A graphics memory constraint graph neural network block reasoning method for rapid prediction of aero-engine turbine blade heat flow field

Abstract

This invention provides a memory-constrained graph neural network block inference method for rapid prediction of the thermal flow field of aero-engine turbine blades, relating to the field of aero-engine turbine blade design. The method includes: acquiring turbine blade mesh data and converting it into graph structure data; estimating the memory usage of a single subgraph inference, determining the number of core nodes and the number of neighborhood expansion layers; performing boundary-sensitive block division of the graph data based on boundary-sensitive weights; constructing an inference subgraph by expanding the neighborhood nodes of each core block outwards as peripheral expansion nodes; after each subgraph is inferred by the graph neural network, extracting the prediction results of the core nodes and scattering them back to the complete mesh according to the original node numbers; performing consistency checks and corrections on the joint areas of adjacent blocks, and verifying engineering physical quantities, outputting the complete three-dimensional thermal flow field prediction results of the turbine blade. This invention enables rapid prediction of the thermal flow field of large-scale turbine blades under limited memory conditions, ensuring full-domain node coverage and block boundary continuity.

Description

Technical Field

[0001] This invention relates to the fields of aero-engine turbine blade design, computational fluid dynamics simulation, artificial intelligence proxy models, and graph neural network technology, specifically to a memory-constrained graph neural network block inference method for rapid prediction of the flow and heat field of aero-engine turbine blades. Background Technology

[0002] Aero-engine turbine blades operate under conditions of high temperature, high pressure, high speed, strong thermal gradients, and complex aerodynamic load coupling. Their aerodynamic performance, cooling capacity, and structural reliability directly affect the engine's thrust-to-weight ratio, thermal efficiency, lifespan, and safety. Traditional turbine blade design typically relies on high-fidelity computational fluid dynamics simulations to obtain pressure, temperature, velocity, density within the blade passage, and the thermal load distribution on the blade surface. When there are many design variables, operating condition variables, or cooling structure parameters, repeated geometric modeling, mesh generation, boundary condition setting, numerical iteration, and post-processing analysis are required, resulting in long design cycles and high computational costs.

[0003] In recent years, graph neural networks have been able to directly process unstructured meshes and complex geometries, making them suitable for learning complex flow and heat transfer mappings on the nodes and edge connections of CFD (Computational Fluid Dynamics) meshes. Unlike convolutional neural networks, which require interpolating complex blades into regular images or voxel spaces, graph neural networks can preserve the topological connections and local geometric relationships of unstructured meshes, making them suitable for predicting three-dimensional flow and heat fields in complex regions such as turbine blades, film cooling holes, endwalls, and tip clearances.

[0004] However, real turbine blade CFD meshes typically contain a large number of nodes and edges. Graph neural networks need to preserve node hidden features, edge hidden features, and intermediate activations during multi-step message passing. If full-graph inference is performed directly on the complete graph, the memory requirements may exceed the memory limit of a single GPU (Graphics Processing Unit). To reduce memory usage, existing methods may employ random sampling, uniform downsampling, or simple block inference. However, random sampling cannot guarantee that all nodes in the entire region are predicted; uniform downsampling will lose key local structures such as near-wall surfaces, cooling hole outlets, wakes, and secondary flows; and simple block inference is prone to missing neighborhood information at block boundaries, leading to discontinuities in the cloud map such as patches, jagged edges, or color intrusion.

[0005] Furthermore, turbine blades exhibit significant engineering uniqueness compared to general external flow problems. Their flow field is simultaneously influenced by inlet total pressure and temperature, outlet pressure, rotational speed, periodic boundaries, wall thermal boundaries, cooling gas supply, endwall secondary flow, tip leakage flow, and film gas jet. If the segmentation boundary cuts through cooling holes, near-wall high gradient regions, or wake regions, it will further amplify the segmentation error. Therefore, a graph neural network segmentation inference method for turbine blades is needed that can adapt to memory constraints while maintaining global node coverage and segmentation seam continuity. Summary of the Invention

[0006] In view of this, this application provides a memory-constrained graph neural network block inference method for rapid prediction of flow thermal field of aero-engine turbine blades, in order to solve the problems of excessive memory consumption of full graph inference of large-scale turbine blade three-dimensional unstructured mesh graph neural network, inability of random sampling prediction to cover all nodes, easy generation of discontinuity between blocks in simple block prediction, unreasonable segmentation of complex boundaries and high gradient regions of turbine blades, and difficulty in directly using prediction results for engineering post-processing.

[0007] This method is applicable to the rapid prediction of pressure field, temperature field, velocity field, density field, wall temperature field, cooling efficiency field, heat transfer coefficient field, or heat flux density field of objects such as high-pressure turbine guide vanes, turbine moving blades, blades with tip clearance, blades with endwall structures, and blades with internal cooling or film cooling structures.

