An automobile external flow field pressure distribution prediction, distributed computing system, device and medium

Abstract

This application provides a distributed computing system, device, and medium for predicting the pressure distribution of an external flow field in a vehicle. Applied to a distributed computing system with multiple computing nodes, the system first obtains a global computational domain for the external flow field of the vehicle, consisting of grid points carrying physical parameters and geometric information. An attention dependency graph is constructed based on the hydrodynamic relationships between grid points, completing the partitioning of the computational domain and assigning it to corresponding computing nodes. Each node generates a candidate set based on its spatial neighborhood, calculates attention weights, and filters to obtain the final attention dependency set. Each node completes cross-node data transmission according to the dependency set, obtains the dependency data required for computation, performs attention-weighted summation within the corresponding range, and finally summarizes the computation results of all nodes to complete the prediction of the pressure distribution of the external flow field in the vehicle. Using this method, the parallel computing efficiency and pressure prediction accuracy of vehicle external flow field pressure simulation under large-scale unstructured grids can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of automotive external flow field simulation and distributed parallel computing technology, specifically to an automotive external flow field pressure distribution prediction, distributed computing system, device and medium. Background Technology

[0002] Automotive external flow field simulation is a crucial tool for developing automotive aerodynamic shapes. By solving for the pressure distribution in the external flow field, aerodynamic performance such as vehicle drag and driving stability can be accurately assessed. Currently, the industry widely employs CFD (Computational Fluid Dynamics) technology for flow field simulation. However, with the continuous increase in simulation mesh size, the computing power of a single computing device is insufficient to meet the solution requirements. Therefore, distributed computing systems are widely used for solving large-scale flow field data. Simultaneously, artificial intelligence models are increasingly being applied to flow field prediction. Among them, the Transformer model, with its excellent feature modeling capabilities, is frequently used for automotive external flow field pressure prediction tasks.

[0003] Currently, the basic distributed Transformer computing scheme used for automotive external flow field simulation adopts a geometrically evenly divided segmentation method. This scheme only uniformly divides the global grid based on the geometric position of the grid space and directly distributes the segmented grid blocks to each computing node. In the model attention calculation stage, a fixed global communication method is used, and all grid feature data are transmitted between computing nodes without difference. Attention weight calculation is completed by relying on global data interaction, and finally the flow field pressure results are obtained by summarizing.

[0004] This traditional geometrically distributed parallel scheme has obvious technical drawbacks. Because the slicing process does not take into account the correlation characteristics of fluid dynamics, it ignores the physical law of local coupling of pressure in the external flow field of the vehicle. This causes a large number of far-field meshes without physical correlation to continuously participate in cross-node communication and calculation, resulting in a massive amount of invalid communication data in the system, which occupies network transmission bandwidth, increases communication latency, and leads to low overall parallel computing efficiency of the distributed system. It is difficult to adapt to the simulation conditions of high-precision, large-scale unstructured meshes in the external flow field of the vehicle. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, device, equipment and storage medium for predicting the pressure distribution of the external flow field of a vehicle, which can effectively improve the parallel computing efficiency and pressure prediction accuracy of the pressure simulation of the external flow field of a vehicle under large-scale unstructured grids.

[0006] In a first aspect, embodiments of this application provide a method for predicting the pressure distribution of an external flow field in a vehicle, applied to a distributed computing system, the distributed computing system comprising multiple computing nodes; the method includes: Obtain the global computational domain defined by the unstructured mesh describing the external flow field of the vehicle, wherein the unstructured mesh contains multiple mesh points carrying physical parameters and geometric information; An attention dependency graph is constructed based on the fluid dynamic correlation between grid points. The global computation domain is divided into multiple sub-computation domains according to the attention dependency graph, and each sub-computation domain is assigned to a corresponding computation node. Each computing node generates a candidate set by combining the candidate generation function of the spatial neighborhood with the grid points within its own sub-computation domain. Attention weights are calculated for each computing node within the candidate set, and the final attention dependency set for each grid point is obtained according to a preset filtering rule. Each computing node initiates cross-node data transmission to grid points belonging to other computing nodes in its final attention dependency set to obtain the dependency data required for computation. Each computing node performs attention weighted summation calculation based on the acquired dependency data, within the scope defined by its own final attention dependency set. The calculation results of each calculation node are summarized to obtain the predicted pressure distribution of the external flow field of the vehicle.

[0007] Optionally, the step of constructing an attention dependency graph based on the hydrodynamic correlation between grid points, dividing the global computational domain into multiple sub-computational domains according to the attention dependency graph, and assigning each sub-computational domain to a corresponding computational node includes: Construct an attention dependency graph with global grid points as vertices and the expected dependency strength between two grid points as edge weights; Based on the edge weights of the attention dependency graph, a graph partitioning algorithm is used to solve the partitioning objective to obtain the solution result. The partitioning objective is to minimize the weighted sum of the weights of high-weight dependency edges across subdomains and the weighted sum of the load imbalance term. Based on the solution results, the global computing domain is divided into multiple sub-computing domains, and each sub-computing domain is assigned to a computing node in the distributed computing system, so that the computing node subsequently processes only the grid points within the assigned sub-computing domain.

[0008] Optionally, the step of generating a candidate set by each computing node for grid points within its own sub-computation domain, in conjunction with a candidate generation function of the spatial neighborhood, includes: For each grid point within its own sub-computation domain, the lightweight candidate generation function is invoked; The candidate set is generated by the candidate generation function based on the spatial neighborhood information of the grid point, the candidate points in the boundary cache, and the highly relevant subdomains output by the lightweight router. The candidate set is used as the search range for subsequent precise attention calculation.

[0009] Optionally, the step of calculating attention weights within the candidate set through each computing node and obtaining the final attention dependency set for each grid point according to a preset filtering rule includes: For each grid point within its sub-computation domain, each computing node computes attention weights only within the range of candidate points contained in the candidate set; According to the preset filtering rules, a preset number of points with the largest attention weight or points with an attention weight greater than a preset threshold are retained from the candidate set as the final attention dependency set of the grid point, and the size of the final attention dependency set is limited to not exceeding a preset upper limit. The final attention dependency set is used for subsequent cross-node communication and local computation.

[0010] Optionally, the step of each computing node initiating cross-node data transmission to grid points belonging to other computing nodes in its final attention dependency set to obtain the dependency data required for computation includes: Each compute node traverses its own final attention dependency set, filters out the grid points that are different from its own compute node, and forms a communication set from these grid points. For each grid point in the communication set, a data request is initiated to the target computing node to which the grid point belongs, in order to obtain the feature vectors of these grid points as dependency data; Multiple requests destined for the same target computing node are aggregated, packaged, and sent asynchronously, and frequently accessed dependency data is cached locally.

