimal intervention strategy based on the feature vector of the key flow structure and the preset neural controller; and generating a virtual control field as a momentum source term and an energy source term based on the optimal intervention strategy.

[0010] Optionally, adding the virtual control field as a source term to the flow field simulation process for iterative calculation to obtain the optimized flow field state includes: superimposing the virtual control field in the flow field simulation and performing iterative simulation calculations, monitoring the flow field variables in real time; stopping the calculation and outputting the optimized flow field state when the flow field variables reach a stable state.

[0011] Optionally, generating the final impeller geometry parameters based on the optimized flow field state and the virtual control field includes: merging the optimized flow field state and the virtual control field to generate joint input data; processing the joint input data through a generative neural network to generate a candidate set of impeller geometry parameters; and selecting a set with optimized performance from the candidate set of impeller geometry parameters as the final impeller geometry parameters.

[0012] Optionally, constructing a three-dimensional impeller model based on the final impeller geometry parameters includes: calculating and generating three-dimensional point cloud data of the impeller based on the final impeller geometry parameters; performing a surface reconstruction operation on the three-dimensional point cloud data of the impeller to generate a smooth surface model, and outputting the smooth surface model as the final three-dimensional impeller model.

[0013] Optionally, the performance verification of the impeller 3D model includes: performing high-precision flow field simulation on the final impeller 3D model to obtain performance index data; comparing the performance index data with a preset target value; and when the performance index data does not reach the preset target value, using the final impeller 3D model as a new initial impeller geometric parameter, iteratively executing until the performance index data reaches the preset target value.

[0014] Based on the same inventive concept, this invention also provides an impeller optimization design system based on flow field simulation. The system includes: a flow field data acquisition and simulation module for acquiring initial impeller geometric parameters and performing flow field simulation on the impeller to obtain initial flow field data reflecting the distribution of velocity, pressure, turbulence, and vortices in the flow field; a flow field feature extraction module for analyzing the initial flow field data using a model for identifying flow field structures and extracting feature vectors of key flow structures in the flow field; an intervention strategy generation module for inputting the feature vectors into a preset neural controller to generate an intervention strategy and obtain a virtual control field based on the intervention strategy; a flow field optimization calculation module for adding the virtual control field as a source term to the flow field simulation process for iterative calculation to obtain an optimized flow field state; a geometric parameter generation module for establishing an inverse mapping network between the optimized flow field state and the virtual control field to geometric shapes, and generating final impeller geometric parameters based on the optimized flow field state and the virtual control field; and a three-dimensional model construction and performance verification module for constructing a three-dimensional impeller model based on the final impeller geometric parameters and verifying the performance of the three-dimensional impeller model.

[0015] Compared with the prior art, the present invention has the following advantages: 1. This invention improves the efficiency and quality of design data generation by constructing a novel data processing architecture that deeply embeds artificial intelligence models into the design process. By directly intervening in flow field data within the simulation data domain, it avoids the time-consuming process of repeatedly modifying geometric input data and restarting the entire calculation, as in traditional methods. This transforms the originally discrete, serial optimization iteration into continuous, efficient online data processing, thereby achieving a fundamental improvement in computational resource consumption and time cost.

[0016] 2. The data processing method provided by this invention changes the underlying logic of computer-aided design, realizing the reverse generation from target effect data to causal geometric data. By first actively shaping an ideal flow field data state, and then using a reverse mapping network to deduce the geometric design data that can generate that state, the design paradigm is transformed from the traditional "forward verification" to "reverse generation," making the design process more goal-oriented and fundamentally ensuring that the final generated design data can meet performance expectations.

[0017] 3. This invention enhances the interpretability of AI decision-making during the design process. By generating and utilizing a virtual control field as an intermediate data product, not only is the optimization task completed, but also an analyzable and understandable physical insight data is provided. Visual analysis of the virtual control field data allows for understanding in which regions AI exerts what kind of influence to improve the flow field, making the previously black-box intelligent optimization process transparent and enhancing the efficiency and trustworthiness of human-machine collaborative design.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an impeller optimization design method based on flow field simulation according to an embodiment of the present invention.

[0021] Figure 2 These are comparison diagrams showing the improvement of key flow field characteristics in embodiments of the present invention.

[0022] Figure 3This is a comparison chart of impeller performance curves in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of an impeller optimization design system based on flow field simulation according to an embodiment of the present invention.

