A method for fast simulation of multi-physical fields of a compressor blisk

Abstract

The application provides a compressor blisk multi-physical field rapid simulation method, and relates to the technical field of aero-engines, and comprises the following steps: a parameterized geometric model of the compressor blisk is established, and aerodynamic flow field grids and blisk structure grids are automatically divided; based on the parameterized geometric model, automatic simulation analysis including aerodynamic performance, temperature field, static strength, mode and life is sequentially performed to obtain blisk gas-thermal-solid multi-disciplinary coupling evaluation results; based on multi-dimensional data in the evaluation process, an intelligent simulation proxy model based on a physical information neural network, a flow field rapid prediction proxy model based on a graph convolutional neural network and an artificial intelligence proxy model based on deep learning are respectively constructed to realize rapid coupling simulation of the compressor blisk multi-physical field. The application provides a compressor blisk multi-disciplinary rapid simulation method related to aerodynamic performance, structural design, heat transfer analysis and strength analysis, and realizes rapid simulation of the compressor blisk multi-physical field.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engines, and in particular to a multi-physical field rapid simulation method for an integral blade disk of a compressor. Background Art

[0002] As a high-speed rotating component in aircraft engines, the compressor blisk is characterized by high inlet temperature and pressure, and a harsh and complex operating environment. Furthermore, it is particularly sensitive to flow parameters, making it prone to surge and other issues. Its design challenges integrate multiple disciplines, including gas dynamics, heat transfer, and mechanical strength, with deep intersections and high coupling among these disciplines. The design of high-performance, high-efficiency, and high-load compressor blisks, in particular, requires comprehensive consideration of interdisciplinary conflicts and the search for an optimized design solution within the design feasible domain.

[0003] The current design and engineering calculation methods for compressor blisks are serial: aerodynamic design is performed first, followed by aerodynamic performance calculations and analysis. Following aerodynamic design, the compressor blade structure / compressor disk design is performed based on the resulting aerodynamic shape of the compressor blades. The blades and disks are then combined to form a preliminary blisk structural design. Once the preliminary blisk structural design is complete, a heat transfer analysis is performed to obtain the temperature field. This is done to determine whether the selected material for the compressor blisk is operating at excessive temperatures and to use the temperature field as the thermal boundary for strength analysis. Finally, a strength analysis is performed on the compressor blisk to determine whether its static strength reserve, overspeed burst speed reserve, low-cycle fatigue life, high-cycle fatigue life, and vibration modes meet requirements. As can be seen from the above description, the traditional engineering design approach for compressor blisks is characterized by independent design processes and decoupling between disciplines. This makes rapid iterative design and analysis nearly impossible. Furthermore, according to the engine development process, the design of a compressor blisk involves a series of steps: one-dimensional aerodynamic design in the conceptual phase, two-dimensional structural design in the engineering phase, and three-dimensional performance analysis in the engineering phase. This makes it difficult to ensure the dimensionality and accuracy of data transfer between disciplines throughout the design cycle. In general, the traditional decoupled / serial approach to compressor blisks presents the following limitations:

[0004] (1) Each discipline conducts design independently without comprehensively considering the impact of other disciplines. The design results given are only feasible results of the discipline itself, which is naturally unsuitable for solving the multidisciplinary design contradictions of high-performance, high-load compressor integral disks;

[0005] (2) Low iteration efficiency. Serial design is carried out between disciplines. After the previous design step is completed, its design results are used as input for the next design step. This design process makes it almost impossible to conduct rapid iterative design and find a comprehensive optimization solution. Once it is difficult to find a feasible solution for the subsequent design and it is necessary to iterate the next step, it is necessary to break the design results of the previous step and redesign. This will bring huge manpower and material costs, and greatly extend the design cycle. These costs are often unacceptable in the early stages of engine development.

[0006] (3) The mismatch in data transfer dimensions between disciplines leads to low simulation analysis accuracy;

[0007] (4) Serial / decoupled design and weakly coupled analysis and calculation among multiple disciplines lead to long design and analysis cycles;

[0008] (5) The life assessment deviation of the compressor integral blade strength reserve level is large.

[0009] Therefore, the traditional decoupled / serial design and simulation calculation methods of compressor components have become increasingly difficult to meet the high efficiency, high precision, high load and long life low stress multi-physics field coupling design requirements and the needs of rapid iterative optimization of the integral blade. Summary of the Invention

[0010] In view of this, an embodiment of the present application provides a multi-physics field rapid simulation method for an integral blade disk of a compressor to solve the problem that the existing engineering design method for an integral blade disk of a compressor is difficult to comprehensively consider multiple disciplines such as gas, heat and solid, and lacks a multi-disciplinary rapid iterative analysis method, thereby improving design efficiency and simulation accuracy.

[0011] The present application provides the following technical solution: a method for rapid multi-physics simulation of a compressor blade disk, comprising:

[0012] A parametric geometric model of the compressor blisk is established, and the aerodynamic flow domain grid and the blisk structural grid are automatically divided. Based on the parametric geometric model, automatic simulation analysis including aerodynamic performance, temperature field, static strength, modal and life span is sequentially performed to obtain the results of the multidisciplinary coupled aero-thermal-solid evaluation of the compressor blisk.

