A three-dimensional design and optimization method for centrifugal pump impeller

Abstract

The application is a kind of centrifugal pump impeller three-dimensional design and optimization method, relates to the field of centrifugal pump design and fluid mechanics numerical simulation technology, comprising: using high-order conservative interpolation to realize lossless mapping of physical quantity when grid dynamically changes, and reconstructing polygon intersection matrix of dynamic and static interface. In the application, starting from the core pain point of centrifugal pump flow field numerical simulation, an adaptive grid system based on memory octree is constructed, through a double-threshold determination strategy driven by flow field characteristics, intelligent decision of grid refinement and coarsening is realized, which can accurately adapt to strong distortion flow areas such as impeller trailing edge vortex, tongue jet, blade cavitation, etc., and solve the contradiction that traditional static grid precision and efficiency cannot be considered; through forced execution of 2:1 grid level constraint, numerical divergence caused by adjacent grid level difference exceeding limit is avoided, and the stability of flow field calculation is ensured.

Description

Technical Field

[0001] This invention relates to the field of centrifugal pump design and fluid dynamics numerical simulation technology, and in particular to a method for three-dimensional design and optimization of centrifugal pump impellers. Background Technology

[0002] As a core component of fluid transport systems, the three-dimensional configuration of the impeller in a centrifugal pump directly determines its hydraulic efficiency, cavitation performance, and operational stability. Modern high-precision centrifugal pump design relies entirely on computational fluid dynamics (CFD) numerical simulations for flow field analysis and iterative optimization. Traditional centrifugal pump numerical simulations typically employ static structured or unstructured meshes, which cannot achieve adaptive mesh matching based on the spatial distortion characteristics of unsteady and complex flows such as impeller trailing-edge vortex shedding, volute jet flow, and blade leading-edge cavitation. Globally refined meshes lead to a sharp depletion of redundant computing power, while locally coarse meshes struggle to accurately capture flow details, resulting in a core technical contradiction between computational accuracy and solution efficiency.

[0003] Existing mesh dynamic adjustment techniques suffer from defects in physical quantity transfer and interface coupling in the numerical simulation of centrifugal pump flow fields. Low-order interpolation algorithms are often used during mesh refinement and coarsening, which can easily lead to non-conservation of key physical quantities such as mass and momentum and numerical oscillations, resulting in distortion of the flow field calculation. At the same time, the topology of the dynamic-static interface between the centrifugal pump's rotating impeller and stationary volute changes dynamically with the mesh. Traditional interface data transfer methods cannot reconstruct the polygon intersection relationship in real time, and flux transfer suffers from losses and numerical jumps, which directly affect the convergence of the flow field solution and the reliability of the simulation, making it difficult to support high-precision impeller design optimization.

[0004] Numerical simulation of the entire flow field of large-scale centrifugal pumps relies on the MPI parallel computing architecture. However, existing technologies have not achieved dynamic matching between grid computing load and parallel processes. Different computing nodes experience severe load imbalance due to differences in grid quantity and computational load, resulting in a significant reduction in computing resource utilization. At the same time, grid data relies on repeated reading and writing of external files and discrete memory storage, resulting in high disk I / O consumption and low memory management efficiency. It is impossible to achieve efficient dynamic reconstruction of grid topology, leading to long impeller design optimization iteration cycles and high computational costs. Existing fluid dynamics numerical simulation systems cannot simultaneously meet the high precision, high efficiency, and high stability requirements of centrifugal pump impeller design.

[0005] Therefore, a three-dimensional design and optimization method for centrifugal pump impellers is proposed to address the aforementioned problems. Summary of the Invention

[0006] The purpose of this invention is to provide a three-dimensional design and optimization method for centrifugal pump impellers in order to solve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for three-dimensional design and optimization of centrifugal pump impellers, comprising: Read the three-dimensional geometric data of the centrifugal pump and perform spatial discretization. Construct an octree data structure with a globally unique identifier in computer memory to complete the initial mapping between the fluid domain and the data structure. After solving the flow field at each time step, the leaf nodes are traversed, and the error indicator is calculated through the flow field characteristics. After double threshold judgment, a list of mesh refinement and coarsening to be processed is generated. Complete the memory allocation and splitting refinement, memory release and merging coarsening of the mesh according to the list of pending tasks, and enforce the 2:1 mesh level constraint. High-order conservation interpolation is used to achieve lossless mapping of physical quantities when the mesh changes dynamically, and the intersection matrix of the dynamic-static interface polygon is reconstructed. Real-time monitoring of MPI process load imbalance, abstracting the grid into an undirected graph and completing cross-node data migration through graph partitioning.