[0011] Optionally, the step of performing attention weighted summation calculation based on the acquired dependency data within the scope defined by each computing node's final attention dependency set includes: For each grid point within its sub-computation domain, each computing node constructs a local feature vector from the physical parameters and geometric information of that grid point, and uses the feature vectors from the acquired dependency data as the feature vectors for cross-node dependency points. The local feature vectors correspond to the local points in the final attention dependency set, and the feature vectors from the dependency data correspond to the cross-node dependency points in the final attention dependency set. Using the local feature vector as the query vector, and using the local feature vector and the feature vector in the dependent data together as the key vector and value vector, the attention weighted sum is calculated to obtain the attention output feature of the current grid point.

[0012] Optionally, the step of summarizing the calculation results of each calculation node to obtain the pressure distribution prediction result of the vehicle's external flow field includes: Collect the predicted stress values ​​output by each computing node to all grid points within its own sub-computation domain; According to the spatial index order of each grid point in the global computational domain, the pressure prediction values ​​of each sub-computational domain are spliced ​​together to generate a complete pressure distribution prediction result covering the entire external flow field of the vehicle.

[0013] Secondly, embodiments of this application provide a distributed computing system for executing the above-described method for predicting the pressure distribution of the external flow field of a vehicle, the system comprising: Multiple computing nodes, each configured to perform local computing tasks within its own allocated region; The dependency graph partitioning module is configured to obtain the global computational domain defined by the unstructured grid describing the external flow field of the vehicle, construct an attention dependency graph based on the fluid dynamic correlation between grid points, divide the global computational domain into multiple sub-computational domains according to the attention dependency graph, and assign each sub-computational domain to the corresponding computational node; The candidate set generation module is deployed on each computing node and configured to generate a candidate set by combining the candidate generation function of the spatial neighborhood for the grid points within the sub-computation domain under the responsibility of the computing node. The attention calculation module is deployed on each computing node and configured to calculate attention weights within the candidate set and obtain the final attention dependency set for each grid point according to a preset filtering rule. The communication module, deployed on each computing node, is configured to initiate cross-node data transmission to grid points belonging to other computing nodes in the final attention dependency set of the computing node itself, in order to obtain the dependency data required for computation. The local computation execution module is deployed on each computing node and configured to perform attention weighted summation calculation based on the acquired dependency data within the scope of the final attention dependency set of that computing node. The results aggregation module is configured to collect the calculation results output by each computing node and integrate them to obtain the pressure distribution prediction results of the external flow field of the vehicle.

[0014] Optionally, the dependency graph partitioning module is specifically used for: Construct an attention dependency graph with global grid points as vertices and the expected dependency strength between two grid points as edge weights; Based on the edge weights of the attention dependency graph, a graph partitioning algorithm is used to solve the partitioning objective, which is to minimize the weighted sum of the weights of high-weight dependency edges across subdomains and the weighted sum of the load imbalance term. Based on the solution results, the global computing domain is divided into multiple sub-computing domains, and each sub-computing domain is assigned to a computing node in the distributed computing system, so that the computing node subsequently only processes the grid points within the assigned sub-computing domain.

[0015] Optionally, the candidate set generation module is specifically used for: For each grid point within the sub-computation domain that the computing node is responsible for, call the lightweight candidate generation function; The candidate generation function generates a candidate set based on the spatial neighborhood information of the grid point, candidate points in the boundary cache, and the highly relevant subdomains output by the lightweight router. The candidate set is used as the search range for subsequent precise attention calculation.

[0016] Optionally, the attention calculation module is specifically used for: For each grid point within the sub-computation domain of the computing node, attention weights are calculated only within the range of candidate points included in the candidate set; According to the preset filtering rules, a preset number of points with the largest attention weight or points with an attention weight greater than a preset threshold are retained from the candidate set as the final attention dependency set of the grid point, and the size of the final attention dependency set is limited to not exceeding a preset upper limit. The final attention dependency set is used for subsequent cross-node communication and local computation.

[0017] Optionally, the communication module is specifically used for: Iterate through the final attention dependency set of the computing node itself, filter out the grid points that are different from the current node, and form a communication set for these grid points. For each grid point in the communication set, a data request is initiated to the target computing node to which the grid point belongs, in order to obtain the feature vectors of these grid points as dependency data; Multiple requests destined for the same target computing node are aggregated, packaged, and sent asynchronously, and frequently accessed dependency data is cached locally.

[0018] Optionally, the local computation execution module is specifically used for: For each grid point within the sub-computation domain of the computing node, the physical parameters and geometric information of the grid point are used to construct a local feature vector, and the feature vector in the acquired dependency data is used as the feature vector of the cross-node dependency point. The local feature vector corresponds to the local point in the final attention dependency set, and the feature vector in the dependency data corresponds to the cross-node dependency point in the final attention dependency set. Using the local feature vector as the query vector, and using the local feature vector and the feature vector in the dependent data together as the key vector and value vector, the attention weighted sum is calculated to obtain the attention output feature of the current grid point.

[0019] Optionally, the result aggregation module is specifically used for: Collect the predicted stress values ​​output by each computing node to all grid points within its own sub-computation domain; According to the spatial index order of each grid point in the global computational domain, the pressure prediction values ​​of each sub-computational domain are spliced ​​together to generate a complete pressure distribution prediction result covering the entire external flow field of the vehicle.

[0020] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the vehicle external flow field pressure distribution prediction method described in any of the optional embodiments of the first aspect are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the vehicle external flow field pressure distribution prediction method described in any of the optional embodiments of the first aspect.

[0022] The technical solution provided in this application includes, but is not limited to, the following beneficial effects: By acquiring the global computational domain defined by the unstructured grid and collecting the physical parameters and geometric information of the grid points, it is possible to completely retain all the flow field information of the external flow field of the vehicle, accurately characterize the flow field features of different regions such as the vehicle boundary layer, wake region, and far field, and provide a complete and reliable data foundation for subsequent flow field analysis, zonal calculation and attention solution, thus ensuring the data integrity of pressure prediction.

[0023] An attention dependency graph is constructed based on the fluid dynamic correlation between grid points, and sub-computation domains are partitioned and nodes are assigned according to the dependency graph. Unlike the traditional geometric random partitioning method, this method can partition grids with strong fluid correlations to the same computation node, reduce the cross-node distribution of strongly coupled grids, and reduce the frequency of cross-node data interaction from the source.

[0024] Each computing node generates a candidate set by combining the spatial neighborhood candidate generation function, performs coarse screening of grid dependencies, narrows the search range of attention calculation, eliminates grids that have no influence at a distance, reduces the data base of subsequent attention calculations, effectively reduces the amount of invalid computation, and improves the overall computing speed.