[0013] Based on the multidimensional data in the gas-thermal-solid multidisciplinary coupling evaluation process of the compressor integral blade, an intelligent simulation agent model based on a physical information neural network, a flow field rapid prediction agent model based on a graph convolutional neural network, and an artificial intelligence agent model based on deep learning are constructed respectively. Based on the intelligent simulation agent model, the flow field rapid prediction agent model, and the artificial intelligence agent model, the rapid coupling simulation of the multi-physical fields of the compressor integral blade is realized.

[0014] According to one embodiment of the present application, establishing a parameterized geometric model of a compressor blade disk includes:

[0015] According to the preset geometric structure parameters of the compressor integral disk, the parametric geometric model of the compressor integral disk is automatically generated through the CAD engine, and the aerodynamic flow domain grid and the disk structure grid are automatically divided.

[0016] According to an embodiment of the present application, the preset geometric structure parameters include aerodynamic boundaries, air system boundaries, rotational speed boundaries, material parameters and structural constraint boundaries.

[0017] According to one embodiment of the present application, based on the parameterized geometric model, automatic simulation analysis including aerodynamic performance, temperature field, static strength, mode and life is performed in sequence, including:

[0018] Calculating the aerodynamic load distribution on the surface of the blisk according to the preset geometric structure parameters to evaluate the aerodynamic performance of the compressor blisk;

[0019] The aerodynamic load distribution on the blisk surface is used as a thermal boundary condition to solve the blade-disk temperature field boundary to evaluate the compressor blisk temperature field;

[0020] Calculating the stress distribution of the compressor blisk based on the aerodynamic load distribution on the blisk surface and the temperature field boundary obtained by the aero-thermal coupling calculation to perform a static strength assessment of the compressor blisk;

[0021] Calculating the static frequency and dynamic frequency of the compressor blisk based on the aerodynamic load distribution on the surface of the blisk, the temperature field boundary, and the stress distribution to perform a modal evaluation of the compressor blisk;

[0022] Based on the static strength assessment results and modal assessment results, and according to the material parameters of the compressor blisk, the compressor blisk life assessment is carried out.

[0023] According to one embodiment of the present application, it also includes: judging whether the temperature field, stress distribution and lifespan meet the set requirements based on the gas-thermal-solid multidisciplinary coupling evaluation results of the compressor integral blade; if there are abnormal problems including local overtemperature, excessive local stress or short life, feeding back and optimizing the parameterized geometric model of the compressor integral blade, and iterating again until a compressor integral blade design result that meets the design criteria is obtained.

[0024] According to one embodiment of the present application, an intelligent simulation agent model based on a physical information neural network is constructed, including:

[0025] Based on the multi-scale Fourier network architecture, a neural attention mechanism is introduced to clarify the multiplicative interaction between different input dimensions and enhance the hidden state through residual connections;

[0026] The weights are adaptively updated through the learning rate annealing method to balance the learning speed of each loss.

[0027] According to one embodiment of the present application, constructing an intelligent simulation agent model based on a physical information neural network also includes:

[0028] The time domain is divided into multiple subdomains, the PDE residual loss is reformulated for the subdomains, and the physical causal relationship during model training is clarified to perform causal training on the model.

[0029] According to one embodiment of the present application, a flow field rapid prediction agent model based on a graph convolutional neural network includes:

[0030] The image data of the compressor blades is transformed from the physical space grid to a regular Cartesian grid in the computational space through coordinate transformation, so that the data can be adapted to the convolutional neural network to achieve rapid prediction of the flow field.

[0031] According to one embodiment of the present application, the flow field rapid prediction agent model based on the graph convolutional neural network also includes:

[0032] The simulation data is stored in a topological connection structure based on graph representation, and the information is transmitted and aggregated on the topological connection graph through graph convolution operations to complete the prediction of the flow field solution.

[0033] According to one embodiment of the present application, building an artificial intelligence agent model based on deep learning includes:

[0034] Parameterization of geometry and boundary conditions through modeling engines, model parameters, and model templates;

[0035] Using a neural network solver, parameterized geometry and boundary conditions are used as input for training the neural network to achieve rapid prediction of the physical field under specific parameter structures or boundary conditions.

[0036] The embodiments of the present invention provide a multidisciplinary rapid simulation method for a compressor blade disk, involving aerodynamic performance, structural design, heat transfer analysis, and strength analysis. This method addresses the difficulty of existing compressor blade disk engineering design methods in comprehensively considering multiple disciplines, such as aerodynamics, thermal solids, and the lack of multidisciplinary rapid iterative analysis methods. The present invention also develops intelligent agent model construction methods, such as physics-guided neural operators and graph neural networks, that are adapted to standard models of compressor blade disks. This method establishes a multi-physics rapid calculation model for the compressor blade disk, addressing the issues of large computational resources and long computational time consumed by multidisciplinary simulation analysis, and laying a model foundation for the rapid design, analysis, and virtual verification of compressor blade disks.