[0008] Preferably, the step of reading the three-dimensional geometric data of the centrifugal pump and performing spatial discretization, constructing an octree data structure with a globally unique identifier in computer memory, and completing the initial mapping between the fluid domain and the data structure specifically includes: Read the pre-generated 3D CAD geometric boundary data of the centrifugal pump impeller, hub, volute, tongue, and inlet / outlet flow channels, and perform standardized preprocessing on the geometric data to obtain a closed fluid computational domain geometric model without topological errors; The preprocessed centrifugal pump full-fluid computational domain is spatially partitioned to generate a background coarse mesh; the node coordinates, surface topology, and volume topology of all coarse meshes reside in allocated contiguous memory. Each independent background coarse hexahedral mesh cell is defined as a root node. A root node index matrix is ​​constructed in the form of a two-dimensional sparse matrix. The row and column indices of the matrix correspond to the topological positions of the coarse mesh on the X, Y, and Z axes in space. The matrix elements store the globally unique integer ID code of the root node. The ID code adopts a combination rule of spatial coordinates plus mesh number. The processor constructs a global octree data dictionary based on a hash table in memory. Each tree node contains fixed, complete, and directly callable memory data fields, including: node ID, spatial coordinate boundary, physical quantity array pointer, parent node pointer, child node pointer, and grid level identifier. After initialization, the octree data dictionary and the root node matrix are bidirectionally bound.

[0009] Preferably, the step of traversing leaf nodes after solving the flow field at each time step, calculating the error indicator through flow field characteristics, and generating a list of meshes to be refined or coarsened after double threshold determination specifically includes: At the current time step After the Navier-Stokes fluid control equations are solved iteratively, the traversal algorithm traverses the global octree and selects all leaf nodes that have no lower-level child nodes and directly participate in the numerical calculation. During the traversal, the velocity vector and pressure scalar physical quantity data in the corresponding memory block are read through the node physical quantity pointer and temporarily stored in the register for subsequent calculations; For unsteady and complex flow inside a centrifugal pump, a central difference scheme is used to calculate the velocity curl and pressure spatial gradient of the flow field inside the leaf node, quantifying the severity of flow field distortion. It then calls a preset error evaluation function to calculate the node error indicator. The calculation formula is as follows: ; in: For the first Error indicators for each leaf node; For the first The velocity curl of each node; For the first The grid spatial characteristic scale of each node.

[0010] Preferably, the process further includes dual threshold comparison and generation of a list of items to be processed: The processor calls the pre-stored refinement threshold in memory. With coarsening threshold Error indicators for all leaf nodes Each value is compared with the two thresholds, and two linearly stored lists of items to be processed are generated based on the judgment results: like > If the flow field distortion at a node is deemed severe, it is added to the list of nodes to be split in the mesh refinement process. like < If the flow field at a node is determined to be stable, it is added to the list of nodes to be coarsened and merged. like ≤ ≤ The decision nodes do not need to be adjusted; the current grid state remains unchanged.

[0011] Preferably, the step of completing the memory allocation and splitting refinement, memory release and merging coarsening of the mesh according to the list to be processed, and forcibly implementing the 2:1 mesh level constraint, specifically includes: The processor sequentially traverses the list of mesh refinements to be split, performing a normalization refinement operation on each parent node to be split: Memory allocation: Request 8 contiguous physical memory blocks from the system through memory management functions to serve as child node storage units; Child node generation: According to the octree space equal division rule, the parent node hexahedral mesh is equally divided into 8 sub-mesh, the spatial coordinate boundary of the child node is calculated, a globally unique child node ID is assigned, and the mesh level is set to the parent node + 1 level; Pointer update: The pointers of the parent node's eight child nodes are pointed to the newly allocated child node memory address one by one, thus completing the parent-child node binding; Queue update: Add 8 child nodes to the fluid computation queue, remove the original parent node, and the child nodes participate in the next time step computation as new leaf nodes; The processor sequentially traverses the list of mesh coarsening to be merged, performing validity checks, memory releases, state rollbacks, and log entries.