[0008] This application provides the following technical solution: a memory-constrained graph neural network block inference method for rapid prediction of flow thermal field of aero-engine turbine blades, including: Acquire three-dimensional computational fluid dynamics mesh data of turbine blades, and convert the mesh data into graph structure data. The graph structure data includes nodes and edges. Each computational fluid dynamics mesh node corresponds to a graph node, and the connection relationship between mesh nodes corresponds to graph edges. Based on the available video memory of the graphics processor, the number of parameters of the graph neural network model, the number of message passing steps, and the graph size of the graph structure data, the video memory usage of a single subgraph inference is estimated; the estimated video memory usage is compared with a preset safety threshold, and the number of core nodes and the number of neighborhood expansion layers of each subgraph are adaptively adjusted and determined according to the comparison results, so that the video memory usage of a single subgraph inference does not exceed the preset safety threshold. Calculate the boundary sensitivity weight for each node in the graph structure data, and perform boundary sensitivity segmentation on the graph structure data based on the boundary sensitivity weight to divide all nodes of the graph structure data into multiple non-overlapping core segments; For each core block, the neighborhood expansion layer is extended outward along the graph edge, and the neighboring nodes obtained from the expansion are used as peripheral expansion nodes; an inference subgraph containing core nodes and peripheral expansion nodes is constructed, and the inference subgraph is input into the graph neural network for inference. During the inference process, the peripheral expansion nodes participate in message passing but do not output prediction results. Extract the inference prediction results of the core nodes in each inference subgraph, scatter the prediction results back to the complete turbine blade mesh according to the original node number, and reconstruct the predicted physical quantities on the complete turbine blade mesh; The predicted physical quantities of adjacent core segment joint areas are continuously checked and corrected to ensure that the prediction results of the whole domain meet the joint consistency and obtain the complete prediction results after reconstruction. The complete prediction results are then checked for consistency of engineering physical quantities. After the check is passed, the complete three-dimensional flow and thermal field prediction data of the turbine blade is output.

[0009] According to one embodiment of this application, the memory usage for single-time subgraph inference is estimated using the following formula: The memory usage estimation method for the single subgraph inference is as follows:

[0010] in, This represents the estimated memory usage of the b-th subgraph, where γ is an empirical coefficient. Indicates the number of nodes in the subgraph. Indicates the number of edges in the subgraph. Indicates the dimension of the node's input features. Indicates the dimension of the input feature. Let represent the hidden dimension of the graph neural network, K represent the number of message passing steps, and L represent the number of layers in the multilayer perceptron.

[0011] According to one embodiment of this application, the boundary sensitivity weight is calculated using the following formula:

[0012] in, For nodes Boundary-sensitive weights Represents a node Whether it belongs to the blade wall region is a binary indicator. Represents a node Whether it belongs to the air film pore region is a binary indicator quantity. Represents a node Whether it belongs to the leaf tip region is a binary indicator. Represents a node Whether it belongs to the end-wall region (binary indicator); and Let be the pressure and temperature at node i, respectively. For gradient operators, Represents a node Pressure gradient modulus at the location, Represents a node Temperature gradient modulus at that location to These are the weighting coefficients.

[0013] According to one embodiment of this application, the node set of the reasoning subgraph is:

[0014] in, Indicates the first A set of nodes in a reasoning subgraph. Indicates the first A set of nodes for each core block. Indicates outward expansion The set of peripheral extended nodes obtained from the layer neighborhood.

[0015] According to one embodiment of this application, the number of neighborhood expansion layers... The relationship between the number of message passing steps K in the graph neural network and the following is satisfied: ≥K; the number of neighborhood expansion layers The subgraph is adaptively determined based on the type of the region it belongs to.

[0016] According to one embodiment of this application, the seam consistency check includes: detecting the continuity of predicted physical quantities in the boundary region of adjacent core blocks, calculating the seam error of the boundary region, and triggering correction when the seam error exceeds a preset seam error threshold; the correction method includes increasing the number of neighborhood expansion layers of the corresponding boundary region and then re-inferring, moving the block boundary to a low gradient region and then re-blocking, performing weighted fusion of prediction results within the seam zone, performing local secondary inference on abnormal regions, and marking abnormal regions as at least one of the regions requiring computational fluid dynamics verification.

[0017] According to one embodiment of this application, the seam error between adjacent core block a and core block b is... The calculation method is as follows:

[0018] in, This represents the predicted physical quantity gradient on the seam side of core block a. This represents the predicted physical quantity gradient on the joint side of core block b. Represents the L2 norm, This indicates a positive number that prevents the denominator from being zero.

[0019] According to one embodiment of this application, the engineering physical quantity includes at least one of the following: outlet mass flow rate, pressure ratio, outlet average total pressure, outlet average total temperature, blade surface average temperature, maximum wall temperature, wall heat flux density, cooling efficiency, and temperature non-uniformity; when the engineering physical quantity deviates from a preset threshold, at least one of the following operations is performed: increasing the number of neighborhood expansion layers and then re-predicting, physical constraint correction, or marking it as requiring computational fluid dynamics verification.

[0020] According to one embodiment of this application, in the graph structure data, the node features of each node include at least one of the following: node three-dimensional coordinates, node type identifier, boundary condition parameters, working condition parameters, and wall distance; the edge features of each edge include at least one of the following: edge relative coordinates, edge length, and direction vector.

