Fluid machinery design methods, devices, equipment, storage media, and computer program products
By identifying and removing distortions in the initial geometric parameter population and using a physical field prediction model to predict performance parameters, the problem of model anomalies and mesh distortion caused by distortion in fluid machinery design is solved, thereby improving design efficiency and accuracy.
Patent Information
- Application Number
- CN202611401079.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-10
AI Technical Summary
In existing fluid machinery design, model anomalies and mesh distortion caused by distorted geometric parameters affect design efficiency and contaminate the training of surrogate models, leading to a decrease in prediction accuracy.
By obtaining an initial geometric parameter population, a geometric structure classification model is used to remove distorted individuals, an optimized geometric parameter population is constructed, and a physical field prediction model is used to predict performance parameters. The design values of geometric parameters are then determined by combining optimization rules.
This approach avoids wasting high-cost CFD simulation resources, improves design efficiency and prediction accuracy, and ensures the reliability and accuracy of fluid machinery design.
Smart Images

Figure CN123069987A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of fluid machinery design technology, and in particular to a fluid machinery design method, apparatus, equipment, storage medium, and computer program product. Background Technology
[0002] In fluid machinery design, data-driven and surrogate modeling techniques are being used to replace or accelerate traditional fluid dynamics simulation (CFD). Typical methods include: constructing parametric geometric models, combining CFD to generate performance data samples, and training machine learning models such as neural networks to achieve performance prediction and optimization design.
[0003] Currently, the Latin hypercube sampling method is typically used to perform discrete sampling in a high-dimensional parameter space, then construct a set of parameterized geometric models. A CFD solver is then used to solve these parameterized geometric models to obtain the corresponding simulation data. However, in this process, the geometric parameter combinations obtained from discrete sampling may experience surface self-intersection and Boolean operation failures during geometric model reconstruction, leading to modeling tool malfunctions or crashes, or mesh distortion during the meshing of the constructed geometric model. These anomalies in the geometric modeling and meshing stages directly cause subsequent CFD simulations to fail or produce invalid results, affecting not only overall design efficiency but also contaminating the training of surrogate models, resulting in decreased prediction accuracy. Summary of the Invention
[0004] The main objective of this application is to provide a fluid machinery design method, apparatus, equipment, storage medium, and computer program product, which aims to solve the technical problem that the low quality of training samples in surrogate models leads to a decrease in the prediction accuracy of surrogate models.
[0005] To achieve the above objectives, this application proposes a fluid machinery design method, comprising: Obtain the initial geometric parameter population of the target fluid machinery; The trained geometric structure classification model is used to classify the initial geometric parameter population and remove distorted individuals to obtain an optimized geometric parameter population. Fluid dynamics simulations are performed on some combinations of geometric parameters in the optimized geometric parameter population. Based on the simulation data, a training set sample for the physical field prediction model is constructed, and the physical field prediction model is trained using the training set sample to obtain the trained physical field prediction model. Based on the preset optimization rules, the trained physical field prediction model is invoked to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population. Based on the calculated performance parameter prediction values of each geometric parameter combination, the geometric parameter design values of the target fluid machinery are determined.
[0006] In one embodiment, the population of initial geometric parameters for the target fluid machinery includes: Based on the performance parameters of the target fluid machinery, the core geometric parameters and non-core geometric parameters of the target fluid machinery are divided. Initial geometric parameter samples are generated based on core and non-core geometric parameters.
[0007] In one embodiment, the process of classifying the core geometric parameters and non-core geometric parameters of the target fluid machinery based on its performance parameters includes: Based on the global sensitivity analysis method, the total effect index of each geometric parameter is calculated; The core and non-core geometric parameters of the target fluid machinery are classified based on the total effect index.
[0008] In one embodiment, the training method for the geometric structure classification model includes: Distortion analysis and labeling are performed on a portion of the initial geometric parameter population, and a training sample set for the geometric structure classification model is constructed. The geometric structure classification model is trained using the training sample set of the geometric structure classification model to obtain the trained geometric structure classification model.
[0009] In one embodiment, training the geometric structure classification model includes: When the dimension of the core geometric parameters is greater than the preset dimension, the geometric structure classification model based on random forest is trained. When the dimension of the core geometric parameters is less than or equal to the preset dimension, the geometric structure classification model based on support vector machine is trained.
[0010] In one embodiment, the step of performing hydrodynamic simulations on a subset of geometric parameter combinations from the optimized geometric parameter population, and constructing a training set sample for the physical field prediction model based on the simulation data, includes: Fluid dynamics simulations were performed on some combinations of geometric parameters in the optimized geometric parameter population to calculate simulation data of the performance parameters of each geometric parameter combination. By associating the various geometric parameter combinations with the corresponding simulation data, a training set sample for the physical field prediction model is constructed.
[0011] In one embodiment, the training set samples for constructing the physics prediction model by associating the combinations of geometric parameters and the corresponding simulation data include: Three-dimensional modeling is performed on some geometric parameter combinations in the optimized geometric parameter population to obtain three-dimensional models of each geometric parameter combination; The three-dimensional model with various geometric parameter combinations is converted into an SDF matrix. Each SDF matrix is associated with the corresponding simulation data to establish a training sample set for the physical field prediction model.