[0012] Preferably, the method further includes enforcing a 2:1 mesh hierarchy balance constraint: After topology reconstruction, traverse all adjacent leaf nodes and enforce a 2:1 mesh hierarchy constraint: The mesh level difference between any two adjacent leaf nodes is limited to ≤1, and fine meshes and coarse meshes are prohibited from being directly adjacent; If the difference between adjacent grid levels is detected to exceed the limit, the coarser grid at the lower level will be automatically passively refined until the constraint conditions are met.

[0013] Preferably, the step of using high-order conserved interpolation to achieve lossless mapping of physical quantities when the mesh dynamically changes, and reconstructing the intersection matrix of the dynamic-static interface polygon, specifically includes: In the instantaneous state of mesh splitting / merging, the lossless mapping of fluid physical quantities between parent and child nodes is completed, and integral conservation constraints are executed; After interpolation is completed, a conservation error check is automatically performed. An allowable error threshold is set. If the error exceeds the allowable threshold, the interpolation operation is re-executed until the conservation requirement is met.

[0014] Preferably, the method further includes dynamic reconstruction of the static-dynamic interface and flux transfer: The interface between the centrifugal pump's rotating impeller and the stationary volute serves as the data transfer interface. Dynamic additions and deletions to the mesh dynamically alter the interface topology, prompting the processor to perform the following operations: Real-time retrieval of grid nodes located at the interface between dynamic and static regions in an octree, and marking of the interface topological boundary;

[0015] Calculate the spatial overlap area of ​​the dynamic and static grids on both sides of the interface, and generate a polygon intersection matrix. The matrix elements are the area weights and overlap coefficients of the interface grids. Data coupling and transfer between dynamic and static regions is accomplished based on polygon intersection matrix.

[0016] Preferably, the real-time monitoring of MPI process load imbalance, which abstracts the mesh as an undirected graph and completes cross-node data migration through graph partitioning, specifically includes: The main thread monitors the running metrics of all MPI parallel processes in real time, including: single-process computation time, number of grid nodes managed, CPU utilization, and memory usage. The quantitative load imbalance is calculated using the formula: Load imbalance = (Maximum process time - Minimum process time) / Average process time; When the load imbalance exceeds the preset tolerance, the dynamic load balancing process is automatically triggered. Construction of an undirected graph data model for the computational domain: Abstract all leaf nodes of the current octree into a parallel computational undirected graph model: Vertex: Each leaf grid node corresponds to a graph vertex; Edge: The topological connection between adjacent grids corresponds to an undirected edge in the graph; Vertex weight: The computational cost of a single node.

[0017] Preferably, the method further includes: With the goal of ensuring that the sum of the vertex weights in each parallel partition is equal, the undirected graph is re-partitioned. Pack the node IDs, coordinates, physical quantity arrays, pointer information, and topological relationships of cross-boundary meshes into standardized data packets; Data message transmission is accomplished through network communication protocols; The receiving end completes data unpacking, re-establishes octree pointer binding and memory mapping, and verifies the load balancing effect.

[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention addresses the core challenges of numerical simulation of centrifugal pump flow fields by constructing an adaptive mesh system based on a memory octree. Through a dual-threshold decision-making strategy driven by flow field characteristics, it achieves intelligent decision-making for mesh refinement and coarsening, accurately adapting to highly distorted flow regions such as impeller trailing edge vortices, tongue jets, and blade cavitation, thus resolving the contradiction between accuracy and efficiency in traditional static meshes. By forcibly enforcing a 2:1 mesh hierarchy constraint, it avoids numerical divergence caused by excessive differences between adjacent mesh levels, ensuring the stability of flow field calculations.

[0019] 2. This invention achieves lossless mapping of physical quantities such as mass and momentum during dynamic grid adjustment by employing a high-order conservation interpolation algorithm. It reconstructs the intersection matrix of the dynamic-static interface polygon to ensure flux conservation and transmission, eliminates physical quantity distortion and numerical jumps, and significantly improves the convergence and simulation accuracy of the Reynolds-averaged Navier-Stokes equations. This provides reliable technical support for the accurate analysis of complex flow fields inside centrifugal pumps. Attached Figure Description

[0020] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0022] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0023] Example 1

[0024] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0025] Appendix Figure 1 The flowchart of a three-dimensional design and optimization method for a centrifugal pump impeller provided in this embodiment of the invention shows the complete steps from reading the three-dimensional geometric data of the centrifugal pump and completing spatial discretization to abstracting the mesh into an undirected graph and completing cross-node data migration through graph partitioning.