[0021] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: First, this invention transforms large-scale turbine blade full-graph inference into multiple controllable-scale subgraph inferences through memory budget-driven core-halo block inference, significantly reducing the GPU memory requirement per run and enabling ordinary engineering graphics cards to complete the prediction of three-dimensional flow and thermal fields of turbine blades with millions of nodes.

[0022] Secondly, this invention uses full-node core partitioning and core-only scatter reconstruction to ensure that each original grid node is predicted and written only once, avoiding node omissions caused by random sampling prediction and duplicate writing caused by simple overlapping partitioning.

[0023] Third, this invention uses halo nodes to participate in message passing, providing core nodes with cross-block neighborhood context, which can reduce errors caused by missing neighborhood information near block boundaries.

[0024] Fourth, this invention introduces a turbine blade boundary sensitive weight, which makes the segment boundary avoid the wall, film air hole, blade tip, end wall, wake and high gradient region as much as possible, which is beneficial to maintain the prediction continuity of cooling jet, end wall secondary flow, blade tip leakage flow and near wall thermal boundary layer.

[0025] Fifth, this invention uses a joint consistency check and correction mechanism to automatically diagnose the predicted physical quantities and gradients at the junctions of blocks, which can reduce phenomena such as cloud map patches, blocky faults and color intrusion, and improve the visualization quality and engineering credibility of the prediction results.

[0026] Sixth, this invention enables the graph neural network prediction results to directly participate in turbine blade design evaluation, scheme selection, optimization iteration and post-processing analysis through engineering physical quantity consistency verification and original CFD file write-back mechanism, thereby shortening the turbine blade flow and heat performance evaluation cycle. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the overall process of rapid prediction of the thermal field of turbine blades in this invention. Figure 2 This is a flowchart of the memory-constrained core-halo block inference process of the present invention; Figure 3 This is a flowchart of the joint consistency inspection and correction process of the present invention; Figure 4 This is a flowchart illustrating the engineering write-back and verification process for the prediction results of this invention. Detailed Implementation

[0029] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0030] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Without departing from the core idea of ​​this invention, those skilled in the art can make equivalent substitutions or adjustments according to the specific turbine blade type, mesh size, model structure, memory conditions, and post-processing software.

[0031] This invention provides a memory-constrained graph neural network block inference method for rapid prediction of flow thermal fields in aero-engine turbine blades, including: Acquire three-dimensional computational fluid dynamics mesh data of turbine blades, and convert the mesh data into graph structure data. The graph structure data includes nodes and edges. Each computational fluid dynamics mesh node corresponds to a graph node, and the connection relationship between mesh nodes corresponds to graph edges. Based on the available video memory of the graphics processor, the number of parameters of the graph neural network model, the number of message passing steps, and the graph size of the graph structure data, the video memory usage of a single subgraph inference is estimated; the estimated video memory usage is compared with a preset safety threshold, and the number of core nodes and the number of neighborhood expansion layers of each subgraph are adaptively adjusted and determined according to the comparison results, so that the video memory usage of a single subgraph inference does not exceed the preset safety threshold. Calculate the boundary sensitivity weight for each node in the graph structure data, and perform boundary sensitivity segmentation on the graph structure data based on the boundary sensitivity weight to divide all nodes of the graph structure data into multiple non-overlapping core segments; For each core block, the neighborhood expansion layer is extended outward along the graph edge, and the neighboring nodes obtained from the expansion are used as peripheral expansion nodes; an inference subgraph containing core nodes and peripheral expansion nodes is constructed, and the inference subgraph is input into the graph neural network for inference. During the inference process, the peripheral expansion nodes participate in message passing but do not output prediction results. Extract the inference prediction results of the core nodes in each inference subgraph, scatter the prediction results back to the complete turbine blade mesh according to the original node number, and reconstruct the predicted physical quantities on the complete turbine blade mesh; The predicted physical quantities of adjacent core segment joint areas are continuously checked and corrected to ensure that the prediction results of the whole domain meet the joint consistency and obtain the complete prediction results after reconstruction. The complete prediction results are then checked for consistency of engineering physical quantities. After the check is passed, the complete three-dimensional flow and thermal field prediction data of the turbine blade is output.

[0032] This invention discloses a memory-constrained graph neural network block inference method for rapid prediction of the thermal flow field of aero-engine turbine blades. The method first reads the turbine blade's three-dimensional computational fluid dynamics mesh, boundary conditions, geometric parameters, operating parameters, and high-fidelity simulation field quantities, converting the unstructured mesh into graph data containing node and edge features. Then, based on the graph neural network model size, message passing steps, number of nodes and edges, and the target graphics processor's memory budget, it automatically determines the core node size and neighborhood expansion layer number of each prediction subgraph. Further, it combines information from the turbine blade wall, endwall, tip, film canopy, wake, and high-gradient region to perform boundary-sensitive connectivity block division of the complete mesh graph. For each core block, a halo neighborhood is expanded outwards and input into the graph neural network for inference. Halo nodes only participate in message passing and do not participate in writing the final result. The prediction results of each core node are scattered back to the complete mesh according to the original node number, and consistency checks and corrections are performed on the joint areas of adjacent blocks and engineering integral quantities to obtain a complete and continuous prediction result of the three-dimensional thermal flow field of the turbine blade. This invention enables rapid prediction of the three-dimensional flow and thermal field of large-scale turbine blades under limited video memory conditions, reduces the video memory requirements for full-graph inference, improves the problem of discontinuity at block boundaries, and facilitates writing the prediction results back to engineering post-processing files such as CGNS (Common Data Format for Computational Fluid Dynamics), VTK (Visualization Toolkit Data Format), or Tecplot (Engineering Science Visualization Software).