[0024] Figure 5 This is a comparison diagram before and after flow field feature diagnosis and intervention according to an embodiment of the present invention.

[0025] Figure 6 This is a performance verification data diagram of the final impeller model in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 One embodiment of this invention proposes an impeller optimization design method based on flow field simulation. It employs a novel data processing architecture that first directly intervenes in and optimizes the flow field data within the simulation data domain, and then reverse-maps the optimized flow field data to generate geometric parameters. This improves design efficiency and fundamentally enables intelligent design driven by the target flow field. This method transforms the originally discrete, serial optimization iteration into continuous, efficient online data processing, achieving improvements in computational resource consumption and time costs. Furthermore, it changes the underlying logic of computer-aided design by first actively shaping an ideal flow field data state and then reverse-deriving the geometric design data that can generate that state.

[0028] The method described in this embodiment specifically includes: The initial impeller geometry parameters are obtained, and the flow field is simulated on the impeller to obtain initial flow field data reflecting the distribution of velocity, pressure, turbulence and vortex in the flow field. Specifically, initial impeller geometry parameters are obtained using 3D modeling software; these parameters serve as input data for flow field simulation. The impeller is then numerically simulated using fluid dynamics simulation software to solve the Navier-Stokes equations. This process yields initial flow field data, including velocity, pressure, turbulence, and vortex distribution at various locations within the flow field.

[0029] The initial flow field data is analyzed using a model for identifying flow field structures, and feature vectors of key flow structures in the flow field are extracted. Specifically, the initial flow field data is processed into visualized flow field contour maps, which are then analyzed using a model for identifying flow field structures. This model can identify undesirable flow structures in the flow field, such as vortices and flow separation occurring on blade surfaces or within flow channels. After identification, these key flow structures are converted into digital features, and these digital features are combined into a feature vector of a key flow structure.

[0030] The feature vector is input into a preset neural controller to generate an intervention strategy, and a virtual control field is obtained based on the intervention strategy. Specifically, the feature vectors of key flow structures are input into a pre-set neural controller. This controller, after pre-training, can automatically generate an optimal intervention strategy based on the input feature vectors to resolve undesirable flow problems in the flow field. Based on this intervention strategy, a virtual control field is obtained. This virtual control field is used to directly intervene in the flow field data, rather than modifying the impeller's geometric parameters.

[0031] The virtual control field is added as a source term to the flow field simulation process, and iterative calculations are performed to obtain the optimized flow field state. Specifically, a virtual control field, such as momentum and energy sources, is added to the flow field simulation equations. This allows the flow field simulation to be adjusted in real time based on the influence of the virtual control field during the solution process. The simulation will perform iterative calculations and monitor the flow field variables in real time. When the flow field variables reach a steady state, the calculation will stop and the optimized flow field state will be output. This process transforms the original flow field data processing into an online, controllable iterative calculation.

[0032] Establish an inverse mapping network for the mapping relationship between the optimized flow field state and the virtual control field to the geometric shape, and generate the final impeller geometric parameters based on the optimized flow field state and the virtual control field; Specifically, after optimizing the flow field state, a reverse mapping network is used to input the optimized flow field state and the virtual control field. The reverse mapping network processes this data and derives the impeller geometry parameters that generate the optimized flow field state. This process is the core of this method; it transforms the original forward design process "from parameters to flow field" into a reverse generation process "from flow field to parameters," thus directly driving the design with performance objectives.

[0033] A three-dimensional model of the impeller is constructed based on the final impeller geometry parameters, and the performance of the three-dimensional impeller model is verified.

[0034] Specifically, using the final impeller geometry parameters, a completely new 3D impeller model is reconstructed in 3D modeling software. Then, high-precision flow field simulations are performed on this new model to verify whether its performance indicators, such as efficiency and head, have met the expected design goals. This verification process is a necessary step to ensure that the generated design scheme is physically feasible.

[0035] By adopting a method that involves acquiring initial flow field data, extracting key flow field features, generating a virtual control field through a neural controller, adding the virtual control field as a source term to the flow field simulation to obtain the optimized flow field state, and generating the final impeller geometry parameters through an inverse mapping network, the complex impeller optimization process can be transformed into an intelligent iterative problem, improving design efficiency and impeller performance, and shortening the design cycle.