[0026] In this embodiment, it includes: Read the three-dimensional geometric data of the centrifugal pump and perform spatial discretization. Construct an octree data structure with a globally unique identifier in computer memory to complete the initial mapping between the fluid domain and the data structure. Specifically, it includes: This step lays the foundation for the data structure of the entire solution, completing the one-time initialization mapping of the centrifugal pump fluid computation domain from physical geometry to computer memory data structure, and building a globally dynamic octree data management system. All operations are completed in computer memory, without relying on repeated reading and writing of external mesh files.

[0027] Centrifugal pump 3D geometric boundary data reading and preprocessing: The central processing unit reads pre-generated complete 3D CAD geometric boundary data of the centrifugal pump impeller, hub, volute, tongue, inlet and outlet channels, etc., through a high-speed I / O interface, and performs standardized preprocessing on the geometric data: By removing duplicate geometric vertices, repairing broken geometric surfaces and non-manifold edges, and unifying the geometric coordinate units to the International System of Units (SI), a closed fluid computational domain geometric model without topological errors is obtained after preprocessing. The preprocessing results are temporarily stored in a cache to avoid repeated reading.

[0028] Spatial discretization and contiguous memory allocation with a coarse background grid: The processor allocates contiguous heap memory blocks in the computer's physical memory according to the preset initial mesh precision parameters. It uses a structured hexahedral mesh discretization method to spatially partition the preprocessed centrifugal pump full fluid computation domain, generating a uniform background coarse mesh. All node coordinates, surface topology, and volume topology relationships of the coarse mesh reside in the allocated contiguous memory in the form of single-precision / double-precision floating-point arrays, without generating local mesh files, thus reducing computational losses caused by disk I / O interactions.

[0029] Root node matrix construction and globally unique identifier encoding: Each independent background coarse hexahedral mesh cell is defined as a root node. A root node index matrix is ​​constructed in the form of a two-dimensional sparse matrix. The row and column indices of the matrix strictly correspond to the topological positions of the coarse mesh on the X, Y, and Z axes in space. The matrix elements store the globally unique integer ID code of the root node. The ID code adopts a combination rule of spatial coordinates plus mesh number to ensure that there are no duplicates of all mesh nodes and that they can be quickly addressed and located.

[0030] Global octree data dictionary initialization and field definition: The processor constructs a global octree data dictionary based on a hash table in memory, which serves as the core management structure for the entire flow field mesh. Each tree node contains fixed, complete, and directly accessible memory data fields, the specific fields and functions of which are as follows: Node ID: A globally unique integer identifier used for fast retrieval in the hash dictionary; Spatial coordinate boundary: storage node corresponds to hexahedral mesh , , , , , Extreme coordinates uniquely define the spatial range of nodes; Pointer to the array of physical quantities: points to the starting address of a contiguous memory block storing fluid physical quantities such as pressure, velocity vector, density, vorticity, and turbulent kinetic energy; Parent pointer: Points to the physical memory address of the current node's parent node. The root node's parent pointer is set to null by default. 8 child node pointers: Initially, all are set to null pointers, and valid memory addresses are allocated only when mesh refinement is triggered; Grid level identifier: Records the refinement level of the current node in the octree. The root node defaults to level 0.

[0031] After initialization, the octree data dictionary and the root node matrix are bidirectionally bound, forming a dynamically growing fluid data management system.

[0032] After solving the flow field at each time step, the leaf nodes are traversed, and the error indicator is calculated through the flow field characteristics. After double threshold judgment, a list of mesh refinement and coarsening to be processed is generated. Specifically, it includes: This step is the core of intelligent decision-making for dynamic grid adjustment. It is executed automatically after the fluid equation is solved at each time step, completing the quantification of flow field characteristics, grid adjustment judgment, and generation of the list of tasks to be processed. No manual intervention is required, thus achieving precise allocation of computing resources.