[0037] Compared with the traditional method, the at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects:

[0038] (1) Multi-physics coupling simulation method

[0039] The analysis method proposed in the embodiment of the present invention considers the multidisciplinary coupling conditions of the integral blade of the compressor. It considers that each point of the integral blade of the compressor is affected by the gas temperature or pressure and the solid strength, and couples these parameters to ultimately determine the actual service life of the integral blade of the compressor. It truly solves the multi-physical field coupling problem of the integral blade of the compressor, such as gas, heat, and solid, and realizes high-dimensional and high-precision data transmission between disciplines, greatly reducing the design iteration cycle of the integral blade of the compressor and improving the design accuracy.

[0040] (2) Intelligent simulation method

[0041] This embodiment of the present invention applies physical information neural networks, operator learning techniques, and graph convolutional neural networks to construct an intelligent simulation method for engine components. Compared with traditional methods, this intelligent simulation method reduces or eliminates the need for meshing; embeds physical information constraints within the neural network, improving model reliability; enables natural multi-physics coupled simulation; and can use parameterized geometry or boundary conditions as input, providing technical support for the construction of intelligent agent models.

[0042] (3) Efficient neural network architecture

[0043] The embodiment of the present invention is oriented towards the compressor integral blade disk structure and designs a neural network architecture to improve the efficiency and accuracy of the intelligent simulation method.

[0044] (4) Intelligent agent model technology

[0045] The embodiment of the present invention utilizes the powerful fitting capabilities of PINNs and operator learning technology, takes the parameters of geometric or boundary conditions as the input of the neural network, and constructs an artificial intelligence agent model embedded with physical constraint equations for the compressor integral disk. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0047] Figure 1 2. It is a schematic flow chart of a multi-physics field rapid simulation method for a compressor blade disk according to an embodiment of the present invention;

[0048] Figure 2 2. It is a schematic diagram of the multidisciplinary coupling simulation process of the compressor blade in an embodiment of the present invention;

[0049] Figure 3 2. It is a schematic diagram of a method for establishing a multidisciplinary rapid simulation model of a compressor blade according to an embodiment of the present invention;

[0050] Figure 4 Schematic diagram of the Fourier network fitting of high-frequency features in an embodiment of the present invention;

[0051] FIG5( a ) is a schematic diagram of a multi-scale Fourier network architecture according to an embodiment of the present invention; FIG5( b ) is a schematic diagram of a spatiotemporal multi-scale Fourier network architecture according to an embodiment of the present invention;

[0052] Figure 6 is an example of a one-dimensional wave equation in an embodiment of the present invention; wherein the lower line represents causal training; the upper line represents conventional PINNs;

[0053] Figure 7 Schematic diagram of the application of a convolutional neural network model based on coordinate transformation in rapid prediction of a turbine blade physical field in an embodiment of the present invention;

[0054] Figure 8 Schematic diagram of flow field data representation based on grid topology connection in an embodiment of the present invention;

[0055] Figure 9 Schematic diagram of conversion of flow field data stored in the grid center into a topological connection graph in an embodiment of the present invention;

[0056] Figure 10 1 is a schematic diagram of a flow field data format based on topological connection in an embodiment of the present invention;

[0057] Figure 11 Schematic diagram of graph convolution operation in an embodiment of the present invention;

[0058] Figure 12 Schematic diagram of a deep graph convolutional network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0060] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the features in the following embodiments and embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0061] like Figure 1 As shown, an embodiment of the present invention provides a multi-physics field rapid simulation method for a compressor blade disk, comprising:

[0062] A parametric geometric model of the compressor blisk is established, and the aerodynamic flow domain grid and the blisk structural grid are automatically divided. Based on the parametric geometric model, automatic simulation analysis including aerodynamic performance, temperature field, static strength, modal and life span is sequentially performed to obtain the results of the multidisciplinary coupled aero-thermal-solid evaluation of the compressor blisk.

[0063] Based on the multidimensional data in the gas-thermal-solid multidisciplinary coupling evaluation process of the compressor integral blade, an intelligent simulation agent model based on a physical information neural network, a flow field rapid prediction agent model based on a graph convolutional neural network, and an artificial intelligence agent model based on deep learning are constructed respectively. Based on the intelligent simulation agent model, the flow field rapid prediction agent model, and the artificial intelligence agent model, the rapid coupling simulation of the multi-physical fields of the compressor integral blade is realized.