[0036] Optionally, the method further includes: Acquire historical flow field data containing flow field data and flow field characteristics, and combine it with the initial impeller geometry parameters to form a historical flow field simulation dataset; Specifically, through large-scale flow field simulation, a large amount of flow field data and flow field characteristics under different impeller geometry parameters are obtained, thereby constructing a historical flow field simulation dataset.

[0037] Based on the historical flow field simulation dataset, a supervised learning method was used to train a model and an inverse mapping network for identifying flow field structures. Specifically, when training a model for recognizing flow field structures, a deep convolutional neural network (DCNN) architecture can be employed. The input to the DCNN model is flow field contour data, and the output is a vector representing the flow field features. The training process minimizes the loss function between the predicted and true features. The model parameters are optimized using numbers, and the loss function is expressed as follows: , in, It is the number of training samples. It is the first The true feature vector of each sample It is the first The model predicts the feature vector for each sample.

[0038] Based on the historical flow field simulation dataset, a neural controller for generating intervention strategies is trained using reinforcement learning.

[0039] Specifically, the neural controller used to generate the intervention strategy can be trained using the Deep Deterministic Policy Gradient (DDPG) algorithm. During this training process, the agent receives the current flow field state. As input, and according to its policy π, output the virtual control field action. The environment depends on and Give the agent a reward signal and return to the next state. The reward function is designed as follows: , in, These are weighting coefficients. This represents the decrease in the total energy of the flow field. This represents the squared decrease in vortex intensity in the flow field. This represents the reduction in the flow separation region. The agent maximizes the cumulative reward. To optimize the strategy, among which It is a discount factor, indicating the importance of future rewards.

[0040] Optionally, the step of analyzing the initial flow field data using a model for identifying flow field structures and extracting feature vectors of key flow structures in the flow field includes: The initial flow field data is processed into flow field cloud map data; Specifically, after completing the flow field simulation calculations, the acquired initial flow field data, such as velocity, pressure, and turbulent kinetic energy distributed on the grid nodes, are rendered into two-dimensional or three-dimensional flow field contour maps using post-processing software. These contour maps represent flow field variables of different values ​​with different colors or grayscale levels, intuitively reflecting the distribution of the flow field. For example... Figure 2 As shown, the method of the present invention intuitively demonstrates the significant improvement effect of the method on key undesirable flow structures in the impeller flow field.

[0041] The flow field cloud map data is used to extract features by a convolutional neural network to obtain digital features that characterize eddies and flow separation. Specifically, a pre-trained convolutional neural network (CNN) is used to extract features from the flow field contour data. The CNN contains multiple convolutional and pooling layers, enabling it to automatically learn and recognize visual patterns associated with eddies and flow separation in the flow field contour maps. For example, in eddy regions, streamlines form closed or spiral patterns, and pressure and velocity gradients exhibit specific distributions. The CNN captures these patterns through its filters and transforms them into a series of digital features that characterize information such as the intensity, size, and location of eddies and flow separation.

[0042] The digital features are combined to generate feature vectors for key flow structures.

[0043] Specifically, multiple numerical features extracted by the convolutional neural network, such as numerical values ​​representing eddy current intensity and the area of ​​the flow separation region, are arranged in a predetermined order and combined into a feature vector of a key flow structure. This feature vector of the key flow structure can comprehensively describe the key flow structures in the flow field in a compact, machine-readable format.

[0044] Optionally, the feature vector is input to a preset neural controller to generate an intervention strategy, and a virtual control field is obtained based on the intervention strategy, including: The optimal intervention strategy is calculated based on the feature vector of the key flow structure and the preset neural controller. Specifically, the feature vectors of the key flow structures are used as input to a pre-defined neural controller. The neural controller can employ a trained deep neural network, which calculates and outputs an optimal intervention strategy based on the input feature vectors of the key flow structures. The intervention strategy includes information such as the direction, magnitude, and location of the virtual force applied in the flow field, aiming to maximize the improvement of impeller performance.

[0045] Based on the optimal intervention strategy, a virtual control field is generated as a momentum source and an energy source.

[0046] Specifically, based on the optimal intervention strategy, the virtual control field is materialized in the flow field simulation mesh. The virtual control field includes virtual force vectors as momentum sources and virtual heat or pressure gradients as energy sources. These sources are precisely applied to specific regions of the flow field, such as adding a reverse virtual force at the flow separation point on the back of a blade, or applying a virtual torque to the vortex core region in the flow channel. These sources are additional inputs to the flow field simulation, used to directly and accurately intervene in and correct fluid flow.