[0012] In one embodiment, the step of calling the trained physical field prediction model based on a preset optimization rule to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the geometric parameter design values of the target fluid machinery based on the calculated performance parameter prediction values of each geometric parameter combination, includes: The trained physics prediction model is invoked to predict the performance parameters of all geometric parameter combinations in the optimized geometric parameter population, so as to obtain the predicted performance parameter values of all geometric parameter combinations. Based on preset evaluation criteria, the predicted performance parameters of all geometric parameter combinations are sorted, and the optimal geometric parameter combination is determined as the geometric parameter design value of the target fluid machinery.
[0013] In one embodiment, the step of calling the trained physical field prediction model based on a preset optimization rule to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the geometric parameter design values of the target fluid machinery based on the calculated performance parameter prediction values of each geometric parameter combination, includes: Based on the iterative optimization algorithm and the trained physical field prediction model, iterative optimization is performed in the population of optimized geometric parameters until the convergence condition is met. The combination of geometric parameters that meets the convergence condition is determined as the design value of the geometric parameters of the target fluid machinery.
[0014] In one embodiment, the step of calling the trained physical field prediction model based on a preset optimization rule to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the geometric parameter design values of the target fluid machinery based on the calculated performance parameter prediction values of each geometric parameter combination, includes: In the coarse search phase, a representative subset is selected from the population of optimized geometric parameters. The trained physics prediction model is then called to predict the performance parameters of all geometric parameter combinations in the representative subset, so as to obtain the predicted performance parameter values of each geometric parameter combination in the representative subset. The predicted performance parameters of each geometric parameter combination in the representative subset are sorted, and the individuals with the best performance parameters in the first predetermined proportion are selected as the candidate starting point set for the local optimization stage. In the local optimization stage, for each starting point in the candidate starting point set, iterative optimization is performed in its neighborhood with that starting point as the center until the convergence condition is met. The combination of geometric parameters that meets the convergence condition is determined as the geometric parameter design value of the target fluid machinery.
[0015] Furthermore, to achieve the above objectives, this application also proposes a fluid machinery design apparatus, comprising: The sample screening module is used to obtain the initial geometric parameter population of the target fluid machinery; classify the initial geometric parameter population using the trained geometric structure classification model and remove distorted individuals to obtain an optimized geometric parameter population; perform fluid dynamics simulation on some geometric parameter combinations in the optimized geometric parameter population; construct a training set sample for the physical field prediction model based on the simulation data; and train the physical field prediction model using the training set sample to obtain the trained physical field prediction model. The optimization module is used to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population based on preset optimization rules and by calling the trained physical field prediction model. Based on the calculated performance parameter prediction values of each geometric parameter combination, the module determines the geometric parameter design values of the target fluid machinery.
[0016] In addition, to achieve the above objectives, this application also proposes an apparatus comprising: a memory, a processor, and a computer program stored on the memory and running on the processor, the computer program being configured to implement the steps of the method described above.
[0017] In addition, to achieve the above objectives, this application also proposes a storage medium that is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the method described above.
[0018] In addition, to achieve the above objectives, this application also proposes a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method described above.
[0019] One or more technical solutions proposed in this application have at least the following technical effects: A fluid machinery design method is proposed. It obtains an initial set of geometric parameters for the target fluid machinery and uses a geometric structure classification model with distortion recognition capabilities to classify the combinations of geometric parameters in the initial set, thus obtaining an optimized set of geometric parameters. Only a subset of geometric parameter combinations from the optimized set are selected for fluid dynamics simulation, avoiding the waste of high-cost CFD simulation resources. Based on the simulation data, a training set of physical field prediction models is constructed, and these training set samples are used to train the physical field prediction model, thus avoiding contamination of the model training by distorted samples. Based on preset optimization rules, the trained physical field prediction model is invoked to predict the performance parameters of the geometric parameter combinations in the optimized set. Based on the calculated predicted performance parameter values of each geometric parameter combination, the design values of the target fluid machinery's geometric parameters are determined, significantly improving optimization efficiency. Attached Figure Description
[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0021] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating the first embodiment of the fluid machinery design method of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application. To better understand the technical solutions of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.
[0024] This embodiment proposes a fluid machinery design method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the fluid machinery design method of this application. The fluid machinery design method may include steps S10 to S40: Step S10: Obtain the initial geometric parameter population of the target fluid machinery.
[0025] In one feasible implementation, step S10 may include steps S101 to S102: Step S101: Based on the performance parameters of the target fluid machinery, divide the geometric parameters of the target fluid machinery into core geometric parameters and non-core geometric parameters.
[0026] It should be noted that the performance parameters are the performance items of the target fluid machinery. For example, when the target fluid machinery is an impulse turbine, the performance parameters can be configured as hydraulic efficiency and / or wear distribution. The core geometric parameters are multiple items that have a significant impact on the performance parameters, as determined by sensitivity analysis of all geometric parameters. The rest are non-core geometric parameters, resulting in a compressed set of core geometric parameters.