[0064] The method of the embodiment of the present invention mainly includes the following contents:

[0065] (1) Establish the gas-thermal-solid coupling simulation process of the compressor blade, such as Figure 2 As shown;

[0066] Establish a strongly coupled aero-thermal-solid simulation process for the compressor blade. The simulation process is based on a set of parameterized compressor blade geometry models. CAD software is used to automatically perform geometric modeling of the aerodynamic flow domain and structural domain, and automatically perform meshing of the flow domain and structural domain, and transfer domain boundary data. Parameterized aerodynamic boundaries, air system boundaries, speed boundaries, engineering material libraries, and result constraints are preset. Automatic simulation analysis is performed based on the model and boundaries. Aerodynamic / temperature / static strength / modal results are extracted according to preset post-processing, and design results such as performance, structural reserve, life, and resonance are analyzed and verified. It has the following functions:

[0067] a. Generate a geometric model of the compressor blisk based on the preset geometric structural parameters of the compressor blisk using CAD software;

[0068] b. Obtain the aerodynamic boundary of the compressor blade and determine the actual temperature and / or pressure at each point on the blade;

[0069] c. determining the static strength of all points of the compressor blisk based on the actual temperatures and pressures at all points;

[0070] d. Determine the actual service life of the compressor blade based on the static strength at all points.

[0071] (2) Establish a multidisciplinary intelligent agent model for the compressor blade, such as Figure 3 As shown;

[0072] In response to the problems of weak intelligent simulation foundation and difficulty in constructing agent models in the field of aero-engines, we have carried out research on the construction of artificial intelligence agent models for rapid simulation of blades based on artificial intelligence, made breakthroughs in deep learning model architecture design technology, developed physical-guided neural operators and graph neural networks that are suitable for standard models of compressor blades, and mastered technologies such as data dimensionality reduction and dynamic modeling. It has the following functions:

[0073] a. Data collection, supplementation and processing

[0074] Multidimensional data from simulations and tests is collected and organized, covering geometric parameters (compressor blisk geometry), boundary conditions (aerodynamic parameters), operating conditions (speed), and performance indicators (material properties). Data cleaning and standardization techniques are employed to ensure data quality. Furthermore, feature selection and dimensionality reduction methods are used to extract key features and reduce redundancy. Finally, a parameter list is generated through a design of experiments (DOE).

[0075] Through multidisciplinary simulation, a sample set is formed. Existing test and simulation data are classified. One portion is used for training to build an artificial intelligence proxy model, and the other portion is used to verify the accuracy of the model. If the model accuracy does not meet the requirements, the DOE method is changed, the number of samples is increased, and the parameter list is regenerated. To address the problem of insufficient test data, an intelligent simulation method is used to supplement the training data. Since simulation results usually contain more comprehensive physical field data and can fully cover the working conditions being studied, this gives simulation results an important advantage in building artificial intelligence proxy models. Therefore, the data for training the model mainly comes from numerical simulation results, while the test results are mainly used to evaluate the accuracy error of the proxy model.

[0076] b. Intelligent simulation technology for compressor blades

[0077] This embodiment conducts research on intelligent simulation technology for integral blades of compressors, including three technical paths: physical information neural network, operator learning, and graph convolutional neural network. For integral blades of compressors, this embodiment adopts an intelligent simulation method based on physical information neural network, a flow field rapid prediction method based on graph convolutional neural network, and an artificial intelligence agent model construction method based on deep learning, and establishes a multidisciplinary simulation intelligent model for integral blades of compressors. Finally, the accuracy of the multidisciplinary simulation intelligent model is verified. This embodiment develops an efficient neural network architecture suitable for multi-scale problems, and uses a hierarchical network structure and a multi-level feature extraction mechanism to achieve comprehensive simulation from micro to macro. In view of the coupling characteristics of multiple physical fields, an efficient weight adaptive balancing strategy is developed to dynamically adjust the mutual influence weights between each physical field, and realize the coupled simulation of multiple physical fields under the same framework to ensure the accuracy and stability of the simulation results.

[0078] c. Artificial Intelligence Agent Model Technology

[0079] This example combines simulation and test data, utilizing parametric modeling techniques to construct a geometric model of a compressor blisk, enabling it to flexibly adapt to varying design parameters. This model defines parameterized representations of boundary conditions, including variables such as temperature and speed. These boundary conditions are then adjusted using test data to ensure they align more closely with actual operating conditions.

[0080] Based on this, this embodiment employs deep learning technology to design a rational network architecture. The input layer receives parameterized geometric data and boundary conditions, while the output layer predicts various performance indicators. To enhance the physical interpretability of the model, physical constraints (such as heat conduction equations and fluid dynamics equations) are used as part of the loss function to help guide the model to learn relevant physical laws. This embeds physical information into the deep learning model, improving its accuracy and reliability.

[0081] Furthermore, the dataset was split into training, validation, and test sets to ensure the model's performance on unseen data. Cross-validation and hyperparameter tuning were used to improve the model's generalization and prediction accuracy. This systematic strategy not only optimizes compressor blisk design but also provides a solid foundation for future research and promotes the development of intelligent simulation technology in engineering applications.