[0047] Optionally, the virtual control field is added as a source term to the flow field simulation process for iterative calculation to obtain the optimized flow field state, including: The virtual control field is superimposed in the flow field simulation, and iterative simulation calculations are performed to monitor flow field variables in real time. Specifically, in fluid dynamics simulations, a virtual control field is added as a source term to the momentum and energy equations. The momentum source term can be expressed as... The energy source term can be represented as These source terms are incorporated into the discrete form of the Navier-Stokes equations, allowing the simulator to update the flow field variables at each iteration step based on the influence of the virtual control field. The simulation will monitor flow field variables such as velocity, pressure, turbulence, and temperature in real time.

[0048] When the flow field variables reach a steady state, the calculation stops and the optimized flow field state is output.

[0049] Specifically, convergence criteria are set to determine whether the flow field variables have reached a steady state. Convergence criteria can include the following aspects: Residual convergence: The discrete form of the Navier-Stokes equations produces residuals at each iteration step. When the residual value... Less than the preset convergence threshold At that time, the simulation is considered to have converged, that is... Stable physical quantities at monitoring points: Monitoring points are set at key locations. When the fluctuations of the physical quantities at these points within a certain number of consecutive iterations are less than a preset threshold... At that time, it is considered that a stable state has been reached, that is... ,in It represents any physical quantity. This is the current iteration step. It monitors the iteration step size. When any convergence criterion is met, the calculation stops, and the optimized flow field state data at that point is output. For example... Figure 3 As shown, the performance curves of the impeller before and after optimization are compared, intuitively demonstrating the improvement in isentropic efficiency and pressure ratio.

[0050] Optionally, generating the final impeller geometry parameters based on the optimized flow field state and the virtual control field includes: The optimized flow field state and the virtual control state are combined to generate joint input data; Specifically, after the flow field reaches an optimized state, all flow field data in the optimized state are integrated with the corresponding virtual control field data, such as integrating the flow velocity and pressure at each grid node with the applied virtual force vectors. The integrated data is formatted as a high-dimensional joint vector or a three-dimensional tensor, which serves as the input to the subsequent generative neural network. This data merging can simultaneously preserve the physical properties of the flow field and the information on the agent's intervention in the flow field, providing a comprehensive basis for inverse generation.

[0051] The joint input data is processed by a generative neural network to generate a candidate set of impeller geometry parameters; Specifically, a generative neural network is used to process the joint input data. The generative neural network can be an architecture such as a generative adversarial network or a variational autoencoder. The generative neural network can learn the complex mapping relationship from flow field states and control information to impeller geometry. When the joint input data is fed into the network, the network will output one or more sets of parameters representing different impeller geometry designs, thereby generating a candidate set of impeller geometry parameters.

[0052] Select a set of performance-optimized parameters from the candidate set of impeller geometry parameters as the final impeller geometry parameters.

[0053] Specifically, a rapid performance evaluation is performed on each design scheme in the candidate set of impeller geometry parameters, such as using a lightweight flow field surrogate model or a simplified flow field solver. The evaluation will score or rank each candidate scheme according to preset performance indicators. Then, the set of parameters with the highest score or best performance is selected from the candidate set as the final impeller geometry parameters.

[0054] Optionally, constructing a three-dimensional impeller model based on the final impeller geometry parameters includes: Based on the final impeller geometry parameters, calculate and generate three-dimensional point cloud data of the impeller; Specifically, after obtaining the final impeller geometry parameters, a parametric calculation program is used to generate three-dimensional point cloud data of the impeller surface based on the principles of impeller geometry. These parameters include the impeller hub radius, rim radius, blade angle, and blade thickness distribution. Through mathematical equations and algorithms, densely distributed discrete three-dimensional coordinate points on the impeller surface are calculated, and these coordinate points together constitute the impeller's three-dimensional point cloud data.

[0055] Perform a surface reconstruction operation on the impeller 3D point cloud data to generate a smooth surface model, and output the smooth surface model as the final impeller 3D model.