[0027] In a specific implementation process, step S101 may include step S1011: Step S1011: Obtain the set of geometric parameters of the target fluid machinery and construct an orthogonal experimental scheme; perform fluid dynamics simulation (CFD) on each experimental group in the orthogonal experimental scheme to determine the performance parameters corresponding to each experimental group; perform variance analysis on the performance parameters corresponding to each orthogonal experimental scheme, and based on the variance analysis results, classify the core geometric parameters and non-core geometric parameters in the geometric parameters of the target fluid machinery.
[0028] Specifically, analysis of variance (ANOVA) is performed on the performance parameters corresponding to each orthogonal experimental scheme to calculate the main effect contribution rate of each geometric parameter. Based on the preset contribution rate threshold, the core geometric parameters and non-core geometric parameters of the target fluid machinery are divided.
[0029] In one specific implementation, step S101 may include: Based on the global sensitivity analysis method, the total effect index of each geometric parameter is calculated; The core and non-core geometric parameters of the target fluid machinery are classified based on the total effect index.
[0030] Specifically, if there are highly nonlinear and strong interaction effects among the parameters, a global integral operation is performed on the set of geometric parameters to calculate the total effect index of each geometric parameter. This index fully captures the second-order and higher-order interaction effects of a single variable and its coupling with all other variables. When the total effect index of a geometric parameter exceeds a preset threshold, it is determined to be a core geometric parameter; otherwise, it is determined to be a non-core geometric parameter.
[0031] Understandably, the second specific implementation process is applied to nonlinear, highly interactive scenarios, such as physical characteristics where the geometric parameters of an impulse turbine have a highly nonlinear impact on performance and exhibit strong interactive effects.
[0032] In this implementation, the threshold for the total effect index used to determine whether a parameter is a core geometric parameter is configured to be 0.05.
[0033] Step S102: Generate initial geometric parameter samples based on core geometric parameters and non-core geometric parameters.
[0034] It should be noted that the initial geometric parameter samples generated based on core geometric parameters and non-core geometric parameters aim to use core geometric parameters as the main variables, while using non-core geometric parameters as preset constants or non-main variables (within a small range of preset constants), thereby constructing a set of candidate parameters in the dimensionality reduction space that covers different geometric shapes and whose subsequent modeling costs are controllable.
[0035] Understandably, initial geometric parameter samples generated based on core and non-core geometric parameters have a higher spatial filling density than sampling in the original high-dimensional space, with the same sampling amount. This allows the limited sample budget to cover the core parameter design range relatively evenly and to hit the distortion region with a higher probability. This provides a sufficient number of positive and negative samples for subsequent classification model training, avoiding underfitting or overfitting of the classification model due to missing distortion samples in the high-dimensional sparse space.
[0036] Step S20: The initial geometric parameter population is classified using the trained geometric structure classification model, and distorted individuals are removed to obtain an optimized geometric parameter population.
[0037] In one feasible implementation, the training method for the geometric structure classification model may include steps A1-A2: Step A1 involves performing distortion analysis and labeling on a portion of the initial geometric parameter population, and constructing a training sample set for the geometric structure classification model.
[0038] It should be noted that the distortion analysis of geometric parameter samples can include the distortion analysis of the geometric model constructed based on the geometric parameter samples (e.g., whether surface self-intersection occurs, and / or whether Boolean operation failure occurs) and / or the mesh distortion analysis after the geometric model is meshed (based on one or more of the mesh aspect ratio, twist, Jacobian, orthogonal mass and volume).
[0039] In one specific implementation process, after generating the initial geometric parameter samples, CAD is called to perform 3D reconstruction of the geometric parameter combinations, and the corresponding geometric parameter combinations are labeled as "legitimate topology" and "distorted geometry" according to the CAD execution log; the mesh generation program is called to perform mesh generation on the geometric model labeled "legitimate topology", and if the distortion rate of the generated mesh is higher than the threshold, the geometric parameter combination is labeled as "distorted geometry".
[0040] Step A2: Train the geometric structure classification model using the training sample set of the geometric structure classification model to obtain the trained geometric structure classification model.
[0041] It should be noted that geometric structure classification models can be built based on machine learning algorithms such as Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Tree (GBT), or Artificial Neural Network (ANN).
[0042] In one feasible implementation, training the geometric structure classification model includes: When the dimension of the core geometric parameters is greater than the preset dimension, the geometric structure classification model based on random forest is trained. When the dimension of the core geometric parameters is less than or equal to the preset dimension, the geometric structure classification model based on support vector machine is trained.
[0043] In this embodiment, the preset dimension is configured as 12.
[0044] In one specific implementation, the geometric structure classification model is built upon support vector machines (SVMs), employing a soft-margin SVM combined with a radial basis function (RBF) to construct the decision boundary. For the input geometric parameter vector, the optimal hyperplane is determined by solving the following objective function:
[0045] To handle the complex nonlinear geometric constraints of fluid mechanics, a radial basis function (RBF) is introduced for high-dimensional mapping.