[0025] Attention weights are calculated within the candidate set, and the final attention dependency set is determined according to the screening rules. This enables fine-grained screening of dependency grids, retaining only the associated grids that have a real impact on the pressure of the target grid, further compressing the calculation range, ensuring that the attention calculation conforms to the physical laws of the external flow field of the vehicle, and improving the calculation effectiveness.

[0026] Cross-node data transmission is initiated based on the final attention dependency set, requesting only the necessary cross-node dependency grid data. This abandons the traditional global indiscriminate communication mode, precisely limits the communication objects, reduces redundant data transmission, and lowers the communication bandwidth consumption and communication latency of the distributed system.

[0027] Attention-weighted summation is performed within the final attention dependency set, strictly constraining the computational boundary. Operations are conducted only based on the effective dependency grid, avoiding interference from irrelevant grid features. This aligns with the physical characteristics of local coupling in the pressure field, thereby improving the computational accuracy of solving the pressure in the external flow field of a vehicle.

[0028] The complete pressure distribution is obtained by summarizing the subdomain calculation results of each computing node. The results are spliced ​​and integrated based on the independent calculation results of each node, without the need for global data synchronization, which simplifies the result integration process. At the same time, it can continuously restore the pressure distribution of the external flow field of the whole vehicle, ensuring the spatial continuity and integrity of the flow field results.

[0029] In summary, this application, through a series of steps including full-domain data acquisition, physical association partitioning, candidate coarse screening, dependency fine screening, precise communication, local computation, and partition aggregation, reduces invalid communication and redundant computation without adding complex hardware or auxiliary algorithms. It optimizes the load of the distributed system, aligns with the physical characteristics of the vehicle's external flow field, and effectively improves the parallel computing efficiency and pressure prediction accuracy of pressure simulation of the vehicle's external flow field under large-scale unstructured grids.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a method for predicting the pressure distribution in the external flow field of a vehicle, as provided in Embodiment 1 of this application, is shown. Figure 2A flowchart of a sub-computation domain allocation method provided in Embodiment 1 of this application is shown; Figure 3 A flowchart of a candidate set generation method provided in Embodiment 1 of this application is shown; Figure 4 A flowchart of a method for determining the final attention dependency set provided in Embodiment 1 of this application is shown; Figure 5 A flowchart of a dependency data acquisition method provided in Embodiment 1 of this application is shown; Figure 6 A flowchart of an attention calculation method provided in Embodiment 1 of this application is shown; Figure 7 A flowchart of a method for determining pressure distribution prediction results provided in Embodiment 1 of this application is shown; Figure 8 This paper presents a bar chart comparing communication performance according to Embodiment 1 of this application. Figure 9 This paper presents a comparison chart of inference latency and video memory usage provided in Embodiment 1 of this application; Figure 10 This paper presents a bar chart comparing the accuracy of a model according to Embodiment 1 of this application. Figure 11 This paper presents a bar chart comparing the accuracy of pressure field prediction according to Embodiment 1 of this application. Figure 12 This paper shows a schematic diagram of the structure of a distributed computing system provided in Embodiment 2 of this application; Figure 13 A schematic diagram of the structure of a computer device provided in Embodiment 3 of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] Example 1 This method belongs to the intersection of high-performance computing and distributed deep learning, and is suitable for large-scale CFD (Computational Fluid Dynamics) numerical simulation scenarios such as automotive external flow fields, aerospace flow fields, and ship hydrodynamics calculations. It is specifically designed for the parallel optimization of Transformer (Transformer model, Transformer neural network model) in scientific computing. The core innovation is the neighborhood consistency constraint mechanism, which accurately fills the technical gap between communication behavior and attention computation requirements. It fundamentally solves the three core pain points of traditional distributed Transformers: ineffective global communication, failure to utilize physical neighborhood characteristics, and inconsistency between the communication domain and the computation domain. It provides an efficient and reliable parallel solution for the engineering implementation of Transformer models in large-scale scientific computing tasks.

[0035] To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart illustrating the method for predicting the pressure distribution of an external flow field in a vehicle, as shown in Embodiment 1 of this application, provides a detailed description of Embodiment 1 of this application.

[0036] See Figure 1 As shown, Figure 1 A flowchart of a method for predicting the pressure distribution of an external flow field in a vehicle, as provided in Embodiment 1 of this application, is shown. The method is applied to a distributed computing system, which includes multiple computing nodes. The method includes steps S101 to S107: S101: Obtain the global computational domain defined by the unstructured mesh describing the external flow field of the vehicle, wherein the unstructured mesh contains multiple mesh points carrying physical parameters and geometric information.

[0037] Specifically, global calculation is denoted as It corresponds to the complete flow field space of the external flow field of the car, and fully covers the car body surface, near-wall boundary layer, high-pressure area at the stagnation point at the front of the car, low-pressure wake area at the rear of the car, flow separation area on both sides of the car body and the outer far field region. It is the core solution space for subsequent physical quantities such as flow field pressure, velocity, and density.

[0038] The unstructured mesh adopts tetrahedral body mesh + triangular surface mesh. The mesh is densified on the vehicle surface, near-wall boundary layer, front stagnation area, rear wake area and separation area on both sides of the vehicle. The mesh is moderately sparse in the far field area. The distortion rate of the control unit is strictly controlled to be ≤5% to ensure the accuracy of the flow field physical quantity calculation.

[0039] Global grid point count Each grid point corresponds to a spatial location in the flow field and carries two types of core information: one is physical parameters, including static pressure, velocity vector, density, vorticity, and viscosity coefficient; the other is geometric information, including three-dimensional spatial coordinates, vehicle surface normal vector, and grid topology connection relationship, providing complete physical input for subsequent flow field feature construction and attention calculation.

[0040] S102: Construct an attention dependency graph based on the fluid dynamic correlation between grid points, divide the global computation domain into multiple sub-computation domains according to the attention dependency graph, and assign each sub-computation domain to the corresponding computation node.

[0041] Specifically, the fluid dynamic correlation directly corresponds to the intensity of the physical influence of pressure gradient transfer, velocity field convection transport, and viscous diffusion between grid points in the external flow field of a vehicle. The stronger the correlation, the tighter the coupling of the physical quantities of the flow field between two points, thus constructing an attention-dependent graph. Completely different from the traditional method of simply dividing by geometric position, it achieves adaptive partitioning driven by the physical flow field.