[0033] Leaf node traversal and flow field physical quantity extraction: At the current time step After the Navier-Stokes (NS) fluid control equations are iteratively solved, the processor uses a depth-first traversal algorithm to traverse the global octree and select all leaf nodes that have no lower-level child nodes and directly participate in the numerical calculation. Iterative solution process of fluid control equations: at the current time step Given that the mesh topology reconstruction has been completed, the Reynolds-averaged Navier-Stokes (RANS-NS) control equations for incompressible fluids are spatially discretized using the finite volume method based on all leaf node meshes in the octree. This transforms the partial differential control equations into a sparse linear algebraic equation system that can be directly computed by a computer. Then, the SIMPLE / SIMPLEC pressure-velocity coupled iterative algorithm is used to iteratively solve the algebraic equation system in memory. First, the velocity field is estimated using the pressure field at the current moment, and the pressure correction equation is constructed and solved. Then, the velocity field and flow coefficient are updated using the pressure correction value, and the physical quantities such as pressure, velocity, and turbulent kinetic energy of each leaf node are corrected successively. After each iteration, the global residual is calculated and compared with the preset convergence threshold until the residual decreases to within the convergence threshold and the physical quantities no longer change significantly. This indicates that the iterative solution of the fluid control equations for this time step t is complete, and the solved physical quantities are written back to the memory array corresponding to the octree node. Then, the subsequent flow field feature extraction and error indicator calculation stages are entered. The solution process is a direct reference to existing technology and will not be elaborated here.

[0034] During the traversal, the velocity vector and pressure scalar physical quantity data in the corresponding memory block are read through the node physical quantity pointer and temporarily stored in the register for subsequent calculations; Calculation of flow field characteristic gradient and error indicator: For the unique unsteady and complex flow inside the centrifugal pump (blade trailing edge vortex shedding, blade leading edge cavitation, volute tongue jet, secondary flow, etc.), the central difference scheme is used to calculate the velocity curl and pressure spatial gradient of the flow field inside the blade node, and to quantify the severity of flow field distortion. It then calls a preset error evaluation function to calculate the node error indicator. The core calculation formula is as follows: ; in: For the first The error indicator for each leaf node is positively correlated with the degree of flow field distortion; For the first The velocity curl of each node characterizes the eddy current intensity; For the first The grid spatial characteristic scale of each node, i.e., the side length of the current hexahedral grid.

[0035] It also includes dual threshold comparison and generation of a list of tasks to be processed: The processor calls the pre-stored refinement threshold in memory. With coarsening threshold (The threshold is a configurable parameter. In this embodiment, the default refinement threshold is 0.05 and the coarsening threshold is 0.01. It can be dynamically modified according to the calculation accuracy requirements.) Error indicators for all leaf nodes. Each value is compared with the two thresholds, and two linearly stored lists of items to be processed are generated based on the judgment results: like > If the flow field distortion at a node is deemed severe, it is added to the list of nodes to be split in the mesh refinement process. like < If the flow field at a node is determined to be stable, it is added to the list of nodes to be coarsened and merged. like ≤ ≤ The decision nodes do not need to be adjusted; the current grid state remains unchanged.

[0036] The list of items to be processed is stored in a sequential storage structure, which facilitates batch traversal and execution of subsequent topology reconstruction steps.

[0037] Complete the memory allocation and splitting refinement, memory release and merging coarsening of the grid according to the list of tasks to be processed, and enforce the 2:1 grid level constraint to ensure computational stability; Specifically, it includes: This step focuses on the underlying logic of dynamic memory allocation, pointer updates, topology reconstruction, and resource reclamation in computers.

[0038] Complete execution flow of mesh splitting (refinement): The processor sequentially traverses the list of mesh refinements to be split, performing a normalization refinement operation on each parent node to be split: Memory allocation: Eight contiguous physical memory blocks are requested from the system through memory management functions to serve as child node storage units. Contiguous memory can improve CPU cache hit rate and reduce data access latency. Child node generation: According to the octree space equal division rule, the parent node hexahedral mesh is equally divided into 8 sub-mesh, the spatial coordinate boundary of the child node is calculated, a globally unique child node ID is assigned, and the mesh level is set to the parent node + 1 level; Pointer update: The pointers of the parent node's eight child nodes are pointed to the newly allocated child node memory address one by one, thus completing the parent-child node binding; Queue update: Add 8 child nodes to the fluid computation queue, remove the original parent node, and the child nodes participate in the next time step computation as new leaf nodes; The complete execution flow of mesh merging (coarsening): The processor sequentially traverses the list of meshes to be coarsened and merged, performing validity checks, memory releases, state rollbacks, and logging, as detailed below: Legality check: Determine whether the child nodes to be merged belong to the same parent node, and whether all 8 child nodes under the parent node meet the coarsening conditions; none of them can be missing. Memory release: After the verification is passed, the memory release function is called to reclaim the physical memory occupied by the 8 child nodes and clear the child node pointers to avoid memory leaks and resource waste; State rollback: The computation weight is redirected to the parent node, the parent node is restored to a leaf node, and it is re-added to the computation queue; Log recording: Records the node ID and time step of the merge operation for easy subsequent calculation and debugging.