[0046] In the formula, : is the normal vector of the hyperplane, which determines the characteristic direction that distinguishes between "legal" and "distorted" schemes in the geometric design space.
[0047] b : This is the bias term, which determines the displacement and intercept of the classification decision boundary in the geometric parameter space.
[0048] : is a slack variable used to quantify the degree to which some edge geometric sample points deviate from the ideal classification hyperplane, thereby greatly enhancing the system's tolerance to nonlinear topological noise (such as curvature changes in a critical state).
[0049] : is the penalty coefficient, used to balance the accuracy of geometric screening with the structural complexity of the SVM model, and to optimize the robustness of sample classification.
[0050] and : These represent different combinations of geometric parameters generated by the preceding LHS (Latin Hypercube Sampling) algorithm.
[0051] : The bandwidth coefficient of the RBF kernel function, which precisely controls the distribution shape after mapping complex geometric features.
[0052] In one specific implementation, the sample covariance matrix is calculated by principal component analysis (PCA) and eigenvalue decomposition is performed to map high-dimensional parameters to an orthogonal low-dimensional space to eliminate noise generated by nonlinear coupling. Subsequently, the bootstrapping mechanism of a random forest (RF) composed of multiple decision trees is used for integrated discrimination to ensure the robustness of interception under high-dimensional extreme conditions.
[0053] It should be noted that the continuous parameter space of fluid machinery is relatively sparse. SVM, based on the strict structural risk minimization theory, can find the hyperplane that maximizes the margin classification by relying on only a few key "support vectors". It effectively overcomes the defect of deep neural networks (such as MLP) being prone to overfitting and has excellent boundary characterization ability in high-dimensional sparse space. With the help of the RBF kernel function, SVM can implicitly map complex geometric parameters to high-dimensional or even infinite-dimensional space, accurately separating the nonlinear critical state of "legitimate topology" and "distorted geometry" (such as extremely small curvature abrupt changes or hidden surface self-intersections). When the dimension of the core geometric parameters is still high, the PCA+RF dual mechanism ensures that even if the geometric parameter distribution is extremely uneven, it can still maintain stable interception accuracy, which greatly enhances the reliability of industrial-grade automated production lines.
[0054] Step S30: Perform fluid dynamics simulation on some combinations of geometric parameters in the optimized geometric parameter population. Based on the simulation data, construct a training set sample for the physical field prediction model, and use the training set sample to train the physical field prediction model to obtain the trained physical field prediction model.
[0055] In one feasible implementation, performing fluid dynamics simulations on a subset of geometric parameter combinations from the optimized geometric parameter population, and constructing a training set sample for the physical field prediction model based on the simulation data, may include steps B1-B2: Step B1: Perform fluid dynamics simulation on some combinations of geometric parameters in the optimized geometric parameter population to calculate the simulation data of the performance parameters of each geometric parameter combination.
[0056] Specifically, after constructing an optimized geometric parameter population, three-dimensional reconstruction and mesh generation can be performed on each combination of geometric parameters. Then, a fluid dynamics simulation solver can be called to carry out three-dimensional transient simulation to obtain simulation data of the performance parameters of each combination of geometric parameters, such as the hydraulic efficiency and surface wear gradient of an impulse turbine.
[0057] Step B2: Associate the combinations of geometric parameters with the corresponding simulation data to construct the training set samples for the physical field prediction model.
[0058] In one specific implementation, step B2 may include: Three-dimensional modeling is performed on some geometric parameter combinations in the optimized geometric parameter population to obtain three-dimensional models of each geometric parameter combination; The three-dimensional model with various geometric parameter combinations is converted into an SDF matrix. Each SDF matrix is associated with the corresponding simulation data to establish a training sample set for the physical field prediction model.
[0059] It should be noted that the SDF (Signed Distance Field) matrix refers to the signed numerical value of each voxel storing its nearest distance to the surface of the 3D model.
[0060] Specifically, analytical geometric mapping algorithms can be used to calculate the minimum Euclidean distance from global voxel nodes to solid surfaces, thus losslessly transforming discrete geometric parameters into a three-dimensional symbolic distance field (SDF) matrix.
[0061] In one feasible implementation, the physics prediction model is built on an MT-CNN network with dual-path concurrent output capability. It is a multi-task model, in which one branch is configured as a performance parameter prediction branch (including global pooling and depth mapping) to achieve rapid regression of the target performance parameters; the other branch is configured as a physics reconstruction branch (including transposed convolution operator) to capture spatial distribution patterns and output a cloud map corresponding to the input geometric parameters.
[0062] Preferably, the loss function of the physical field prediction model is constructed based on a dynamic weighting mechanism of homoscedastic uncertainty, and the mathematical expression of the loss function is as follows:
[0063] In the formula, The joint total loss during network training; This is the internal weight parameter matrix of a multi-task convolutional neural network; This is the prediction error term for the performance evaluation branch; This is the prediction error term for the physical field reconstruction branch; and The system introduces a learnable noise parameter (i.e., an adaptive weight adjustment factor). By updating this parameter, the system automatically penalizes subtasks with excessive errors, thereby finding the optimal gradient descent direction among multimodal outputs with vastly different dimensions.