[0082] In some embodiments of the present invention, the multi-physics field rapid simulation method for a compressor blade disk includes the following steps:

[0083] 1. Establish the gas-thermal-solid coupling simulation process for the compressor blade

[0084] Principle: This embodiment realizes the simulation process of integrated air-heat-solid integration of the compressor blade. The aerodynamic evaluation of the compressor blade can realize the evaluation of the aerodynamic load, blade flow field and blade performance of the blade; the heat transfer evaluation and analysis stage of the compressor blade can automatically obtain the aerodynamic parameter design boundary and carry out the temperature field evaluation and analysis of the compressor blade and the wheel of the compressor blade; the static strength evaluation of the compressor blade can realize the automatic loading of the static strength calculation boundary conditions according to the aerodynamic load boundary and temperature field boundary obtained by the air-heat coupling calculation, and realize the static strength evaluation and analysis of the compressor blade; the modal evaluation and analysis of the compressor blade can automatically inherit the static strength calculation load data and constraint boundaries from the upstream static strength calculation, realize the static frequency and dynamic frequency calculation of the compressor blade, and finally realize the evaluation of the life of the compressor blade by obtaining the equivalent stress distribution and deformation of the compressor blade.

[0085] The specific process of the gas-thermal-solid multidisciplinary coupling analysis method for the compressor blade of this embodiment is as follows: Figure 2 As shown, the following steps are included:

[0086] Step 1: Generate and process the parametric geometry model of the compressor blisk. This involves processing the compressor blisk geometry model based on real geometry. While simplifying the geometry model, it also maximizes the restoration of the real geometry and improves simulation evaluation accuracy.

[0087] Step 2: Automatically mesh the aerodynamic flow domain and the bladed disk structure. Meshing is performed based on the simplified geometric model, providing the initial geometry and mesh input for the compressor bladed disk aero-thermal-solid multidisciplinary coupled analysis platform.

[0088] Step 3: Conduct aerodynamic performance evaluation. Calculate the aerodynamic loads on the blisk using the flow analysis module in software such as CFX / Fluent.

[0089] Step 4: Conduct temperature field evaluation. Use the thermal coupling module in software such as CFX / Fluent to evaluate the temperature field of the entire blade disk.

[0090] Step 5: Conduct a static strength assessment. Within the structural strength analysis module of software such as ANSYS / Abqus, the obtained aerodynamic load boundary and temperature field boundary of the compressor blisk are used as input boundaries for the static strength calculation module. This automatically applies load boundary loading to the compressor blisk, performs a static strength assessment of the compressor blisk, and obtains the compressor blisk stress distribution and deformation.

[0091] Step 6: Conduct modal evaluation. Using the structural strength analysis module of software such as ANSYS / Abqus, the aerodynamic load distribution and solid surface temperature field on the compressor blisk are obtained. Simultaneously, the surface stresses of the compressor blisk obtained in the static strength calculation module are used as input boundaries for the compressor blisk static and dynamic frequency calculation module. This allows for calculation of the static and dynamic frequencies of the compressor blisk, obtaining the blade Campbell diagram, and performing vibration analysis of the compressor blisk.

[0092] Step 7: Conduct life assessment. Based on the results obtained from static strength calculations and vibration analysis, and according to the compressor blisk material parameters, the compressor blisk life assessment can be completed.

[0093] Step 8: Determine whether the compressor blisk temperature field, stress distribution, and lifespan meet the requirements. Based on the results of the compressor blisk's multidisciplinary gas-thermal-solid coupled assessment, determine and analyze whether the compressor blisk's temperature field, stress distribution, and lifespan meet the requirements. If any abnormalities such as local overheating, excessive local stress, or short lifespan are detected, timely feedback is provided and the compressor blisk parameter model is optimized. Further iterations are performed to ultimately obtain a compressor blisk design that meets all design criteria.

[0094] 2. Intelligent simulation method based on physical information neural network

[0095] PINNs (Physics-Informed Neural Networks) use physical laws (such as mechanical constitutive relations) to constrain neural network outputs, combining sparse data with physical information (such as differential equations and their physical parameters) to predict physical processes. In scientific fields with limited data but comprehensive physical knowledge, this approach offers superior predictive performance compared to purely data-driven neural network algorithms. This physics-informed machine learning approach offers a new approach to solving various engineering problems and is promoting advancements in various fields.

[0096] In general, PINNs incorporates physical equation loss and boundary condition loss into the loss function. By optimizing the loss function, the predicted value of the neural network gradually approaches the physical and boundary condition constraints, thereby realizing physical field simulation.

[0097] The compressor blade has multi-scale and multi-physics coupling problems. In the process of applying PINNs, on the one hand, it is necessary to design a reasonable network architecture, and on the other hand, it is necessary to apply adaptive weights, precise boundary conditions, causal training and other technologies to achieve efficient and accurate fitting of neural networks to multi-physics fields. Specifically, it includes:

[0098] 1) Multi-scale Fourier network architecture

[0099] Neural networks tend to learn low-frequency features. The Fourier network architecture adds a Fourier embedding layer before the fully connected network to convert the input into a high-dimensional Fourier space, thereby effectively learning high-frequency features or sharp gradients that are prevalent in multi-scale structures. Figure 4 shown.

[0100] The multi-scale Fourier network embeds the input into several different Fourier feature maps and concatenates them through a linear layer after forward propagation, allowing all frequency components to be learned at the same convergence rate. Compared to the traditional fully connected network PINN model, both proposed architectures do not introduce any additional trainable parameters and do not require more floating-point operations to evaluate their forward or backward passes. Therefore, they can serve as alternatives to traditional fully connected architectures without sacrificing computational efficiency.