[0042] The core objective of the partitioning is to minimize high-weight dependency connections across subdomains while also balancing the load of each node. In practice, the high-pressure area at the front of the vehicle, the side of the vehicle, the low-pressure wake area at the rear of the vehicle, and the near-wall boundary layer are prioritized to be integrated into independent subdomains. This ensures that the strongly coupled flow field is processed within the same node, thereby reducing cross-node flow field data interaction from the source.

[0043] The global computation domain is divided into Number of sub-computation domains, number of grids in each sub-computation domain ≈ The load balancing error is ≤10%, ensuring that the computing tasks of each node are evenly distributed, avoiding local node overload or idleness, and ensuring overall parallel efficiency.

[0044] S103: Each computing node generates a candidate set by combining the candidate generation function of the spatial neighborhood with the grid points within its own sub-computation domain.

[0045] Specifically, a two-stage attention neighborhood determination mechanism is adopted. This step is the first stage of coarse screening. The core purpose is to narrow down the subsequent precise attention calculation range, which is in line with the local dependence characteristics of the pressure field of the external flow field of the car. The pressure of the grid point is only significantly affected by the grid in the near-wall neighborhood, adjacent flow layer, and boundary layer. The influence of distant grids on pressure is negligible and does not need to be included in the candidate range.

[0046] For any grid point The candidate set is denoted as Lightweight (Generate) Candidate generating function generation, formula:

[0047] Formula character explanation: : Index of spatial grid points in the external flow field of a vehicle, corresponding to the solution location for a specific pressure / velocity; Grid points Pressure-related candidate mesh set; Lightweight candidate generating functions for flow field physical constraints; The set of highly relevant subdomains / data blocks selected by the lightweight router; The set of candidate points in the local spatial proximity and boundary buffer; Global Summary (Marker), Regional Center Anchor point gather; Lightweight Candidate generating functions can be obtained through low-dimensional... (Query vector) It can be implemented through (key vector) projection, upper-level attention inheritance, local spatial constraints, or combinations thereof. It does not require global feature interaction and can quickly generate candidate sets based only on local grid information and lightweight routing results. The computational cost is far lower than that of precise attention calculation, which is suitable for the rapid generation of millions of grids.

[0048] S104: Calculate the attention weights of each computing node within the candidate set, and obtain the final attention dependency set of each grid point according to the preset filtering rules.

[0049] Specifically, a second stage of fine screening, based on two-stage neighborhood determination, is performed only in the candidate set. Internal calculation of pressure-related attention weights strictly adheres to the physical laws of the external flow field of a vehicle—grid points right The greater the impact of stress, the lower the attention weight. The higher the value, the less need to calculate grid weights without stress correlation, thus avoiding unnecessary computational overhead.

[0050] The final attention dependency set is denoted as ,formula:

[0051] Formula character explanation: Grid points The final attention depends on the set; : Index of flow field grid points within the candidate set; Grid points and Attention weights between; Top-K filter preset quantity; Attention weighting filtering preset threshold; Ultimately depends on the number of elements in the set; Ultimately depends on the maximum capacity limit of the set; , , The parameters can be set as fixed or trainable parameters according to the flow field characteristics. In actual screening, the accuracy and sparsity of the flow field solution are considered at the same time. Through fine screening, the global dependency of Transformer is transformed into the local sparse dependency of the flow field physics, which matches the CFD characteristics of strong local correlation and weak long-distance correlation of pressure field.

[0052] S105: Each computing node initiates cross-node data transmission to grid points belonging to other computing nodes in its final attention dependency set to obtain the dependency data required for computation.

[0053] Specifically, it strictly adheres to the consistency constraint mechanism between the attention computation scope and the communication scope, and the communication behavior is entirely determined by the final attention dependency set. Driven by eliminating any invalid communication, it meets the physical requirements of cross-subdomain flow field connection in the external flow field of automobiles.

[0054] In actual execution, first determine the grid points The node belongs to, only when belong And with Cross-node communication is triggered only when nodes are not on the same compute node, and only transmits data. Core physical data such as point pressure, velocity, and density; in the external flow field of a vehicle, cross-node communication only occurs in the pressure connection grid, wake extension grid, and separation zone transition grid at the subdomain boundary, avoiding global data interaction.

[0055] S106: Each computing node performs attention weighted summation calculation based on the acquired dependency data within the scope defined by its own final attention dependency set.

[0056] Specifically, strictly adhering to attention range constraints, grid points within each sub-computation domain... Its precise attention calculation is only allowed in It executes internally, prohibiting access to any non-dependent data, ensuring complete consistency between the communication domain and the computing domain at the computational level, and conforming to the local coupling mechanism of the external flow field and pressure field of a car.

[0057] In practice, the local grid physical features are first constructed, and then the dependent grid physical data obtained across nodes are fused together. Sparse attention weighted summation is then performed. Essentially, this modeling is the local coupling effect of pressure and velocity in the flow field, and the grid points are accurately solved. The predicted pressure value.

[0058] S107: Summarize the calculation results of each calculation node to obtain the predicted pressure distribution of the external flow field of the vehicle.

[0059] Specifically, the results aggregation does not require global data synchronization. It only aggregates the pressure prediction values ​​of grid points within each node's sub-computation domain, corresponding to the local pressure results of the vehicle body surface, near-wall region, and wake region, which fits the overall distribution characteristics of the pressure field of the vehicle's external flow field.

[0060] By stitching together the results of each subdomain in the order of the global grid spatial index, the continuous pressure distribution of the high-pressure area of ​​the vehicle body, the low-pressure area of ​​the roof, the low-pressure wake area of ​​the rear of the vehicle, and the separation area on both sides is fully restored, which is suitable for complex flow field scenarios of unstructured grids.

[0061] In an optional implementation, see Figure 2 As shown, Figure 2 The flowchart illustrates a sub-computation domain allocation method provided in Embodiment 1 of this application, wherein the steps of constructing an attention dependency graph based on the hydrodynamic correlation between grid points, dividing the global computation domain into multiple sub-computation domains according to the attention dependency graph, and allocating each sub-computation domain to a corresponding computation node include steps S201~S203: S201: Construct an attention dependency graph with global grid points as vertices and the expected dependency strength between two grid points as edge weights.

[0062] Specifically, attention dependency graph Explanation of the three elements: : The set of all grid points in the external flow field of a vehicle; : The set of candidate or final attention-dependent edges; The expected dependence strength between two grid points is determined by the candidate generation score, historical average attention weight, local physical correlation, or a combination thereof. In actual construction, all grid point pairs are traversed globally, and the hydrodynamic correlation strength between the two points is calculated one by one and assigned as edge weight. The edge weight directly quantifies the strength of the physical influence between grid points in the external flow field of the vehicle. The grid at the front of the vehicle has a high weight with the adjacent grid, while the weight between grids in the far field is low.