[0033] This invention achieves rapid prediction of the complete three-dimensional flow and thermal field of turbine blades under limited GPU memory conditions through memory budget-driven adaptive core-halo (core-peripheral expansion) block division, turbine blade boundary-sensitive block division, halo message passing and core-only (core node only) output, global scatter (scatter write-back) reconstruction, joint consistency correction and engineering physical quantity verification. This improves the availability of AI (artificial intelligence) prediction results in turbine blade design and post-processing.

[0034] In specific implementation, such as Figures 1-4 As shown, the method of the present invention adopts the following technical solution: Step S1: Obtain high-fidelity simulation data of the turbine blades (the three-dimensional computational fluid dynamics mesh data). Read the turbine blade three-dimensional CFD mesh file, node coordinates, element or node connection relationships, boundary types, geometric design parameters, operating parameters, and high-fidelity fluid and thermal field physical quantities required for training. Boundary types include at least one of the following: mains inlet, outlet, blade wall, endwall, hub, casing, periodic surface, blade tip region, cooling hole inlet, cooling hole outlet, and film cooling hole wall. Physical quantities include at least one of the following: density, static pressure, static temperature, velocity component, wall temperature, cooling efficiency, heat transfer coefficient, and heat flux density.

[0035] Step S2: Construct turbine blade graph data. Using CFD mesh nodes as graph nodes, and adjacent mesh nodes, cell edges, face adjacency relationships, or nearest neighbor relationships obtained through distance threshold search as graph edges, the graph structure is obtained as follows:

[0036] in, This represents turbine blade diagram data. Represents a set of nodes. This represents the set of edges. i and j both represent node numbers; each node... Node features Includes at least one of the following: 3D coordinates, node type, boundary condition parameters, operating condition parameters, geometric parameters, initial field quantities, wall distance, and region identifier; each edge edge features It includes at least one of the following: relative coordinates, side length, direction vector, local scale, and whether it crosses a special boundary region.

[0037] Step S3: Establish a subgraph size estimation mechanism driven by GPU memory budget. Based on available GPU memory, the hidden dimension of the graph neural network model, the number of message passing steps, the number of MLP (Multilayer Perceptron) layers, the node feature dimension, the edge feature dimension, the number of nodes and edges in the subgraph, estimate the GPU memory usage for inference in a single core-halo subgraph:

[0038] in, Indicates the first The estimated memory usage of each subgraph, where b represents the subgraph number; These represent empirical coefficients related to data type, automatic mixed precision, intermediate activation caching, and framework overhead. Indicates the number of nodes in the subgraph. Indicates the number of edges in the subgraph. Indicates the dimension of the node's input features. Indicates the dimension of the input feature. This represents the hidden dimension of a graph neural network. Indicates the number of steps in message passing. This indicates the number of MLP layers. The number of core nodes and the number of Halo extension layers are determined based on memory safety constraints.

[0039] in, This indicates that the GPU has available video memory. This indicates the video memory safety factor. If the estimated video memory exceeds... If the memory is sufficient, reduce the number of core nodes or decrease the size of the subgraph in one inference step; if there is sufficient memory, increase the number of core nodes or increase the number of halo layers to improve inference efficiency and boundary continuity.

[0040] Step S4: Construct turbine blade boundary-sensitive weights. To avoid block boundaries crossing high-gradient and strongly coupled regions, for each node... Calculate the boundary-sensitive weights:

[0041] in, Represents a node Boundary-sensitive weights Represents a node Does it belong to the blade wall or near-wall region? Represents a node Does it fall within the area affected by film cooling holes or cooling jets? Represents a node Is it located in the blade tip or gap leakage area? Represents a node Whether it belongs to the end wall or secondary flow influence area; the above four quantities represented by I are all binary region indicators, which are 1 when the node belongs to the corresponding region, and 0 otherwise; and Let i represent the pressure and temperature at node i, respectively. Represents the gradient operator. Represents a node Pressure gradient modulus Represents a node Temperature gradient mode to These are the weighting coefficients. If the gradient of the predicted field is unknown, it can be approximated by the statistical gradient of the training samples, the initial field gradient, the rate of change of the neighborhood boundary, or the geometric curvature.

[0042] Step S5: Generate core blocks with full node coverage. Based on BFS (Breadth-First Search) connectivity expansion, graph partitioning, or region growth strategies, the complete node set is... Divide the original node into multiple non-overlapping sets of core nodes, such that each original node appears exactly once as a core node: ; in, This represents the total number of blocks, where a and b represent the numbers of the different blocks. Indicates the first A set of core nodes for each block. and These represent the core node sets for different blocks. During the block partitioning process, priority is given to ensuring that the high gradient regions such as the wall surface, cooling hole outlet, blade tip leakage, endwall secondary flow, and wake are not cut off, or that the block boundaries fall as close as possible to low gradient and low weight regions.