[0101] 2) Improved Fourier network architecture

[0102] In the Fourier network, a standard fully connected neural network is used as a nonlinear mapping between Fourier features and model output. In the improved network, based on the neural attention mechanism for computer vision and natural language processing tasks, the following framework is proposed:

[0103]

[0104] in, U , V It represents the feature representation of the input after being transformed by two different network branches, which is used for subsequent gated fusion; X is the input tensor of the neural network; W 1 、 W 2 is the weight matrix of the network; b 1 、 b 2 is the corresponding bias term; is the activation function; H (1) is the output of the first hidden layer, W z,1 、 b z,1 are the weights and biases of the initial hidden layer; Z ( k ) is the k The gating tensor of the layer; H ( k ) is the k The hidden state output of the layer; f θ ( x ) is the output of the network;H (L+1) is the final hidden state output; W 、 b are the weights and biases of the final output layer.

[0105] The key features of this framework are: it explicitly accounts for the multiplicative interactions between different input dimensions; and it enhances the hidden state through residual connections. In practical applications of multi-scale problems, the modified Fourier network has been shown to significantly outperform other networks in terms of convergence speed and accuracy. This is shown in Figures 5(a) and 5(b).

[0106] 3) Adaptive weight

[0107] The main approach to training PINNs is to express the initial and boundary constraints as additional penalty terms in the loss function. This is usually done by multiplying the parameters by Each of these terms is adjusted to balance the contribution of each term to the overall loss. However, manually adjusting these parameters is not simple and they need to be treated as constants. This example uses the learning rate annealing method to adaptively update the weights. Assume that the loss function of a certain steady-state problem is:

[0108]

[0109] in, L ( θ ) is the total loss function of the neural network; L residual ( θ ) is the residual loss term; L BC ( θ ) is the boundary condition loss term; λ (i) For the i The penalty coefficient used for weighting at the iteration; θ are all the trainable parameters of the neural network.

[0110] In each training iteration, the weights are updated based on the gradient information:

[0111]

[0112] in, For the i The boundary condition loss weighting coefficient used for adaptive update in the iteration;

[0113] For the i At the iteration, the residual loss function is about the neural network parameters θ gradient; For the iAt the iteration, the boundary condition loss function is about the neural network parameters θ gradient.

[0114] And calculate the exponentially weighted average:

[0115]

[0116] in, λ (i) For the i The weighting coefficients in the iteration; α is the smoothing factor.

[0117] Adaptive updating of weights can effectively balance the learning speed of each loss and improve the convergence speed and accuracy of PINNs.

[0118] 4) Causal Training

[0119] For a transient problem, assume that the following time-dependent partial differential equations need to be solved:

[0120]

[0121] in, is the position controlled by the PDE system; represents the nonlinear differential operator. Continuous-time PINNs models may violate temporal causality and easily converge to incorrect solutions. Causal training aims to address this issue by reformulating the PDE residual loss, explicitly accounting for physical causality during model training.

[0122] The time domain Divided into subdomains , and define the The loss function of the subdomain is:

[0123]

[0124] in, The total causal loss is given by:

[0125]

[0126] in:

[0127]

[0128] in For the i The causal weight of a sub-time domain is inversely proportional to the size of the cumulative residual loss of the previous self-domain, so before all previous losses are minimized, It will not be minimized; This simple algorithm forces PINNs to gradually learn the solution function of the PDE, respecting the intrinsic causal structure of its dynamic evolution. Figure 6 Demonstrated improvements in convergence speed and accuracy with causal training.

[0129] Thermal-fluid-solid coupling simulation is common in the design of compressor blisks, and PINNs offer significant advantages in multi-physics coupling analysis. By directly embedding physical laws into neural networks, PINNs effectively capture the complex interactions between different physical fields, ensuring that simulation results conform to underlying physical principles. PINNs are also capable of processing high-dimensional data and maintaining high accuracy even when data is scarce. Furthermore, the flexibility of PINNs enables them to adapt to a variety of boundary conditions and operating conditions, providing real-time predictive capabilities and significantly improving the efficiency and reliability of multi-physics coupling analysis.

[0130] 3. Fast flow field prediction method based on graph convolutional neural network

[0131] 1) Convolutional Neural Networks

[0132] When applied to CFD predictions, general CNN-type network models require structured input data similar to images. However, CFD simulations often encounter irregular geometric shapes and computational domains. The general approach is to directly pixelate the geometric shapes, computational domains, etc., that is, interpolate this data onto a structured grid, and input each grid point into the CNN as a pixel. A typical example is to calculate the wall function (Signed Distance Function) on a regular Cartesian grid to represent the geometric shape and input it into the CNN model as an image. This approach has both obvious advantages and disadvantages. The advantage is that it is convenient and flexible to handle geometric shapes, while the disadvantage is that the pixelization operation is more likely to lose geometric and physical information.