[0063] S202: Based on the edge weights of the attention-dependent graph, a graph partitioning algorithm is used to solve the partitioning objective to obtain the solution result. The partitioning objective is to minimize the weighted sum of the weights of high-weight dependent edges across subdomains and the weighted sum of the load imbalance term.

[0064] Specifically, the formula for dividing the target is:

[0065] Formula character explanation: Minimize the objective function; : Weight coefficients of cross-subdomain high-weight dependent edges; Cross-subdomain high weight depends on the sum of edge weights; : Load balancing weighting coefficient; Load imbalance across subdomains; In practical solutions, graph partitioning algorithms, load-constrained minimum cut methods, or adaptive spatial clustering methods can be used to prioritize dividing the same side of the vehicle body, continuous boundary layers, and complete wake regions into the same subdomain. Through iterative optimization, communication overhead and load distribution can be balanced to avoid hot spots.

[0066] S203: Based on the solution results, the global computing domain is divided into multiple sub-computing domains, and each sub-computing domain is assigned to a computing node in the distributed computing system, so that the computing node subsequently processes only the grid points within the assigned sub-computing domain.

[0067] Specifically, global computation domain Divided into Sub-computation domain This corresponds to independent flow field regions such as the front area of ​​the vehicle, the left side of the vehicle, the right side of the vehicle, the roof area, the rear wake area, the near-wall boundary layer area, and the far-field area of ​​the vehicle's external flow field.

[0068] Each sub-computation domain ( It contains a set of physically related mesh points with no overlap between nodes; subsequent nodes only execute the functions of their own subdomains. Tasks such as construction, candidate generation, attention calculation, and parameter updating are performed. During the training phase, a combination of local updates and periodic global synchronization is used, while during the inference phase, only local state updates are performed, which is adapted to the steady-state solution requirements of the external flow field of automobiles.

[0069] In an optional implementation, see Figure 3 As shown, Figure 3The flowchart of a candidate set generation method provided in Embodiment 1 of this application is shown. The method involves generating a candidate set by each computing node for grid points within its responsible sub-computation domain, combining this with a candidate generation function for the spatial neighborhood. This includes steps S301-S303. S301: For each grid point within its own sub-computation domain, call the lightweight candidate generation function.

[0070] Specifically, candidate generating functions Lightweight functions, based on low-dimensionality Projection, upper-level attention inheritance, local spatial constraints, or combinations thereof are used to achieve the local dependence of the pressure field in the external flow field of a vehicle.

[0071] In actual execution, the node traverses all grid points within its own subdomain and calls the candidate generation function one by one. The computational overhead is much lower than that of precise attention calculation, making it suitable for quickly generating candidate sets for large-scale flow field grids.

[0072] S302: The candidate set is generated by the candidate generation function based on the spatial neighborhood information of the grid point, the candidate points in the boundary cache, and the highly relevant subdomains output by the lightweight router.

[0073] Specifically, candidate set It consists of three parts, the formula is: It conforms to the external flow field structure of the car.

[0074] In actual generation, grid points are extracted first. The system first obtains spatial neighborhood information, then retrieves frequently used candidate points from the boundary cache, and finally combines this with highly relevant subdomain grids filtered by a lightweight router. After merging and deduplicating these three data points, a candidate set is formed. It ensures local pressure accuracy, caches data to reuse high-frequency information, connects routing results to global correlation, and fully matches the physical characteristics of the external flow field of the vehicle.

[0075] S303: Use the candidate set as the search range for subsequent precise attention calculation.

[0076] Specifically, all subsequent precise attention weight calculations and dependency filtering are limited to P(i), which aligns with the local dependency characteristics of the external flow field and pressure field of the vehicle.

[0077] In actual execution, the system will forcibly filter all grid data outside the candidate set and prohibit any attention-related calculations outside the candidate set, thereby filtering invalid pressure dependencies from the computational level and reducing redundant flow field calculations.

[0078] In an optional implementation, see Figure 4 As shown, Figure 4The flowchart illustrates a method for determining the final attention dependency set according to Embodiment 1 of this application. The method involves calculating attention weights for each computing node within the candidate set and obtaining the final attention dependency set for each grid point based on preset filtering rules, including steps S401-S403: S401: Each computing node computes attention weights for each grid point within its sub-computation domain, only within the range of candidate points contained in the candidate set.

[0079] Specifically, only for Inner grid points and of and Similarity calculations are performed on features to align with the physical laws of CFD (Computational Fluid Dynamics) in the external flow field of a car, generating attention weights. .

[0080] In actual calculations, there is no need to traverse the global grid; feature similarity calculations are performed only on grid points within the candidate set, significantly reducing computational complexity and adapting to the needs of parallel computation of large-scale flow field grids.

[0081] S402: Based on the preset filtering rules, retain a preset number of points with the largest attention weight from the candidate set, or retain points with attention weight greater than a preset threshold, as the final attention dependency set of the grid point, and limit the size of the final attention dependency set to not exceed a preset upper limit value.

[0082] Specifically, the formula for the filtering rules is: It conforms to the sparse characteristics of the external flow field and pressure field of automobiles.

[0083] In actual screening, Top-K filtering and threshold filtering can be used alone or in combination, while also... Limiting the size of the dependency set and setting a threshold based on the boundary layer thickness prevents excessive dependency sparsity from causing wake pressure distortion and allows for precise targeting. Point pressure has a decisive impact on the grid.

[0084] S403: Use the final attention dependency set for subsequent cross-node communication and local computation.

[0085] Specifically, It serves as the basis for communication and computation ranges, achieving a hard constraint of consistency between the communication and computation domains, and meeting the physical continuity requirements of the external flow field in automobiles.

[0086] In actual execution, cross-node communication only obtains... The data spans multiple nodes, and local computations are performed only based on data within N(i). The two are strongly bound together and dynamically adjusted to precisely match the needs of solving the pressure field.

[0087] In an optional implementation, see Figure 5 As shown, Figure 5 The flowchart of a dependency data acquisition method provided in Embodiment 1 of this application is shown. The method involves each computing node initiating cross-node data transmission to grid points belonging to other computing nodes within its own final attention dependency set to obtain the dependency data required for computation. This includes steps S501-S503: S501: Each computing node traverses its own final attention dependency set, filters out the grid points that are different from its own computing node, and forms a communication set from these grid points.

[0088] Specifically, the formula for the communication set is: Formula character explanation: Grid points The communication set; Grid points Identifier of the computing node; During the actual screening process, each node traverses all its final dependent grids, checks the node identifier one by one, and retains only cross-node grids to form a communication set. This set includes only the boundary layer grids, wake extension grids, and separation zone transition grids of the vehicle's external flow field across subdomains, thus eliminating invalid flow field data transmission from the source.