[0064] It should be noted that, due to the huge difference in dimensions between performance parameters (scalars, of very small magnitude) and cloud maps (high-dimensional tensors, of extremely large magnitude), if conventional fixed weights are used for addition, the network gradient will be completely overwhelmed by the high magnitude of cloud map errors.
[0065] Preferably, the physics prediction model embeds a Physical Information Neural Network (PINN) mechanism, using the residuals of the continuity equation and momentum equation of the Navier-Stokes equations as penalty terms, which are then fused into the aforementioned adaptive joint loss function to construct a physics-constrained enhanced loss function.
[0066] In the formula, For enhanced total loss; The multi-task joint total loss is constructed based on the dynamic weighting mechanism of homoscedastic uncertainty, and is used to comprehensively characterize the prediction error of the performance evaluation branch and the physical field reconstruction branch. and These are the residuals of the partial differential equations (PDEs) calculated by substituting the physical fields output by the network into the continuity equation and the momentum equation, respectively. This is the physical constraint penalty coefficient. Through this mechanism, the network is forced to not only fit the CFD data labels when updating gradients during backpropagation, but also to strictly adhere to the underlying fluid dynamics conservation laws, ensuring that the predicted physical field is not only numerically high-fidelity, but also physically logically rigorous.
[0067] Step S40: Based on the preset optimization rules, the trained physical field prediction model is called to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population, and the geometric parameter design values of the target fluid machinery are determined based on the calculated performance parameter prediction values of each geometric parameter combination.
[0068] In one feasible implementation, step S40 may include any one of steps C1 to C3: Step C1: The trained physical field prediction model is invoked to predict the performance parameters of all geometric parameter combinations in the optimized geometric parameter population to obtain the predicted performance parameter values of all geometric parameter combinations. Based on the preset evaluation criteria, the predicted performance parameter values of all geometric parameter combinations are sorted, and the optimal geometric parameter combination is determined as the geometric parameter design value of the target fluid machinery.
[0069] It should be noted that for scenarios where the number of individuals in the population of geometric parameters is small, a high-throughput forward inference optimization method can be adopted, that is, the physical field prediction model is batched and inferred for all combinations of geometric parameters at once, and then the optimal solution is selected directly according to the prediction performance value. The advantages of this method are: (1) The calculation process is simple and direct, without the need for iterative evolution or gradient update, avoiding the debugging and convergence judgment of the algorithm hyperparameters; (2) Global determinism, as long as the sampling is dense enough (the population individuals can cover the core parameter space well), the optimal individual selected by sorting is the global optimum under the current sampling, and there is no risk of getting trapped in local optima; (3) Extremely low inference latency, the physical field prediction model itself is a lightweight neural network, which can complete the millisecond-level forward propagation of thousands of samples at once with the help of GPU parallel computing, and the total time is much less than the evaluation time of dozens of generations required by traditional iterative optimization algorithms (such as genetic algorithms and gray wolf algorithms).
[0070] Step C2 involves iteratively optimizing the geometric parameter population based on the iterative optimization algorithm and the trained physical field prediction model until the convergence condition is met. The geometric parameter combination that meets the convergence condition is then determined as the geometric parameter design value of the target fluid machinery.
[0071] It should be noted that for scenarios with a large population size for optimizing geometric parameters (e.g., tens of thousands to hundreds of thousands of individuals) or high dimensionality of core geometric parameters (e.g., 5 dimensions or more), while high-throughput forward inference directly on all individuals is feasible, it may lead to increased computational resource consumption or sampling discreteness issues that prevent accurate approximation of the optimal solution in the continuous space. Therefore, this implementation adopts a strategy of collaborative work between iterative optimization algorithms and a physics prediction model: the physics prediction model is used as a high-fidelity fitness evaluator, performing multi-generation iterative updates on the population of optimized geometric parameters (or its sampled subset), such as using genetic algorithms (GA), particle swarm optimization (PSO), grey wolf optimization (GWO), or Bayesian optimization. Each generation generates new candidate geometric parameter combinations through selection, crossover, mutation, or a surrogate model acquisition function, and the physics prediction model is called to quickly evaluate its performance until convergence conditions are met (e.g., fitness improvement is less than a preset threshold for multiple consecutive generations, the maximum number of iterations is reached, or the parameter change range converges). The advantages of this approach are: (1) fine search in continuous space, the iterative optimization algorithm can gradually approach the true global optimal solution within the range of continuous parameter changes, making up for the shortcomings of discrete sampling that may miss the optimal point; (2) balance between sample efficiency and accuracy, by guiding sampling rather than blind dense enumeration, a higher optimization accuracy can be achieved with fewer predictions, which is especially suitable for high-dimensional scenarios or scenarios where the evaluation cost is low but the number of calls still needs to be controlled.