[0056] Specifically, the 3D point cloud data of the impeller is input into specialized surface reconstruction software. Using the Poisson reconstruction algorithm, the discrete point cloud data is transformed into a continuous, smooth surface. This operation fills the gaps between point clouds, corrects potential errors, and generates a mesh model with topological and normal information. This mesh model is the final 3D impeller model, which can be directly used for simulation verification, physical manufacturing, and demonstration.

[0057] Optionally, performance verification of the impeller 3D model includes: High-precision flow field simulation was performed on the final three-dimensional model of the impeller to obtain performance index data; Specifically, after constructing the three-dimensional model of the impeller, high-precision fluid dynamics simulation software is used to perform numerical simulations on the model. This simulation will employ fine mesh generation and advanced turbulence models to obtain accurate flow field data. Through post-processing, the impeller's performance indicators, including but not limited to impeller efficiency, head, and total pressure ratio, are calculated and extracted from these flow field data.

[0058] Compare the performance index data with the preset target value; Specifically, the performance indicators are compared with the preset target values ​​set at the initial design stage. The preset target values ​​are performance benchmarks determined based on actual application requirements and engineering experience, such as a design efficiency of not less than 90% and a head of not less than 50 meters.

[0059] If the performance index data does not reach the preset target value, the final three-dimensional impeller model is used as the new initial impeller geometry parameters, and the process is iterated until the performance index data reaches the preset target value.

[0060] Specifically, when the performance index data fails to reach the preset target value, the geometric parameters corresponding to the impeller 3D model can be used as new initial impeller geometric parameters, and all steps in the method can be re-executed. This iterative loop will continue until the performance index data corresponding to the newly generated impeller geometric parameters reaches or exceeds the preset target value.

[0061] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an impeller optimization design system based on flow field simulation, the system comprising: The flow field data acquisition and simulation module is used to acquire the initial impeller geometric parameters and perform flow field simulation on the impeller to obtain initial flow field data that reflects the distribution of velocity, pressure, turbulence and vortices in the flow field. The flow field feature extraction module is used to analyze the initial flow field data using a model for identifying the flow field structure, and extract feature vectors of key flow structures in the flow field. An intervention strategy generation module is used to input the feature vector into a preset neural controller, generate an intervention strategy, and obtain a virtual control field based on the intervention strategy; The flow field optimization calculation module is used to add the virtual control field as a source term to the flow field simulation process, perform iterative calculations, and obtain the optimized flow field state. The geometric parameter generation module is used to establish an inverse mapping network for the mapping relationship between the optimized flow field state and the virtual control field to the geometric shape, and to generate the final impeller geometric parameters based on the optimized flow field state and the virtual control field. The 3D model construction and performance verification module is used to construct a 3D model of the impeller based on the final impeller geometry parameters and to verify the performance of the 3D model of the impeller.

[0062] To verify the feasibility of this invention in practice, it was applied to an impeller optimization design project for a high-load centrifugal compressor. The project aimed to improve the isentropic efficiency of the existing impeller design and broaden its stable operating range. The main technical bottleneck lay in the difficulty of effectively suppressing vortices and flow separation phenomena caused by complex flow within the blade channels using traditional manual geometry modification methods. Traditional design methods rely on engineers repeatedly modifying the geometry and restarting simulations, resulting in long cycles and optimization effects limited by experience.

[0063] In this embodiment, an initial impeller geometry is first used as a starting point. Flow field simulation calculations are performed using a flow field simulation unit to obtain initial flow field data containing information such as velocity and pressure fields. When this initial design is run under design conditions, its flow field exhibits obvious undesirable flow structures.

[0064] To quantify the diagnosis of this problem, the initial flow field data was processed into a series of flow field contour maps and fed into a pre-trained convolutional neural network model serving as the flow field feature extraction module. After model analysis, a significant flow separation zone was automatically identified in the rear part of the blade's suction surface, and a high-intensity vortex structure was detected in the blade tip clearance region. By integrating the vortex quantity within the vortex core region identified by the model, the vortex intensity was calculated to be 15.8 m. 2 / s, while quantifying the projected area of ​​the flow separation zone on the wall as 120mm². 2 These digital features are combined into a feature vector of a key flow structure, serving as a precise diagnostic report of the current flow field problem. This feature vector is then input into a pre-defined neural controller. The controller's integrated reinforcement learning strategy module instantly calculates the optimal intervention strategy to correct the flow field problem based on the input feature vector. This strategy is then visualized, generating a three-dimensional virtual control field as a momentum source. This virtual control field is precisely applied to the vortex and flow separation region located by the model, its vector direction designed to weaken the rotational energy of the vortices and guide the separated boundary layer to reattach to the wall.