[0039] It also includes the enforcement of a 2:1 mesh hierarchy balance constraint: After topology reconstruction, the processor uses a breadth-first search algorithm to traverse all adjacent leaf nodes, enforcing a 2:1 mesh hierarchy constraint. The grid level difference between any two adjacent leaf nodes is limited to ≤1, and direct adjacency between fine and coarse grids is prohibited to prevent numerical calculation divergence. If the difference between adjacent mesh levels is detected to exceed the limit, the coarse mesh at the lower level is automatically passively refined until the constraint conditions are met. This prevents ill-conditioned finite element stiffness matrix from the data structure level and ensures computational stability.

[0040] High-order conservation interpolation is used to achieve lossless mapping of physical quantities when the mesh changes dynamically, and the intersection matrix of the dynamic-static interface polygon is reconstructed to ensure flux conservation and transmission. Specifically, it includes: Higher-order conservation interpolation and integral conservation verification: During the instantaneous state of mesh splitting / merging, the processor invokes either the volume-weighted interpolation algorithm or the Galerkin finite element projection algorithm to complete the lossless mapping of fluid physical quantities between parent and child nodes, strictly enforcing integral conservation constraints. ; in: This represents the total integral of mass and momentum within the parent node element. It is the sum of the integrals of the physical quantities corresponding to the 8 child nodes; For fluid density; For fluid velocity vector; This is the spatial domain of the hexahedral mesh element corresponding to the parent node; For the first The spatial domain of the hexahedral mesh element corresponding to each child node. =1 to 8, the octree splits into 8 sub-units each time; Let be a volume element, indicating that the integration is performed over the entire unit volume; After interpolation is complete, a conservation error check is automatically performed, and the allowable error threshold is set to [value missing]. If the error exceeds the allowable threshold, the interpolation operation will be re-executed until the conservation requirement is met, thus preventing the loss of physical quantity data.

[0041] It also includes dynamic reconstruction of the static-dynamic interface and flux transfer: The interface between the centrifugal pump's rotating impeller (dynamic region) and the stationary volute (static region) is the core data transfer interface. Dynamic additions and deletions of the mesh will change the interface topology in real time, and the processor will perform the following operations: Interface identification: Real-time retrieval of grid nodes located at the interface between dynamic and static regions in the octree, and marking the interface topological boundary; Polygon intersection matrix calculation: Calculate the spatial overlap area of ​​the dynamic and static grids on both sides of the interface, and generate a polygon intersection matrix. The matrix elements are the area weights and overlap coefficients of the interface grids. Flux conservation and transfer: Based on the intersection matrix of polygons, the data coupling and transfer between dynamic and static regions is completed, ensuring that there is no flux loss or numerical jump when flow, momentum and energy pass through the dynamic grid interface, thus guaranteeing the continuity and stability of the dynamic-static interface calculation.

[0042] Real-time monitoring of MPI process load imbalance; abstracting the grid into an undirected graph and completing cross-node data migration through graph partitioning to achieve dynamic load balancing of parallel computing. Specifically, it includes: Real-time performance monitoring and load balancing of parallel processes: The main thread monitors the running metrics of all MPI (Message Passing Interface) parallel processes in real time, including: single-process computation time, number of grid nodes managed, CPU utilization, and memory usage. The quantitative load imbalance is calculated using the formula: Load imbalance = (Maximum process time - Minimum process time) / Average process time; When the load imbalance exceeds the preset tolerance (15% by default in this embodiment), the dynamic load balancing (DLB) process is automatically triggered. Construction of an undirected graph data model for the computational domain: Abstracting all leaf nodes of the current octree into a parallel computational undirected graph model, thus completing the mapping from physical mesh to graph data: Vertex: Each leaf grid node corresponds to a graph vertex; Edge: The topological connection between adjacent grids corresponds to an undirected edge in the graph; Vertex weight: The computational cost of a single node (number of physical quantity solutions, memory read / write frequency, floating-point operation cost); The graph model is stored in memory in the form of an adjacency list, which is adapted to the input requirements of parallel graph partitioning algorithms.