[0072] Step C3: In the coarse search phase, a representative subset is selected from the population of optimized geometric parameters. The trained physical field prediction model is then invoked to predict the performance parameters of all geometric parameter combinations in the representative subset, thereby obtaining the predicted performance parameter values for each geometric parameter combination in the representative subset. The predicted performance parameter values for each geometric parameter combination in the representative subset are sorted, and the individuals with the best performance parameters in the top predetermined proportion are selected as the candidate starting point set for the local optimization phase. In the local optimization phase, for each starting point in the candidate starting point set, iterative optimization is performed in its neighborhood with that starting point as the center until the convergence condition is met. The geometric parameter combinations that meet the convergence condition are determined as the geometric parameter design values for the target fluid machinery.
[0073] In this embodiment, the predetermined proportion is set between 1% and 5%.
[0074] It should be noted that, for the mathematical dilemma that the core parameter combination after dimensionality reduction may still face an exponential explosion of dimensions (such as tens of millions or more) in the continuous space, simply using high-throughput forward inference (step S4021) will be infeasible due to the exponential relationship between sampling density and dimension. Although pure iterative optimization (step S4022) can be searched in the continuous space, if it is executed directly in the global domain, the initial convergence will be slow and it is easy to get trapped in local optima.
[0075] It is understandable that the fluid machinery design method in this embodiment adopts a hybrid collaborative strategy of coarse search and local optimization. First, a representative subset is selected from the population of optimized geometric parameters. The parameters of the representative subset are predicted using a prediction model, and the prediction results are sorted to screen out multiple high-fitness advantageous regions and lock the search range of local optimization. Then, a local fine search is performed within the locked search range of local optimization, and finally the global optimal solution is converged. The accuracy loss of discrete coarse screening is compensated by continuous optimization or high-density sampling, thus balancing the overall efficiency, cost and optimization accuracy.
[0076] In a specific implementation process, the local optimization stage can be combined with gradient-based optimization algorithms or particle swarm optimization algorithms to fine-tune parameters until the convergence condition is met. The convergence condition can be configured as the change in performance parameters within a certain number of consecutive generations being less than a preset threshold, or reaching the maximum number of iterations.
[0077] Specifically, after constructing the candidate starting point set, the inherent differentiability of neural networks is utilized to employ gradient-based optimization algorithms (such as Adam or gradient ascent) to automatically differentiate the input of MT-CNN (SDF tensor) using AutoDiff, continuously approximating the true global optimum; or heuristic algorithms such as Particle Swarm Optimization (PSO) are mounted in the minimal neighborhood for high-density secondary fine-tuning.
[0078] This embodiment of a fluid machinery design method eliminates all illegal individuals that could lead to CAD modeling failure, mesh distortion, or CFD divergence / distortion through a geometric structure classification model, constructing an optimized geometric parameter population. This ensures that the input data for subsequent optimization processes are all geometrically feasible and mesh-reachable "qualified samples," avoiding pipeline interruptions and wasted computational resources caused by distorted samples. Based on this pure and valid population, a physics prediction model trained with high-fidelity CFD data is invoked for millisecond-level performance prediction and physics reconstruction. Combined with preset optimization rules, the globally optimal geometric design is quickly locked. Therefore, this method achieves a fully automated, zero-interruption, high-fidelity intelligent design closed loop from the original high-dimensional geometric parameters to the final optimal design, significantly improving the design efficiency, prediction accuracy, and engineering reliability of fluid machinery.
[0079] For example, the fluid machinery method of this application is used to design an impulse turbine, with the performance parameter configured as hydraulic efficiency. The specific process may include steps A1 to A10: Step A1: Determine the original set of geometric parameters for the impulse turbine, including the bucket width. W Nozzle diameter D Spray angle α , depth of the blade H Export edge thickness t Imported chamfer radius r , water dividing edge angle β , number of bucket leaves Z Gap ratio ε Installation corner γ And so on, and set a reasonable range of values for each parameter.
[0080] Step A2 involves using Sobol global sensitivity analysis to generate 2000 initial samples within the design space, executing automated CFD simulations, and calculating the total effect index for each geometric parameter. STi Parameters with STi > 0.05 are identified as core geometric parameters, while the rest are non-core geometric parameters.
[0081] Step A3: In the core parameter space, use Latin hypercube sampling to generate 2000 candidate parameter combinations, and merge each combination with fixed non-core parameters to form complete geometric parameters.
[0082] Step A4: Perform automated CAD modeling and mesh generation on the above 2000 sets of parameters in sequence. If the modeling results in surface self-intersection or Boolean operation failure, or if the mesh generation results in negative volume or distortion >0.95, it is labeled as "distorted geometry" (label 0); otherwise, it is labeled as "legitimate topology" (label 1), and a training sample set for the geometric structure classification model is constructed.
[0083] Step A5: Train the SVM classifier using the training sample set of the geometric structure classification model to obtain the trained SVM classifier.
[0084] In step A6, within the core parameter space, several sets of candidate parameters are generated again using LHS. The SVM classifier from step A5 is used for pre-interception to remove individuals predicted as distorted, resulting in a valid combination of geometric parameters that constitutes the optimized geometric parameter population.