[0089] S502: For each grid point in the communication set, initiate a data request to the target computing node to which the grid point belongs, in order to obtain the feature vectors of these grid points as dependency data.

[0090] Specifically, the request content only includes cross-node dependent grid points. The eigenvectors contain physical parameters and geometric information, which meet the physical requirements for solving the pressure in the external flow field of a car.

[0091] In actual requests, only the target mesh data request is sent precisely, without transmitting irrelevant flow field data, so as to accurately obtain the physical data necessary for calculation and reduce the amount of data transmitted.

[0092] S503: Aggregates and packages multiple requests destined for the same target compute node and sends them asynchronously, and caches frequently accessed dependency data locally.

[0093] Specifically, aggregation and packaging reduce the number of communication messages, avoid point-to-point small-granularity communication bottlenecks, and fit the characteristics of batch data interaction in the wake region and boundary layer of the vehicle's external flow field.

[0094] In actual execution, the node merges and packages multiple requests to the same target node and sends them asynchronously. At the same time, it establishes a local cache for high-frequency dependent data, reuses duplicate data, and asynchronous communication enables communication operations to be executed in parallel with candidate set generation and local computation, further reducing overhead.

[0095] In an optional implementation, see Figure 6 As shown, Figure 6 The flowchart of an attention calculation method provided in Embodiment 1 of this application is shown. The step of performing attention weighted summation calculation based on the acquired dependency data within the scope defined by each calculation node's final attention dependency set includes steps S601-S602: S601: For each grid point within its sub-computation domain, each computing node constructs a local feature vector from the physical parameters and geometric information of that grid point, and uses the feature vectors in the acquired dependency data as feature vectors for cross-node dependency points. The local feature vectors correspond to local points in the final attention dependency set, and the feature vectors in the dependency data correspond to cross-node dependency points in the final attention dependency set.

[0096] Specifically, local feature vectors Includes grid points The three-dimensional coordinates, pressure, velocity, density, and normal vector comprehensively characterize the local flow field physical state of the external flow field of the automobile.

[0097] In actual construction, the physical parameters and geometric information of grid points are integrated into a unified feature vector. Cross-node features are the physical features of grid points obtained through communication, which represent the connection state of the flow field across subdomains. The two together constitute the input for attention calculation, and no irrelevant flow field features participate in the calculation.

[0098] S602: Using the local feature vector as the query vector, and using the local feature vector and the feature vector in the dependent data together as the key vector and value vector, calculate the attention weighted sum to obtain the attention output feature of the current grid point.

[0099] Specifically, it follows the standard Transformer attention calculation logic: , It aligns with the CFD solution logic for the external flow field of automobiles.

[0100] In actual computation, local features are used as the query, and attention weights are calculated by combining local and cross-node features and then summed in a weighted manner to form a model. The pressure coupling relationship between points and dependent meshes is analyzed, incorporating local viscosity and convection effects, and the output features are directly used for... Point pressure prediction.

[0101] In an optional implementation, see Figure 7 As shown, Figure 7 The flowchart illustrates a method for determining pressure distribution prediction results provided in Embodiment 1 of this application. The step of summarizing the calculation results of each calculation node to obtain the pressure distribution prediction result of the vehicle's external flow field includes steps S701-S702: S701: Collects the pressure prediction values ​​output by each computing node for all grid points within its own sub-computation domain.

[0102] Specifically, only the pressure prediction values ​​of grid points within each node subdomain are collected, corresponding to the local pressure results of the vehicle body surface, boundary layer, and wake region in the external flow field of the vehicle. Redundant data such as intermediate features and weights are not collected.

[0103] During actual data collection, each node only outputs the final pressure prediction results and does not transmit intermediate calculation data, thereby reducing the amount of data transmitted during result aggregation.

[0104] S702: According to the spatial index order of each grid point in the global computational domain, the pressure prediction values ​​of each sub-computational domain are spliced ​​together to generate a complete pressure distribution prediction result covering the entire external flow field of the vehicle.

[0105] Specifically, based on the spatial index mapping relationship during global grid initialization, the results of each subdomain are spliced ​​together in coordinate order to ensure the continuity of pressure values ​​in the vehicle body surface, near-wall region, wake region, and far-field region of the external flow field.

[0106] During actual stitching, the results of each subdomain are stitched together one by one according to the global grid index, which completely restores the typical external flow field pressure distribution of automobiles, including high pressure at the front stagnation point, low pressure on the roof, low pressure wake at the rear, and low pressure in the separation zone on both sides. The physical meaning is clear and consistent with the actual CFD simulation results.

[0107] To visually verify the actual technical effect of the method in this application, a series of comparative charts are presented below, based on experimental data: See Figure 8 As shown, Figure 8A bar chart comparing communication performance provided in Embodiment 1 of this application is shown. The chart compares three communication performance indicators of the distributed optimization method (i.e., the method of this application) that matches the communication domain and computation domain, and the baseline model (i.e., the traditional distributed Transformer method without the optimization scheme of this application). In the legend, blue represents the standard method and red represents the method of this application. Single-step communication data volume (unit: megabytes / step / node): Standard method: 0.125 megabytes / step / node; Method of this application: 0.005 megabytes / step / node; Method of this application significantly reduces the single-step communication data volume. Number of communication messages (messages / step): Standard method: 34 messages / step; Method of this application: 8 messages / step; Method of this application significantly reduces the number of communication messages. Communication time (unit: milliseconds / step): Standard method: 2.01 milliseconds / step; Method of this application: 1.13 milliseconds / step; Method of this application effectively shortens the single-step communication time. This figure verifies that the method proposed in this application achieves the matching between the communication domain and the computation domain through the neighborhood consistency constraint mechanism, which significantly reduces the communication overhead in distributed computing.

[0108] See Figure 9 As shown, Figure 9 A comparison chart of inference latency and memory usage provided in Embodiment 1 of this application is shown. The chart compares the inference latency and memory usage metrics of the method in this application with those of a standard method. In the legend, blue represents the standard method, and red represents the method in this application. End-to-end inference latency (unit: milliseconds / step): The standard method is 95.4 milliseconds / step, while the method in this application is 26.1 milliseconds / step, indicating a significant reduction in end-to-end inference latency. Peak memory usage (unit: gigabytes): The standard method is 0.91 gigabytes, while the method in this application is 0.92 gigabytes, showing only a slight increase in memory usage, almost on par with the standard method. This chart demonstrates that the method in this application significantly improves inference efficiency without introducing significant memory overhead, thus balancing computational efficiency and resource consumption.