[0043] Step S6: Construct the core-halo subgraph. Divide each core into blocks. Expand outwards according to the adjacency relationship shown in the diagram. Layer neighbor nodes serve as halo nodes, forming a subgraph node set used for reasoning:

[0044] in, Indicates the first A set of nodes for a reasoning subgraph Indicates the first A set of halo nodes in each block. This indicates the number of Halo extension layers. Halo nodes are used to provide messaging context to core nodes. Preferably, Message passing steps in graph neural networks satisfy:

[0045] When video memory is insufficient to satisfy the above relationship, it is permissible to use The approximate reasoning is used, and the block boundary error is compensated by seam consistency correction in step S9.

[0046] Step S7, execute the graph neural network block inference. The first... A graph neural network trained on inputs from core-halo subgraphs. The graph neural network includes node encoders, edge encoders, message passing processors, node decoders, and an inverse normalization module. In the first... In this message transmission, the node The hidden features are updated as follows:

[0047] Where k represents the message passing iteration number; Represents a node In the Hidden features following the next message transmission Represents a node The set of neighboring nodes, Representing neighboring nodes In the Hidden features following the next message transmission Represents a node With nodes Edge features between them This indicates the edge message update function. This represents the node state update function. (After...) After this message transmission, the node decoder outputs the subgraph prediction result:

[0048] in, Indicates the first Predicted physical quantities of each subgraph Indicates the training parameters Deterministic graph neural networks, Indicates the first A core-halo subgraph.

[0049] Step S8: Perform halo discarding and core-only output. Since halo nodes may be repeatedly contained in multiple adjacent subgraphs, and their main function is to provide messaging context, only the prediction results of core nodes are retained.

[0050] in, Indicates the first The prediction results for each core block. This represents the portion extracted from the subgraph prediction results by the core node index.

[0051] Step S9: Perform global scatter reconstruction based on the original node numbers. Write the prediction results of each core block back to the complete turbine blade mesh:

[0052] in, This represents the predicted physical quantities on the complete turbine blade mesh. Indicates the first The core node set is the node number in the original complete mesh. Since all original nodes appear exactly once as core nodes, there are no missing or duplicate nodes in the global reconstruction result.

[0053] Step S10: Perform seam consistency check and correction. Continuity detection is performed on the predicted physical quantities and gradients within the boundary areas of adjacent blocks or the preset inspection zone. Adjacent blocks are defined. and Seam tolerance:

[0054] in, Indicates block With block The seam error between them Indicates block Predict the gradient of physical quantities on the seam side. Indicates block Predict the gradient of physical quantities on the seam side. Represents the L2 norm, This prevents positive numbers with a denominator of zero. This indicates the seam error threshold. When... Greater than the preset threshold When this occurs, perform at least one of the following corrections: increase the number of halo layers in the local region and then re-infer; move the block boundary to a low gradient region and then re-block; perform weighted fusion of the results within the seam inspection zone; perform local secondary inference on the abnormal region; mark the region as one that needs CFD verification.

[0055] Step S11: Perform consistency verification of engineering physical quantities. Calculate the mass flow rate, pressure ratio, average total pressure at the outlet, average total temperature at the outlet, average blade surface temperature, maximum wall temperature, wall heat flux density, cooling efficiency, temperature non-uniformity, and cross-sectional average quantity based on the reconstructed complete prediction results. Taking mass flow rate as an example, the predicted mass flow rate at the outlet cross-section is:

[0056] in, This represents the mass flow rate at the outlet section, where q represents the patch or integration point number. This indicates the number of discrete patches or integration points at the exit section. Indicates the first Density at a patch or integration point Indicates the first The velocity vector at each surface patch or integration point This represents the outward normal unit vector at that location. This indicates the area corresponding to the patch or integral weight. When the deviation of the engineering physical quantity exceeds the preset threshold, it triggers halo increase, local re-prediction, physical constraint correction, or high-fidelity CFD verification.

[0057] Step S12: Output and write back to the post-processing file. Save the complete prediction results in a CSV (comma-separated values), NPZ (numerical array compressed file), VTK, CGNS, or Tecplot readable format. Preferably, use the original CFD file as a template, retaining the original mesh, boundaries, partitions, node order, and existing post-processing metadata, only replacing or adding predicted physical quantity data, thus ensuring that the prediction results can be directly viewed in Tecplot, STAR-CCM+ (commercial computational fluid dynamics software), ParaView (open-source scientific visualization software), or other post-processing software.

[0058] Example 1: Rapid Prediction of Three-Dimensional Flow Field in a High-Pressure Turbine Guide Blade Passage This embodiment uses the prediction of a single-channel three-dimensional steady flow field of a high-pressure turbine guide vane for an aero-engine as an example. The vane channel includes the main flow inlet, outlet, suction surface, pressure surface, endwall, periodic surface, and wake region. High-fidelity simulation uses an unstructured mesh with a large number of nodes, and outputs physical quantities including density, static pressure, static temperature, and velocity components in three directions.