[0133] In addition to pixelation, for the problem of relatively regular geometric shapes, this embodiment transforms the physical space grid into a regular Cartesian grid in the computational space through coordinate transformation, so that it can better adapt to the convolutional neural network. For relatively regular geometric shapes such as turbine blades, the grid on the surface of the turbine blade is transformed into a corresponding Cartesian grid, and image-like data is obtained. The prediction model for rapid simulation is trained through the convolutional neural network. The complete process is as follows: Figure 7 shown.

[0134] 2) Graph Convolutional Neural Network Based on Grid Topology Connection

[0135] GCN usually stores CFD data in a graph-based structure. ,in represents the points in the graph, It means the edge connecting the nodes of the graph, and both the nodes and the edges will have corresponding features X v and X e ,like Figure 8 As shown. The graph convolutional network performs convolution operations on these features through their connection relationships and finally obtains the desired output. Taking the grid method storage commonly used in the finite volume method as an example, the flow field solution data (usually including physical quantities such as velocity and pressure) are directly stored at the grid points, that is, the nodes of the topological connection graph. For another commonly used grid center storage in the finite volume method, that is, the flow field solution is stored at the center of the grid unit, Figure 9 A method for converting flow field data into a topological connection graph for storage is proposed: by defining the grid centers as nodes of the graph and then defining the edges connecting the nodes, the flow field data can be converted into a storage format based on the topological connection graph.

[0136] After the above transformation, this embodiment converts the flow field data into a data storage format based on grid topology connection, such as Figure 10 The features of nodes and edges in the topological connection graph are stored in matrix form, with each column being the feature vector of the node or edge. The connection relationship between the nodes represented by each edge in the topological connection graph is also stored in matrix form, with each column corresponding to the node number on each edge.

[0137] Based on the data stored in the above topological connection graph format, the graph convolution operator is defined to realize the information transmission and aggregation on the topological graph to complete the prediction of the flow field solution. The input of the graph neural network includes the feature information on the nodes and edges, as well as the topological connection relationship of the graph. The graph convolution operator defines the operation of information transmission and aggregation between nodes and edges, which can be Figure 11 It can also be written as follows:

[0138]

[0139] in , Generally, a fully connected neural network is used for fitting. and It represents the trainable parameters of the network. Generally speaking, the performance of the neural network model is improved by increasing the number of network layers and the number of features in each layer to build a deep graph convolutional network. Figure 12 shown.

[0140] 4. Method for building an artificial intelligence agent model based on deep learning

[0141] To build an artificial intelligence agent model, you first need to parameterize the geometry and boundary conditions and train a deep learning model. Then, the parameters are trained as input to the neural network. Once the training is completed, you can achieve rapid prediction of the physical field under specific parameter structures or boundary conditions.

[0142] 1) Parameterized geometry and boundary conditions

[0143] Parametric modeling mainly consists of modeling engine, parameters, and model templates.

[0144] Modeling Engine: The modeling engine is the core package of all geometric elements, involving the properties and operations of geometric objects, such as the creation, modification, and Boolean operations of points, lines, surfaces, and bodies. Any 3D geometric model can be created through the modeling engine.

[0145] Parameters: Parameters are model attributes, essentially data. Considering a specific compressor blisk model, these parameters typically represent the geometric dimensions and characteristic parameters associated with the blisk. By driving the model template with parameters, the relevant model entities can be generated.

[0146] Model templates: Model templates are not model entities, but rather a model framework with general properties. Given specific parameter values, a corresponding model entity can be created. The universal nature of model templates makes parameter-driven modeling possible and provides a foundation for rapid modeling of compressor blisks.

[0147] 2) Intelligent agent model construction

[0148] An important advantage of neural network solvers (PINNs or operator learning) over traditional numerical methods is the ability to solve problems with parameterized geometry or boundary conditions by taking parameters as input. To illustrate the basic principles of the artificial intelligence agent model, take a simple equation as an example:

[0149]

[0150] in is a geometric parameter, As the input of the neural network, by training a neural network To achieve different geometric parameters Rapid prediction of physical fields.

[0151] The AI ​​agent model has the following advantages over the traditional agent model:

[0152] (1) Embedding physical information: PINNs directly incorporate physical laws into the model, so that the simulation results not only rely on data but also conform to basic physical principles, thereby improving the reliability of predictions.

[0153] (2) Small or no data requirement: Traditional proxy models usually rely on a large amount of data for training, while PINNs can still operate effectively in scenarios where the experimental cost is high or data is difficult to obtain.

[0154] (3) Low risk of overfitting: Since PINNs incorporate physical constraints, the complexity of the model is effectively controlled, thereby reducing the risk of overfitting.

[0155] (4) Efficient processing of high-dimensional problems: PINNs can effectively process high-dimensional input problems and adapt to complex geometric shapes and multi-physics coupling analysis. Its flexible network structure enables it to effectively model in multi-dimensional space.

[0156] Real-time prediction capability: PINNs have strong real-time prediction capabilities and can quickly generate simulation results under different geometries and working conditions.