[0085] Step A7: Randomly select 2000 valid individuals from the optimized geometric parameter population, and perform high-fidelity CFD simulation on each individual to obtain the hydraulic efficiency scalar. η And a 3D surface wear contour map. At the same time, the 3D model of each individual is converted into an SDF matrix.
[0086] Step A8: Train a multi-task convolutional neural network with the SDF matrix as input and hydraulic efficiency and wear contour maps as output.
[0087] Step A9: Pack the valid geometric parameter combinations obtained in step A6 into tensors and input them in batches into MT. The CNN makes predictions, obtaining the efficiency prediction value for each sample. The samples are sorted in descending order of efficiency, and the top 5% are selected as the candidate starting point set.
[0088] Step A10: For each starting point in the candidate starting point set, perform gradient ascent optimization within its neighborhood using automatic differentiation (AutoDiff) to achieve MT. The CNN is used as a differentiable estimator, iterating 100 times. The most efficient individual among the local optima at each starting point is determined as the final design parameters, and the corresponding 3D model and physics map are output.
[0089] This application also provides a fluid machinery design apparatus, which may include: The sample screening module is used to obtain the initial geometric parameter population of the target fluid machinery; classify the initial geometric parameter population using the trained geometric structure classification model and remove distorted individuals to obtain an optimized geometric parameter population; perform fluid dynamics simulation on some geometric parameter combinations in the optimized geometric parameter population; construct a training set sample for the physical field prediction model based on the simulation data; and train the physical field prediction model using the training set sample to obtain the trained physical field prediction model. The optimization module is used to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population based on preset optimization rules and by calling the trained physical field prediction model. Based on the calculated performance parameter prediction values of each geometric parameter combination, the module determines the geometric parameter design values of the target fluid machinery.
[0090] The fluid machinery design apparatus provided in this application, employing the fluid machinery design method described in the above embodiments, can solve the same technical problem. Compared with related technologies, the beneficial effects of the fluid machinery design apparatus provided in this application are the same as those of the fluid machinery design method provided in the above embodiments, and other technical features in this fluid machinery design apparatus are the same as those disclosed in the fluid machinery design method of the above embodiments, and will not be repeated here.
[0091] This application also provides an apparatus that may include: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the fluid machinery design method in the above embodiments.
[0092] The device may include, but is not limited to, mobile terminals such as mobile phones, tablets, laptops, wearable devices, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), digital broadcast receivers, and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as servers, network terminals, embedded computers, industrial control computers, desktop computers, PCs (Personal Computers), and digital TVs.
[0093] The device provided in this application, employing one of the methods in the above embodiments, can solve the technical problems corresponding to each method. Compared with related technologies, the beneficial effects of the device provided in this application are the same as those of one of the methods provided in the above embodiments, and other technical features of the device are the same as those disclosed in the methods in the above embodiments, and will not be repeated here.
[0094] This application also provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to perform the methods in the above embodiments.
[0095] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory) with one or more wires electrically connected, optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, etc., or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by an instruction execution system or device, or a combination thereof. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.
[0096] The aforementioned computer-readable storage medium may be included in the aforementioned device; or it may exist independently and not assembled into the aforementioned device.
[0097] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the aforementioned device, enable the aforementioned device to perform the functions defined in one of the methods disclosed in the embodiments of this application.
[0098] Computer program code for performing the operations of this application can be written in one or more programming languages or a combination thereof. These programming languages include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or connected to an external computer, such as using an internet connection provided by an internet service provider.
[0099] The storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for performing the above-described methods, thereby solving the corresponding technical problems. Compared with related technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the methods provided in the above embodiments, and will not be repeated here.
[0100] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods described above.
[0101] The computer program product provided in this application can solve the same technical problem. Compared with related technologies, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the methods provided in the above embodiments, and will not be repeated here.
[0102] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.
Claims
1. A fluid machinery design method, characterized in that, The design method includes: Obtain the initial geometric parameter population of the target fluid machinery; The trained geometric structure classification model is used to classify the initial geometric parameter population and remove distorted individuals to obtain an optimized geometric parameter population. Fluid dynamics simulations are performed on some combinations of geometric parameters in the optimized geometric parameter population. Based on the simulation data, a training set sample for the physical field prediction model is constructed, and the physical field prediction model is trained using the training set sample to obtain the trained physical field prediction model. Based on the preset optimization rules, the trained physical field prediction model is invoked to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population. Based on the calculated performance parameter prediction values of each geometric parameter combination, the geometric parameter design values of the target fluid machinery are determined.
2. A fluid machinery design method as described in claim 1, characterized in that, The population for obtaining the initial geometric parameters of the target fluid machinery includes: Based on the performance parameters of the target fluid machinery, the core geometric parameters and non-core geometric parameters of the target fluid machinery are divided. Initial geometric parameter samples are generated based on core and non-core geometric parameters.