[0059] Step 1: Construct a sample database. Multiple turbine blade channel samples are generated using CFD software for different inlet total pressure, inlet total temperature, outlet static pressure, inlet turbulence intensity, and geometric installation angle conditions. Each sample saves the original CGNS file, node coordinates, boundary type, flow field physical quantities, and operating parameters. Training, validation, and test sets are divided according to design schemes to avoid the same geometric scheme appearing in both the training and test sets.

[0060] Step 2: Construct graph data. Read the GridCoordinates, FlowSolution, and boundary Section information from CGNS, treating each grid node as a graph node. Node characteristics include coordinates, node type one-hot encoding, inlet total pressure, inlet total temperature, outlet pressure, initial density, initial temperature, and initial velocity. If a node is located on the blade wall, endwall, inlet, outlet, or periodic surface, it is identified in the node type. Establish undirected edges based on the element connection relationships; each edge's characteristics include relative coordinates, edge length, and direction vector.

[0061] Step 3: Train the graph neural network model. A graph neural network including node encoders, edge encoders, message passing modules, and node decoders is used. Node and edge inputs are first normalized. Target physical quantities include density, hydrostatic pressure, hydrostatic temperature, and three-dimensional velocity components. The training loss uses the weighted mean square error of the normalized physical quantities, while the validation and testing errors use the inversely normalized relative root mean square error and absolute root mean square error. Through training, model parameters suitable for predicting the three-dimensional flow field of turbine blades are obtained. .

[0062] Step 4: Determine the memory constraint block parameters. During the prediction phase, read the available GPU memory. Based on the model's hidden dimensions Message passing steps MLP layer number Current sample node count and edge density estimation for a single block of video memory. If the estimated video memory usage for full-graph inference exceeds the safety threshold... If so, the block inference mode will be automatically selected. Assuming the model message passing steps are 10, the number of Halo layers will be set to 10 first; if there is insufficient GPU memory, the number of core nodes will be reduced first, while still ensuring as much memory as possible. .

[0063] Step 5: Generate boundary-sensitive core blocks. Calculate node weights based on the blade wall, endwall, wake region, pressure gradient, and temperature gradient. When partitioning, the partition boundaries are arranged first from low-weight regions to avoid cutting off the high gradient region at the suction front edge, the wake region at the blade trailing edge, and the secondary flow region at the endwall. All nodes are divided into multiple non-overlapping sets of core nodes, and each node appears only once as a core node.

[0064] Step 6: Construct the core-halo subgraph and perform inference. For each core block, expand the halo neighborhood outward along the graph connectivity to generate a subgraph containing core nodes and halo nodes. Feed this subgraph into the graph neural network model for inference. Halo nodes participate in message passing, providing neighborhood context for core nodes, but do not participate in the final output.

[0065] Step 7: Perform core-only reconstruction. For each subgraph, extract only the predicted values ​​of the core nodes and write back the complete node sequence according to the original CGNS node numbers. Since each core block does not overlap and covers all nodes, a complete prediction array with the same number of nodes as the original CFD grid is finally obtained.

[0066] Step 8, perform a seam consistency check. Check the pressure, temperature, and velocity gradients near the seams of adjacent sections. If there are seam errors... Exceeding the threshold If the problem persists, the number of Halo layers in the local area is increased and the prediction is recalculated, or a weighted fusion of the seam area is performed. After correction, the seam error is recalculated until the threshold requirement is met or the area is marked as requiring CFD verification.

[0067] Step 9: Perform post-processing. Calculate engineering parameters such as outlet mass flow rate, cross-sectional average static pressure, cross-sectional average total temperature, and pressure ratio based on the complete prediction results. If the deviation compared to the training samples or high-fidelity CFD verification results is within the set range, write the prediction results back to the original CGNS file. When writing back, retain the original mesh, boundary, and Solution (solution data node) structure, only replacing or adding density, pressure, temperature, and velocity fields so that Tecplot or STAR-CCM+ can directly open and view them.

[0068] Through the above steps, this embodiment achieves its goal without performing a full... Figure 1 In the case of one-time inference, the complete three-dimensional turbine guide vane flow field prediction results are obtained. Compared with random sampling prediction, this method can cover all nodes; compared with simple block prediction, this method reduces inter-block discontinuities through halo message passing and seam correction; compared with full graph inference, this method significantly reduces single-time memory peaks.

[0069] Example 2: Prediction of thermal field of turbine blade with multiple exhaust film holes This embodiment focuses on a turbine blade with multiple film cooling holes. The predictions include the velocity field, pressure field, temperature field, and overall cooling efficiency of the blade wall in the mixing region of the main flow and the coolant. The object includes multiple film cooling hole inlets, in-hole channels, hole outlets, blade wall, endwall, and downstream cooling gas coverage area, with complex local geometry and large gradients in physical quantities.