[0065] Subsequently, in a real-time flow field intervention loop, the virtual control field is superimposed as an additional source term onto the Navier-Stokes control equations of the flow field simulation, and iterative simulation calculations are continuously performed. In each iteration step, the current flow field data is acquired in real time, the feature vector is updated through the flow field feature extraction module, and the virtual control field is dynamically adjusted by the neural controller. This loop continues until the rate of change of the total turbulent kinetic energy of the flow field is detected to be less than a preset convergence threshold and stabilizes at a low level. At this point, the simulation stops, and the output flow field is the optimized flow field state.

[0066] To obtain the physical geometry that generates the optimized flow field state, the two three-dimensional data fields—the optimized flow field state and the virtual control field applied during intervention—are merged to form a high-dimensional joint input data set, which is then fed into a pre-defined inverse mapping network. The inverse mapping network is a trained generative neural network that decodes the joint input data and directly generates a set of optimal impeller geometry parameters that accurately describe the three-dimensional morphology of the impeller. Specifically, this is represented by a series of new control point coordinates for the B-spline curve defining the blade profile.

[0067] Based on the newly generated impeller geometry parameters, three-dimensional point cloud data of the impeller surface was calculated and generated through parametric modeling scripts. Then, non-uniform rational B-spline surface reconstruction technology was used to generate a smooth and continuous final three-dimensional impeller model.

[0068] Finally, the final 3D impeller model underwent high-precision performance verification. It was imported into CFD software, and simulations were performed using a finer mesh and a higher-order SST k-ω turbulence model than those used in the optimization process. Simulation results show that the newly designed impeller achieves an isentropic efficiency of 91.5% under design conditions, an improvement over the initial design's 88.0%, with a slight increase in pressure ratio. Furthermore, comparisons revealed that the simulation performance indicators met the preset design target values. The entire optimization design process, from initial simulation to the generation of the final verification model, took approximately 48 hours, a significant improvement in efficiency compared to the weeks-long design cycle of traditional methods.

[0069] Figure 5 and Figure 6 The data documented on the practical application of this invention in the optimization design of centrifugal compressor impellers, and detailed the system’s performance in flow field diagnosis, direct intervention, and final performance verification.

[0070] from Figure 5 As can be seen, the method of this invention can accurately and quantitatively diagnose undesirable structures in the flow field and intervene efficiently and directly by applying a virtual control field. In optimizing the flow field state, the initially existing strong eddies and flow separation phenomena are basically eliminated, the eddy intensity is reduced by 86.7%, and the separation zone area is reduced by more than 95%, which directly leads to a decrease in the total turbulent kinetic energy of the flow field, laying a solid physical foundation for subsequent performance improvement.

[0071] Figure 6 This demonstrates that the present invention can successfully translate flow field optimization into performance gains in solid geometry. Through an inverse mapping network, the impeller geometry capable of achieving an ideal flow field is accurately derived. Final verification results show that the optimized impeller model achieves an isentropic efficiency of 91.5%, representing a 3.5 percentage point improvement and successfully exceeding the design target. More importantly, the entire intelligent design process shortens the traditional iteration cycle of several weeks to 48 hours, significantly improving R&D efficiency and fully demonstrating the technical advantages of this invention in achieving rapid, accurate, and intelligent design of high-performance impellers.

[0072] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0073] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for impeller optimization design based on flow field simulation, characterized in that, The method includes: The initial impeller geometry parameters are obtained, and the flow field is simulated on the impeller to obtain initial flow field data reflecting the distribution of velocity, pressure, turbulence and vortex in the flow field. The initial flow field data is analyzed using a model for identifying flow field structures, and feature vectors of key flow structures in the flow field are extracted. The feature vector is input into a preset neural controller to generate an intervention strategy, and a virtual control field is obtained based on the intervention strategy. The virtual control field is added as a source term to the flow field simulation process, and iterative calculations are performed to obtain the optimized flow field state. Establish an inverse mapping network for the mapping relationship between the optimized flow field state and the virtual control field to the geometric shape, and generate the final impeller geometric parameters based on the optimized flow field state and the virtual control field; A three-dimensional model of the impeller is constructed based on the final impeller geometry parameters, and the performance of the three-dimensional impeller model is verified.