3. A fluid machinery design method as described in claim 2, characterized in that, The classification of core and non-core geometric parameters in the geometric parameters of the target fluid machinery based on its performance parameters includes: Based on the global sensitivity analysis method, the total effect index of each geometric parameter is calculated; The core and non-core geometric parameters of the target fluid machinery are classified based on the total effect index.
4. A fluid machinery design method as described in claim 1, characterized in that, The training method for the geometric structure classification model includes: Distortion analysis and labeling are performed on a portion of the initial geometric parameter population, and a training sample set for the geometric structure classification model is constructed. The geometric structure classification model is trained using the training sample set of the geometric structure classification model to obtain the trained geometric structure classification model.
5. A fluid machinery design method as described in claim 4, characterized in that, Training a geometric structure classification model includes: When the dimension of the core geometric parameters is greater than the preset dimension, the geometric structure classification model based on random forest is trained. When the dimension of the core geometric parameters is less than or equal to the preset dimension, the geometric structure classification model based on support vector machine is trained.
6. A fluid machinery design method as described in claim 1, characterized in that, The process of performing fluid dynamics simulations on a subset of geometric parameter combinations from the optimized geometric parameter population, and constructing a training set sample for the physical field prediction model based on the simulation data, includes: Fluid dynamics simulations were performed on some combinations of geometric parameters in the optimized geometric parameter population to calculate simulation data of the performance parameters of each geometric parameter combination. By associating the various geometric parameter combinations with the corresponding simulation data, a training set sample for the physical field prediction model is constructed.
7. A fluid machinery design method as described in claim 6, characterized in that, The training set samples for constructing the physical field prediction model by associating the combinations of geometric parameters and the corresponding simulation data include: Three-dimensional modeling is performed on some geometric parameter combinations in the optimized geometric parameter population to obtain three-dimensional models of each geometric parameter combination; The three-dimensional model with various geometric parameter combinations is converted into an SDF matrix. Each SDF matrix is associated with the corresponding simulation data to establish a training sample set for the physical field prediction model.
8. A fluid machinery design method as described in claim 1, characterized in that, The process of using a pre-defined optimization rule to call a trained physical field prediction model to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the design values of the geometric parameters of the target fluid machinery based on the calculated predicted performance parameter values of each geometric parameter combination, includes: The trained physics prediction model is invoked to predict the performance parameters of all geometric parameter combinations in the optimized geometric parameter population, so as to obtain the predicted performance parameter values of all geometric parameter combinations. Based on preset evaluation criteria, the predicted performance parameters of all geometric parameter combinations are sorted, and the optimal geometric parameter combination is determined as the geometric parameter design value of the target fluid machinery.
9. A fluid machinery design method as described in claim 1, characterized in that, The process of using a pre-defined optimization rule to call a trained physical field prediction model to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the design values of the geometric parameters of the target fluid machinery based on the calculated predicted performance parameter values of each geometric parameter combination, includes: Based on the iterative optimization algorithm and the trained physical field prediction model, iterative optimization is performed in the population of optimized geometric parameters until the convergence condition is met. The combination of geometric parameters that meets the convergence condition is determined as the design value of the geometric parameters of the target fluid machinery.
10. A fluid machinery design method as described in claim 1, characterized in that, The process of using a pre-defined optimization rule to call a trained physical field prediction model to predict the performance parameters of geometric parameter combinations in the optimized geometric parameter population, and determining the design values of the geometric parameters of the target fluid machinery based on the calculated predicted performance parameter values of each geometric parameter combination, includes: In the coarse search phase, a representative subset is selected from the population of optimized geometric parameters. The trained physics prediction model is then called to predict the performance parameters of all geometric parameter combinations in the representative subset, so as to obtain the predicted performance parameter values of each geometric parameter combination in the representative subset. The predicted performance parameters of each geometric parameter combination in the representative subset are sorted, and the individuals with the best performance parameters in the first predetermined proportion are selected as the candidate starting point set for the local optimization stage. In the local optimization stage, for each starting point in the candidate starting point set, iterative optimization is performed in its neighborhood with that starting point as the center until the convergence condition is met. The combination of geometric parameters that meets the convergence condition is determined as the geometric parameter design value of the target fluid machinery.
11. A fluid machinery design device, characterized in that, The device includes: The sample screening module is used to obtain the initial geometric parameter population of the target fluid machinery; classify the initial geometric parameter population using the trained geometric structure classification model and remove distorted individuals to obtain an optimized geometric parameter population; perform fluid dynamics simulation on some geometric parameter combinations in the optimized geometric parameter population; construct a training set sample for the physical field prediction model based on the simulation data; and train the physical field prediction model using the training set sample to obtain the trained physical field prediction model. The optimization module is used to predict the performance parameters of the geometric parameter combinations in the optimized geometric parameter population based on preset optimization rules and by calling the trained physical field prediction model. Based on the calculated performance parameter prediction values of each geometric parameter combination, the module determines the geometric parameter design values of the target fluid machinery.
12. A fluid machinery design device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and running on the processor, the computer program being configured to implement the steps of the method as claimed in any one of claims 1 to 10.
13. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored, which, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 10.
14. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 10.