[0070] Step 1: Establish a film cooling model. For different orifice types such as circular, fan-shaped, Y-shaped (branched), or horseshoe-shaped orifices, vary the orifice diameter, inclination angle, compound angle, orifice spacing, airflow ratio, density ratio, and mainstream Mach number. Conduct high-fidelity CFD or fluid-structure interaction simulations to obtain the physical fields of the blade's fluid and solid domains. For the conjugate heat transfer model, simultaneously preserve the fluid temperature, solid temperature, wall heat flux density, and overall cooling efficiency.

[0071] Step 2: Region Identification. During the graph data construction phase, node type identifiers are set for the air film inlet, inner wall surface of the inlet, outlet of the inlet, near-wall cooling film coverage area, endwall, and blade tip region. For nodes within a certain characteristic length range downstream of the inlet outlet, [the following is missing from the original text] For several layers of mesh nodes in the blade wall normal direction, set These region identifiers are used for subsequent boundary-sensitive weight calculations.

[0072] Step 3: Perform boundary-sensitive segmentation. Due to the strong jet mixing and temperature gradient near the outlet of the film gas vent, segmentation should prioritize avoiding dividing the jet core region of the same vent outlet into multiple core blocks. If the number of nodes in a local area of ​​a film gas vent is less than the upper limit of a single core block, then the vent outlet, the channel inside the vent, and the near-wall jet region should be placed in the same core block as much as possible; if the number of nodes exceeds the upper limit, then the segmentation boundary should be arranged along the low temperature gradient region or the downstream region far from the vent outlet.

[0073] Step 4: Perform memory-adaptive halo expansion. For cooling holes and near-wall regions, due to strong local physical coupling, a larger number of halo layers is preferred; for mainstream regions far from the wall where the flow field changes gently, a smaller number of halo layers can be used. In other words, It can adapt to the type of subgraph region, rather than using a fixed value for all blocks.

[0074] Step 5: Perform block prediction and seam correction. Perform graph neural network inference on each core-halo subgraph, retaining only the prediction results of the core nodes. For the seam region downstream of the orifice outlet, focus on checking the temperature gradient, cooling efficiency gradient, and velocity shear layer continuity. If a break in the cooling film covering zone or color intrusion is found, perform local secondary inference or expand the halo for that region and then re-predict.

[0075] Step 6: Calculate cooling performance indicators. Calculate the overall cooling efficiency based on the predicted wall temperature and the reference temperatures of the mains and cool air. This can be done in the following format:

[0076] in, Indicates overall cooling efficiency. Indicates the mainstream gas reference temperature. This indicates the blade wall temperature. This represents the reference temperature of the cooling air. It calculates the average integrated cooling efficiency, maximum wall temperature, and temperature non-uniformity across the blade surface area, and uses these as design evaluation metrics.

[0077] This embodiment demonstrates that the present invention is applicable to complex turbine blades with film cooling structures. By combining a segmentation strategy with turbine blade boundaries, cooling holes, and high-gradient regions, the continuity of local cooling flow and heat transfer prediction can be improved while reducing memory usage.

[0078] Example 3: Adaptive Prediction Process under Limited Memory Conditions This embodiment illustrates the adaptive execution method of the present invention under different memory conditions. For the same turbine blade test sample, full graph inference is first attempted to estimate the memory requirements. If If the condition is met, then full-graph reasoning is performed; otherwise, block reasoning is performed.

[0079] On a 24 GB GPU, if the estimated memory usage for full-graph inference exceeds a safe threshold, the program automatically sets the number of core nodes to the maximum that the memory can handle and prioritizes setting the number of halo layers to the number of message passing steps. If the number of halo layers is too large and a single block of memory still exceeds the limit, the program further reduces the number of core nodes. If the limit is still exceeded after reducing the number of core nodes, a number of halo layers smaller than the number of message passing steps is allowed, and high-error areas are locally re-inferenced during the seam check phase.

[0080] On GPUs with 48 GB or 80 GB of VRAM, the program can automatically increase the number of core nodes to reduce the number of blocks, or increase the number of halo layers to improve boundary continuity. In this way, the same prediction program can adapt to different hardware conditions without requiring the user to manually set block parameters multiple times through trial and error.

[0081] This embodiment demonstrates the memory constraint adaptive feature of the present invention. This mechanism not only reduces the risk of memory overflow, but also improves the automation and engineering deployability of the large-scale graph neural network prediction process for turbine blades.

[0082] This invention can be applied to aerodynamic design, cooling design, thermal protection design, fluid-thermal coupling performance evaluation, and rapid selection of multiple solutions for aero-engine turbine blades. The method can be embedded into turbine blade parametric design platforms, automated CFD simulation platforms, surrogate model optimization platforms, or digital R&D processes to quickly obtain three-dimensional flow field, temperature field, and cooling performance prediction results, and can form a closed loop with high-fidelity CFD verification and experimental validation. This invention can also be extended to compressor blades, combustion chamber flame tubes, turbine disks, sealing structures, blade passages, and other complex unstructured mesh fluid-thermal problems.

[0083] 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 technical scope 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.