[0157] The embodiment of the present invention proposes a multidisciplinary rapid simulation method for a compressor integral blade disk involving aerodynamic performance, structural design, heat transfer analysis, and strength analysis, thereby realizing rapid simulation of multiple physical fields of the compressor integral blade disk, thereby improving design efficiency and simulation accuracy.

[0158] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A multi-physics field rapid simulation method for a compressor blade disk, characterized in that: include: Establish a parametric geometric model of the compressor blade and automatically divide the aerodynamic flow domain grid and blade structure grid; Based on the parameterized geometric model, automatic simulation analysis including aerodynamic performance, temperature field, static strength, modal and life is carried out in sequence to obtain the gas-thermal-solid multidisciplinary coupling evaluation results of the compressor integral blade; Based on the multidimensional data from the gas-thermal-solid multidisciplinary coupled evaluation process of the compressor blade, an intelligent simulation agent model based on a physical information neural network, a flow field rapid prediction agent model based on a graph convolutional neural network, and an artificial intelligence agent model based on deep learning are respectively constructed. Based on the intelligent simulation agent model, the flow field rapid prediction agent model, and the artificial intelligence agent model, a rapid coupled simulation of the multi-physical fields of the compressor blade is realized; Based on the parameterized geometric model, automatic simulation analysis of aerodynamic performance, temperature field, static strength, modal and life is performed in sequence, including: Based on the preset geometric structural parameters of the compressor blisk, the aerodynamic load distribution on the blisk surface is calculated to evaluate the aerodynamic performance of the compressor blisk; The aerodynamic load distribution on the blisk surface is used as a thermal boundary condition to solve the blade-disk temperature field boundary to evaluate the compressor blisk temperature field; Calculating the stress distribution of the compressor blisk based on the aerodynamic load distribution on the blisk surface and the temperature field boundary obtained by the aero-thermal coupling calculation to perform a static strength assessment of the compressor blisk; Calculating the static frequency and dynamic frequency of the compressor blisk based on the aerodynamic load distribution on the surface of the blisk, the temperature field boundary, and the stress distribution to perform a modal evaluation of the compressor blisk; Based on the static strength assessment results and modal assessment results, and according to the material parameters of the compressor blisk, the compressor blisk life assessment is carried out.

2. The multi-physics field rapid simulation method for a compressor blade according to claim 1 is characterized in that: Establish a parametric geometric model of the compressor blade, including: According to the preset geometric structure parameters of the compressor integral disk, the parametric geometric model of the compressor integral disk is automatically generated through the CAD engine, and the aerodynamic flow domain grid and the disk structure grid are automatically divided.

3. The multi-physics field rapid simulation method for a compressor blade according to claim 2 is characterized in that: The preset geometric structure parameters include aerodynamic boundaries, air system boundaries, rotational speed boundaries, material parameters and structural constraint boundaries.

4. The multi-physics field rapid simulation method for a compressor blade according to claim 1 is characterized in that: Also includes: Based on the results of the multidisciplinary gas-thermal-solid coupled evaluation of the compressor blisk, it is determined whether the temperature field, stress distribution, and lifespan meet the set requirements. If any abnormal problems exist, such as local overtemperature, excessive local stress, or short lifespan, the parameterized geometric model of the compressor blisk is fed back and optimized, and iteration is performed again until a compressor blisk design result that meets the design criteria is obtained.

5. The multi-physics field rapid simulation method for a compressor blade according to claim 1, characterized in that: Construct an intelligent simulation agent model based on physical information neural network, including: Based on the multi-scale Fourier network architecture, a neural attention mechanism is introduced to clarify the multiplicative interaction between different input dimensions and enhance the hidden state through residual connections; The weights are adaptively updated through the learning rate annealing method to balance the learning speed of each loss.

6. The multi-physics field rapid simulation method for a compressor blade according to claim 5, characterized in that: Building an intelligent simulation agent model based on a physical information neural network also includes: The time domain is divided into multiple subdomains, the PDE residual loss is reformulated for the subdomains, and the physical causal relationship during model training is clarified to perform causal training on the model.

7. The multi-physics field rapid simulation method for a compressor blade according to claim 1, characterized in that: The flow field fast prediction agent model based on graph convolutional neural network includes: The image data of the compressor blades is transformed from the physical space grid to a regular Cartesian grid in the computational space through coordinate transformation, so that the data can be adapted to the convolutional neural network to achieve rapid prediction of the flow field.

8. The multi-physics field rapid simulation method for a compressor blade according to claim 7 is characterized in that: The flow field fast prediction agent model based on graph convolutional neural network also includes: The simulation data is stored in a topological connection structure based on graph representation, and the information is transmitted and aggregated on the topological connection graph through graph convolution operations to complete the prediction of the flow field solution.

9. The multi-physics field rapid simulation method for a compressor blade according to claim 1, characterized in that: Build an AI agent model based on deep learning, including: Parameterization of geometry and boundary conditions through modeling engines, model parameters, and model templates; Using a neural network solver, parameterized geometry and boundary conditions are used as input for training the neural network to achieve rapid prediction of the physical field under specific parameter structures or boundary conditions.

Images