[0043] Graph partitioning and cross-node data migration rebinding: Graph partitioning: The Zoltán and ParMETIS standard graph partitioning algorithms are called, and the undirected graph is re-partitioned with the goal of equal sum of vertex weights in each parallel partition. Data Packaging: Pack the node IDs, coordinates, physical quantity arrays, pointer information, and topological relationships of cross-boundary meshes into standardized data packets; Parallel communication: Data packets are transmitted between different CPU cores and computing nodes via the TCP / IP network communication protocol; Rebinding verification: After the receiving end completes data unpacking, it re-establishes octree pointer binding and memory mapping, verifies the load balancing effect, ensures that the computing power of each process is evenly distributed, and maximizes the efficiency of parallel computing.

[0044] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0045] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0046] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0047] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0048] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0049] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0051] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0053] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for three-dimensional design and optimization of centrifugal pump impellers, characterized in that, include: Read the three-dimensional geometric data of the centrifugal pump and perform spatial discretization. Construct an octree data structure with a globally unique identifier in computer memory to complete the initial mapping between the fluid domain and the data structure. After solving the flow field at each time step, the leaf nodes are traversed, and the error indicator is calculated through the flow field characteristics. After double threshold judgment, a list of mesh refinement and coarsening to be processed is generated. Complete the memory allocation and splitting refinement, memory release and merging coarsening of the mesh according to the list of pending tasks, and enforce the 2:1 mesh level constraint. High-order conservation interpolation is used to achieve lossless mapping of physical quantities when the mesh changes dynamically, and the intersection matrix of the dynamic-static interface polygon is reconstructed. Real-time monitoring of MPI process load imbalance, abstracting the grid into an undirected graph and completing cross-node data migration through graph partitioning.

2. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 1, characterized in that, Read the three-dimensional geometric data of the centrifugal pump and perform spatial discretization. Construct an octree data structure with a globally unique identifier in computer memory to complete the initial mapping between the fluid domain and the data structure. Specifically, this includes: Read the pre-generated 3D CAD geometric boundary data of the centrifugal pump impeller, hub, volute, tongue, and inlet / outlet flow channels, and perform standardized preprocessing on the geometric data to obtain a closed fluid computational domain geometric model without topological errors; The preprocessed centrifugal pump full-fluid computational domain is spatially partitioned to generate a background coarse mesh; the node coordinates, surface topology, and volume topology of all coarse meshes reside in allocated contiguous memory. Each independent background coarse hexahedral mesh cell is defined as a root node. A root node index matrix is ​​constructed in the form of a two-dimensional sparse matrix. The row and column indices of the matrix correspond to the topological positions of the coarse mesh on the X, Y, and Z axes in space. The matrix elements store the globally unique integer ID code of the root node. The ID code adopts a combination rule of spatial coordinates plus mesh number. The processor constructs a global octree data dictionary based on a hash table in memory. Each tree node contains fixed, complete, and directly callable memory data fields, including: node ID, spatial coordinate boundary, physical quantity array pointer, parent node pointer, child node pointer, and grid level identifier. After initialization, the octree data dictionary and the root node matrix are bidirectionally bound.

3. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 1, characterized in that, After solving the flow field at each time step, the leaf nodes are traversed, and the error indicator is calculated using flow field characteristics. A list of elements to be processed for mesh refinement and coarsening is generated after a dual-threshold judgment. Specifically, this list includes: At the current time step After the Navier-Stokes fluid control equations are solved iteratively, the traversal algorithm traverses the global octree and selects all leaf nodes that have no lower-level child nodes and directly participate in the numerical calculation. During the traversal, the velocity vector and pressure scalar physical quantity data in the corresponding memory block are read through the node physical quantity pointer and temporarily stored in the register for subsequent calculations; For unsteady and complex flow inside a centrifugal pump, a central difference scheme is used to calculate the velocity curl and pressure spatial gradient of the flow field inside the leaf node, quantifying the severity of flow field distortion. It then calls a preset error evaluation function to calculate the node error indicator. The calculation formula is as follows: ； in: For the first Error indicators for each leaf node; For the first The velocity curl of each node; For the first The grid spatial characteristic scale of each node.

4. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 3, characterized in that, It also includes dual threshold comparison and generation of a list of tasks to be processed: The processor calls the pre-stored refinement threshold in memory. With coarsening threshold Error indicators for all leaf nodes Each value is compared with the two thresholds, and two linearly stored lists of items to be processed are generated based on the judgment results: like > If the flow field distortion at a node is deemed severe, it is added to the list of nodes to be split in the mesh refinement process. like < If the flow field at a node is determined to be stable, it is added to the list of nodes to be coarsened and merged. like ≤ ≤ The decision nodes do not need to be adjusted; the current grid state remains unchanged.

5. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 1, characterized in that, Complete the memory allocation and splitting refinement, memory release and merging coarsening of the mesh according to the pending list, and enforce the 2:1 mesh level constraint, specifically including: The processor sequentially traverses the list of mesh refinements to be split, performing a normalization refinement operation on each parent node to be split: The system requests 8 contiguous physical memory blocks through memory management functions to serve as child node storage units. According to the octree space division rule, the parent node hexahedral mesh is equally divided into 8 sub-mesh, the spatial coordinate boundary of the sub-node is calculated, a globally unique sub-node ID is assigned, and the mesh level is set to the parent node + 1 level; The parent node's 8 child node pointers are pointed to the newly allocated child node memory addresses one by one, thus completing the parent-child node binding; Eight child nodes are added to the fluid computation queue, the original parent node is removed, and the child nodes participate in the next time step computation as new leaf nodes.

6. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 5, characterized in that, It also includes the enforcement of a 2:1 mesh hierarchy balance constraint: After topology reconstruction, traverse all adjacent leaf nodes and enforce a 2:1 mesh hierarchy constraint: The mesh level difference between any two adjacent leaf nodes is limited to ≤1, and fine meshes and coarse meshes are prohibited from being directly adjacent; If the difference between adjacent grid levels is detected to exceed the limit, the coarser grid at the lower level will be automatically passively refined until the constraint conditions are met.

7. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 1, characterized in that, High-order conservation interpolation is used to achieve lossless mapping of physical quantities when the mesh changes dynamically, and the intersection matrix of the dynamic-static interface polygon is reconstructed, specifically including: In the instantaneous state of mesh splitting / merging, the lossless mapping of fluid physical quantities between parent and child nodes is completed, and integral conservation constraints are executed; After interpolation is completed, a conservation error check is automatically performed. An allowable error threshold is set. If the error exceeds the allowable threshold, the interpolation operation is re-executed until the conservation requirement is met.

8. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 7, characterized in that, It also includes dynamic reconstruction of the static-dynamic interface and flux transfer: The interface between the centrifugal pump's rotating impeller and the stationary volute serves as the data transfer interface. Dynamic additions and deletions to the mesh dynamically alter the interface topology, prompting the processor to perform the following operations: Real-time retrieval of grid nodes located at the interface between dynamic and static regions in an octree, and marking of the interface topological boundary; Calculate the spatial overlap area of ​​the dynamic and static grids on both sides of the interface, and generate a polygon intersection matrix. The matrix elements are the area weights and overlap coefficients of the interface grids. Data coupling and transfer between dynamic and static regions is accomplished based on polygon intersection matrix.

9. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 1, characterized in that, Real-time monitoring of MPI process load imbalance; abstracting the mesh into an undirected graph and completing cross-node data migration through graph partitioning; specifically including: The main thread monitors the running metrics of all MPI parallel processes in real time, including: single-process computation time, number of grid nodes managed, CPU utilization, and memory usage. The quantitative load imbalance is calculated using the formula: Load imbalance = (Maximum process time - Minimum process time) / Average process time; When the load imbalance exceeds the preset tolerance, the dynamic load balancing process is automatically triggered. Construction of an undirected graph data model for the computational domain: Abstract all leaf nodes of the current octree into a parallel computational undirected graph model: Vertex: Each leaf grid node corresponds to a graph vertex; Edge: The topological connection between adjacent grids corresponds to an undirected edge in the graph; Vertex weight: The computational cost of a single node.

10. The method for three-dimensional design and optimization of a centrifugal pump impeller according to claim 9, characterized in that, Also includes: With the goal of ensuring that the sum of the vertex weights in each parallel partition is equal, the undirected graph is re-partitioned. Pack the node IDs, coordinates, physical quantity arrays, pointer information, and topological relationships of cross-boundary meshes into standardized data packets; Data message transmission is accomplished through network communication protocols; The receiving end completes data unpacking, re-establishes octree pointer binding and memory mapping, and verifies the load balancing effect.

Images