Claims

1. A graphics memory constraint graph neural network block inference method for rapid prediction of aero-engine turbine blade heat flow field, characterized in that, include: Acquire three-dimensional computational fluid dynamics mesh data of turbine blades, and convert the mesh data into graph structure data. The graph structure data includes nodes and edges. Each computational fluid dynamics mesh node corresponds to a graph node, and the connection relationship between mesh nodes corresponds to graph edges. Based on the available video memory of the graphics processor, the number of parameters of the graph neural network model, the number of message passing steps, and the graph size of the graph structure data, the video memory usage of a single subgraph inference is estimated. The estimated memory usage is compared with a preset safety threshold. Based on the comparison result, the number of core nodes and the number of neighborhood expansion layers for each subgraph are adaptively adjusted and determined so that the memory usage of a single subgraph inference does not exceed the preset safety threshold. Calculate the boundary sensitivity weight for each node in the graph structure data, and perform boundary sensitivity segmentation on the graph structure data based on the boundary sensitivity weight to divide all nodes of the graph structure data into multiple non-overlapping core segments; For each core block, the neighborhood expansion layer is extended outward along the graph edge, and the neighboring nodes obtained from the expansion are used as peripheral expansion nodes; an inference subgraph containing core nodes and peripheral expansion nodes is constructed, and the inference subgraph is input into the graph neural network for inference. During the inference process, the peripheral expansion nodes participate in message passing but do not output prediction results. Extract the inference prediction results of the core nodes in each inference subgraph, scatter the prediction results back to the complete turbine blade mesh according to the original node number, and reconstruct the predicted physical quantities on the complete turbine blade mesh; The predicted physical quantities of adjacent core block seam areas are continuously checked and corrected to ensure that the prediction results of the whole domain meet the seam consistency and obtain the complete prediction results after reconstruction. The complete prediction results are verified for consistency of engineering physical quantities. Once the verification is passed, the complete three-dimensional flow and thermal field prediction data of the turbine blade is output.

2. The method of claim 1, wherein, The memory usage for single-order subgraph inference is estimated using the following formula: The memory usage estimation method for the single subgraph inference is as follows: wherein, represents the estimated GPU memory usage of the b-th subgraph, and γ is an empirical coefficient, represents the number of nodes of the subgraph, represents the number of edges of the subgraph, represents the input feature dimension of the node, represents the input feature dimension of the edge, represents the hidden dimension of the graph neural network, K represents the number of message passing steps, and L represents the number of multi-layer perceptron layers.

3. The method of claim 1, wherein, The boundary sensitivity weights are calculated using the following formula: in, For nodes Boundary-sensitive weights Represents a node Whether it belongs to the blade wall region is a binary indicator. Represents a node Whether it belongs to the air film pore region is a binary indicator quantity. Represents a node Whether it belongs to the leaf tip region is a binary indicator. Represents a node Whether it belongs to the end-wall region (binary indicator); and Let be the pressure and temperature at node i, respectively. For gradient operators, Represents a node Pressure gradient modulus at the location, Represents a node Temperature gradient modulus at that location to These are the weighting coefficients.

4. The method according to claim 1, characterized in that, The node set of the reasoning subgraph is: in, Indicates the first A set of nodes in a reasoning subgraph. Indicates the first A set of nodes for each core block. Indicates outward expansion The set of peripheral extended nodes obtained from the layer neighborhood.

5. The method according to claim 4, characterized in that, Neighborhood expansion layers The relationship between the number of message passing steps K in the graph neural network and the following is satisfied: ≥K; the number of neighborhood expansion layers The subgraph is adaptively determined based on the type of the region it belongs to.

6. The method according to claim 1, characterized in that, The seam consistency check includes: detecting the continuity of predicted physical quantities in the boundary region of adjacent core blocks, calculating the seam error of the boundary region, and triggering correction when the seam error exceeds a preset seam error threshold; the correction methods include increasing the number of neighborhood expansion layers of the corresponding boundary region and then re-inferring, moving the block boundary to a low gradient region and then re-blocking, performing weighted fusion of prediction results within the seam zone, performing local secondary inference on abnormal regions, and marking abnormal regions as at least one of the following: regions requiring computational fluid dynamics verification.

7. The method according to claim 6, characterized in that, Seam error between adjacent core block a and core block b The calculation method is as follows: in, This represents the predicted physical quantity gradient on the seam side of core block a. This represents the predicted physical quantity gradient on the joint side of core block b. Represents the L2 norm, This indicates a positive number that prevents the denominator from being zero.

8. The method according to claim 1, characterized in that, The engineering physical quantities include at least one of the following: outlet mass flow rate, pressure ratio, outlet average total pressure, outlet average total temperature, blade surface average temperature, maximum wall temperature, wall heat flux density, cooling efficiency, and temperature non-uniformity. When an engineering physical quantity deviates from a preset threshold, at least one of the following operations is performed: increasing the number of neighborhood expansion layers and then re-predicting, physical constraint correction, or marking it as requiring computational fluid dynamics verification.

9. The method according to claim 1, characterized in that, In the graph structure data, the node features of each node include at least one of the following: node three-dimensional coordinates, node type identifier, boundary condition parameters, working condition parameters, and wall distance; the edge features of each edge include at least one of the following: edge relative coordinates, edge length, and direction vector.

Images