[0109] See Figure 10 As shown, Figure 10A bar chart comparing the accuracy of a model provided in Embodiment 1 of this application is shown. The chart uses the percentage of relative L2 error as the core indicator to compare the accuracy performance of the method in this application and the standard method during the training and validation phases. In the legend, blue represents the standard method and red represents the method in this application. Training relative L2 error (%): Standard method 34.2%, this application method 32.5%; Validation phase relative L2 error (%): Standard method 36.4%, this application method 35.1%. The relative L2 error of the method in this application is lower than that of the standard method in both the training and validation phases, indicating that the method in this application can maintain or even improve the prediction accuracy of the model while reducing communication overhead, demonstrating reliable accuracy performance in the task of predicting the pressure distribution of the external flow field of a vehicle.

[0110] See Figure 11 As shown, Figure 11 A bar chart comparing the pressure field prediction accuracy provided in Embodiment 1 of this application is shown. The chart uses the mean absolute error of the pressure field (MAE) (which measures the average error between the predicted and actual values; a smaller value indicates higher prediction accuracy) as an indicator to compare the pressure field prediction errors of the method in this application and the standard method. In the legend, blue represents the standard method, and red represents the method in this application. The MAE of the standard method is 31.9; the MAE of the method in this application is 30.4. The MAE of the method in this application is lower than that of the standard method, further verifying that the neighborhood consistency constraint method of this application can effectively reduce prediction errors and improve the accuracy of pressure field prediction in the task of predicting the pressure distribution in the external flow field of a vehicle.

[0111] Example 2 See Figure 12 As shown, Figure 12 This illustration shows a schematic diagram of a distributed computing system according to Embodiment 2 of this application, wherein the system, used to execute the above-described method for predicting the pressure distribution of the external flow field of a vehicle, includes: Multiple computing nodes 1201, each of which is configured to perform local computing tasks within its own allocated region; The dependency graph partitioning module 1202 is configured to obtain the global computational domain defined by the unstructured grid describing the external flow field of the car, construct an attention dependency graph based on the fluid dynamic correlation between grid points, divide the global computational domain into multiple sub-computational domains according to the attention dependency graph, and assign each sub-computational domain to the corresponding computational node. The candidate set generation module 1203 is deployed on each computing node and configured to generate a candidate set by combining the candidate generation function of the spatial neighborhood for the grid points in the sub-computation domain under the responsibility of the computing node. Attention calculation module 1204 is deployed on each computing node and configured to calculate attention weights within the candidate set and obtain the final attention dependency set for each grid point according to a preset filtering rule; The communication module 1205 is deployed on each computing node and is configured to initiate cross-node data transmission to grid points belonging to other computing nodes in the final attention dependency set of the computing node itself, in order to obtain the dependency data required for computing. The local computation execution module 1206 is deployed on each computing node and configured to perform attention weighted summation calculation based on the acquired dependency data within the scope of the final attention dependency set of the computing node itself. The result summary module 1207 is configured to collect the calculation results output by each computing node and integrate them to obtain the pressure distribution prediction results of the external flow field of the vehicle.

[0112] In an optional implementation, the dependency graph partitioning module is specifically used for: Construct an attention dependency graph with global grid points as vertices and the expected dependency strength between two grid points as edge weights; Based on the edge weights of the attention dependency graph, a graph partitioning algorithm is used to solve the partitioning objective, which is to minimize the weighted sum of the weights of high-weight dependency edges across subdomains and the weighted sum of the load imbalance term. Based on the solution results, the global computing domain is divided into multiple sub-computing domains, and each sub-computing domain is assigned to a computing node in the distributed computing system, so that the computing node subsequently only processes the grid points within the assigned sub-computing domain.

[0113] In an optional implementation, the candidate set generation module is specifically used for: For each grid point within the sub-computation domain that the computing node is responsible for, call the lightweight candidate generation function; The candidate generation function generates a candidate set based on the spatial neighborhood information of the grid point, candidate points in the boundary cache, and the highly relevant subdomains output by the lightweight router. The candidate set is used as the search range for subsequent precise attention calculation.

[0114] In an optional implementation, the attention calculation module is specifically used for: For each grid point within the sub-computation domain of the computing node, attention weights are calculated only within the range of candidate points included in the candidate set; According to the preset filtering rules, a preset number of points with the largest attention weight or points with an attention weight greater than a preset threshold are retained from the candidate set as the final attention dependency set of the grid point, and the size of the final attention dependency set is limited to not exceeding a preset upper limit. The final attention dependency set is used for subsequent cross-node communication and local computation.

[0115] In an optional implementation, the communication module is specifically used for: Iterate through the final attention dependency set of the computing node itself, filter out the grid points that are different from the current node, and form a communication set for these grid points. For each grid point in the communication set, a data request is initiated to the target computing node to which the grid point belongs, in order to obtain the feature vectors of these grid points as dependency data; Multiple requests destined for the same target computing node are aggregated, packaged, and sent asynchronously, and frequently accessed dependency data is cached locally.

[0116] In an optional implementation, the local computation execution module is specifically used for: For each grid point within the sub-computation domain of the computing node, the physical parameters and geometric information of the grid point are used to construct a local feature vector, and the feature vector in the acquired dependency data is used as the feature vector of the cross-node dependency point. The local feature vector corresponds to the local point in the final attention dependency set, and the feature vector in the dependency data corresponds to the cross-node dependency point in the final attention dependency set. Using the local feature vector as the query vector, and using the local feature vector and the feature vector in the dependent data together as the key vector and value vector, the attention weighted sum is calculated to obtain the attention output feature of the current grid point.

[0117] In an optional implementation, the result aggregation module is specifically used for: Collect the predicted stress values ​​output by each computing node to all grid points within its own sub-computation domain; According to the spatial index order of each grid point in the global computational domain, the pressure prediction values ​​of each sub-computational domain are spliced ​​together to generate a complete pressure distribution prediction result covering the entire external flow field of the vehicle.

[0118] Example 3 Based on the same application concept, see [link / reference] Figure 13 As shown, Figure 13 This illustration shows a structural schematic diagram of a computer device provided in Embodiment 3 of this application, wherein, as shown... Figure 13As shown, the computer device 1300 provided in Embodiment 3 of this application includes: The system includes a processor 1301, a memory 1302, and a bus 1303. The memory 1302 stores machine-readable instructions executable by the processor 1301. When the computer device 1300 is running, the processor 1301 and the memory 1302 communicate via the bus 1303. When the machine-readable instructions are executed by the processor 1301, they perform the steps of the vehicle external flow field pressure distribution prediction method shown in Embodiment 1 above.

[0119] Example 4 Based on the same concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the vehicle external flow field pressure distribution prediction method described in any of the above embodiments.