2. The impeller optimization design method based on flow field simulation according to claim 1, characterized in that, The method further includes: Acquire historical flow field data containing flow field data and flow field characteristics, and combine it with the initial impeller geometry parameters to form a historical flow field simulation dataset; Based on the historical flow field simulation dataset, a supervised learning method was used to train a model and an inverse mapping network for identifying flow field structures. Based on the historical flow field simulation dataset, a neural controller for generating intervention strategies is trained using reinforcement learning.

3. The impeller optimization design method based on flow field simulation according to claim 2, characterized in that, The step of analyzing the initial flow field data using a model for identifying flow field structures and extracting feature vectors of key flow structures in the flow field includes: The initial flow field data is processed into flow field cloud map data; The flow field cloud map data is used to extract features by a convolutional neural network to obtain digital features that characterize eddies and flow separation. The digital features are combined to generate feature vectors for key flow structures.

4. The impeller optimization design method based on flow field simulation according to claim 2, characterized in that, The feature vector is input into a preset neural controller to generate an intervention strategy, and a virtual control field is obtained based on the intervention strategy, including: The optimal intervention strategy is calculated based on the feature vector of the key flow structure and the preset neural controller. Based on the optimal intervention strategy, a virtual control field is generated as a momentum source and an energy source.

5. The impeller optimization design method based on flow field simulation according to claim 1, characterized in that, The virtual control field is added as a source term to the flow field simulation process for iterative calculation to obtain the optimized flow field state, including: The virtual control field is superimposed in the flow field simulation, and iterative simulation calculations are performed to monitor flow field variables in real time. When the flow field variables reach a steady state, the calculation stops and the optimized flow field state is output.

6. The impeller optimization design method based on flow field simulation according to claim 1, characterized in that, Based on the optimized flow field state and the virtual control field, the final impeller geometry parameters are generated as follows: The optimized flow field state and the virtual control state are combined to generate joint input data; The joint input data is processed by a generative neural network to generate a candidate set of impeller geometry parameters; Select a set of performance-optimized parameters from the candidate set of impeller geometry parameters as the final impeller geometry parameters.

7. The impeller optimization design method based on flow field simulation according to claim 1, characterized in that, Constructing a 3D model of the impeller based on the final impeller geometry parameters includes: Based on the final impeller geometry parameters, calculate and generate three-dimensional point cloud data of the impeller; Perform a surface reconstruction operation on the impeller 3D point cloud data to generate a smooth surface model, and output the smooth surface model as the final impeller 3D model.

8. The impeller optimization design method based on flow field simulation according to claim 1, characterized in that, The performance verification of the impeller 3D model includes: High-precision flow field simulation was performed on the final three-dimensional model of the impeller to obtain performance index data; Compare the performance index data with the preset target value; If the performance index data does not reach the preset target value, the final three-dimensional impeller model is used as the new initial impeller geometry parameters, and the process is iterated until the performance index data reaches the preset target value.

9. An impeller optimization design system based on flow field simulation, characterized in that, The system is used in the impeller optimization design method based on flow field simulation as described in any one of claims 1-8, the system comprising: The flow field data acquisition and simulation module is used to acquire the initial impeller geometric parameters and perform flow field simulation on the impeller to obtain initial flow field data that reflects the distribution of velocity, pressure, turbulence and vortices in the flow field. The flow field feature extraction module is used to analyze the initial flow field data using a model for identifying the flow field structure, and extract feature vectors of key flow structures in the flow field. An intervention strategy generation module is used to input the feature vector into a preset neural controller, generate an intervention strategy, and obtain a virtual control field based on the intervention strategy; The flow field optimization calculation module is used to add the virtual control field as a source term to the flow field simulation process, perform iterative calculations, and obtain the optimized flow field state. The geometric parameter generation module is used to establish an inverse mapping network for the mapping relationship between the optimized flow field state and the virtual control field to the geometric shape, and to generate the final impeller geometric parameters based on the optimized flow field state and the virtual control field. The 3D model construction and performance verification module is used to construct a 3D model of the impeller based on the final impeller geometry parameters and to verify the performance of the 3D model of the impeller.

Images