[0120] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0121] The computer program product for predicting the pressure distribution of the external flow field of a vehicle provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0122] The distributed computing system provided in this application embodiment can be specific hardware on a device or software or firmware installed on the device. The device provided in this application embodiment has the same implementation principle and technical effects as the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0123] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0124] 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.

[0125] In addition, the functional units in the embodiments provided in 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.

[0126] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0128] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in this application, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application.

Claims

1. A method for predicting the pressure distribution in the external flow field of a vehicle, characterized in that, The method is applied to a distributed computing system, which includes multiple computing nodes; the method includes: Obtain the global computational domain defined by the unstructured mesh describing the external flow field of the vehicle, wherein the unstructured mesh contains multiple mesh points carrying physical parameters and geometric information; An attention dependency graph is constructed based on the fluid dynamic correlation between grid points. The global computation domain is divided into multiple sub-computation domains according to the attention dependency graph, and each sub-computation domain is assigned to a corresponding computation node. Each computing node generates a candidate set by combining the candidate generation function of the spatial neighborhood with the grid points within its own sub-computation domain. Attention weights are calculated for each computing node within the candidate set, and the final attention dependency set for each grid point is obtained according to a preset filtering rule. Each computing node initiates cross-node data transmission to grid points belonging to other computing nodes in its final attention dependency set to obtain the dependency data required for computation. Each computing node performs attention weighted summation calculation based on the acquired dependency data, within the scope defined by its own final attention dependency set. The calculation results of each calculation node are summarized to obtain the predicted pressure distribution of the external flow field of the vehicle.

2. The method according to claim 1, characterized in that, The process involves constructing an attention dependency graph based on the fluid dynamic correlation between grid points, dividing the global computational domain into multiple sub-computational domains according to the attention dependency graph, and assigning each sub-computational domain to a corresponding computational node, including: Construct an attention dependency graph with global grid points as vertices and the expected dependency strength between two grid points as edge weights; Based on the edge weights of the attention dependency graph, a graph partitioning algorithm is used to solve the partitioning objective to obtain the solution result. The partitioning objective is to minimize the weighted sum of the weights of high-weight dependency edges across subdomains and the weighted sum of the load imbalance term. Based on the solution results, the global computing domain is divided into multiple sub-computing domains, and each sub-computing domain is assigned to a computing node in the distributed computing system, so that the computing node subsequently processes only the grid points within the assigned sub-computing domain.

3. The method according to claim 1, characterized in that, The process of generating a candidate set by each computing node for grid points within its own sub-computation domain, combined with a candidate generation function for the spatial neighborhood, includes: For each grid point within its own sub-computation domain, the lightweight candidate generation function is invoked; The candidate set is generated by the candidate generation function based on the spatial neighborhood information of the grid point, the candidate points in the boundary cache, and the highly relevant subdomains output by the lightweight router. The candidate set is used as the search range for subsequent precise attention calculation.

4. The method according to claim 1, characterized in that, The step of calculating attention weights within the candidate set through each computing node and obtaining the final attention dependency set for each grid point according to a preset filtering rule includes: Each computing node computes attention weights for each grid point within its sub-computation domain, only within the range of candidate points contained in the candidate set; According to the preset filtering rules, a preset number of points with the largest attention weight or points with an attention weight greater than a preset threshold are retained from the candidate set as the final attention dependency set of the grid point, and the size of the final attention dependency set is limited to not exceeding a preset upper limit. The final attention dependency set is used for subsequent cross-node communication and local computation.

5. The method according to claim 1, characterized in that, The process involves each computing node initiating cross-node data transmission to grid points belonging to other computing nodes within its own final attention dependency set, in order to obtain the dependency data required for computation. This includes: Each compute node traverses its own final attention dependency set, filters out the grid points that are different from its own compute node, and forms a communication set from these grid points. For each grid point in the communication set, a data request is initiated to the target computing node to which the grid point belongs, in order to obtain the feature vectors of these grid points as dependency data; Multiple requests destined for the same target computing node are aggregated, packaged, and sent asynchronously, and frequently accessed dependency data is cached locally.

6. The method according to claim 1, characterized in that, The step of performing attention-weighted summation calculation based on the acquired dependency data within the scope defined by each computing node's final attention dependency set includes: For each grid point within its sub-computation domain, each computing node constructs a local feature vector from the physical parameters and geometric information of that grid point, and uses the feature vectors from the acquired dependency data as the feature vectors for cross-node dependency points. The local feature vectors correspond to the local points in the final attention dependency set, and the feature vectors from the dependency data correspond to the cross-node dependency points in the final attention dependency set. Using the local feature vector as the query vector, and using the local feature vector and the feature vector in the dependent data together as the key vector and value vector, the attention weighted sum is calculated to obtain the attention output feature of the current grid point.

7. The method according to claim 1, characterized in that, The summarizing of the calculation results from each calculation node yields the predicted pressure distribution of the vehicle's external flow field, including: Collect the predicted stress values ​​output by each computing node to all grid points within its own sub-computation domain; According to the spatial index order of each grid point in the global computational domain, the pressure prediction values ​​of each sub-computational domain are spliced ​​together to generate a complete pressure distribution prediction result covering the entire external flow field of the vehicle.

8. A distributed computing system for executing the vehicle external flow field pressure distribution prediction method of claim 1, characterized in that, include: Multiple computing nodes, each of which is used to execute local computing tasks in a corresponding allocated region; The dependency graph partitioning module is used to obtain the global computational domain defined by the unstructured grid describing the external flow field of the car, construct an attention dependency graph based on the fluid dynamic correlation between grid points, partition the global computational domain into multiple sub-computational domains according to the attention dependency graph, and assign each sub-computational domain to a matching computational node. The candidate set generation module is deployed on each computing node and is used to generate a candidate set for the grid points in the sub-computation domain of the computing node and the corresponding candidate generation function of the spatial neighborhood. The attention calculation module is deployed on each computing node to perform attention weight calculation within the corresponding candidate set and determine the final attention dependency set corresponding to each grid point according to the preset filtering rules. The communication module is deployed on each computing node and is used to initiate cross-node data interaction with grid points belonging to other computing nodes based on its own determined final attention dependency set, in order to obtain the dependency data required for computation. Local computation execution modules are deployed on each computation node and are used to perform attention weighted summation operations by combining the acquired dependency data within the scope defined by the final attention dependency set. The results aggregation module is used to collect the local calculation results output by all computing nodes and integrate them to obtain the complete prediction results of the pressure distribution of the external flow field of the vehicle.

9. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the method for predicting the pressure distribution of the external flow field of a vehicle as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method for predicting the pressure distribution of the external flow field of a vehicle as described in any one of claims 1 to 7.

Images