A collaborative optimization method and system for conceptual design of gas turbine turbine blades

Abstract

The application provides a collaborative optimization method and system for conceptual design of a gas turbine turbine blade, the method comprising: constructing a full-parameterized main model for uniformly describing blade aerodynamic, cooling and structural characteristics through a design parameter vector, and presetting correlation and constraint rules among parameters; defining a multi-disciplinary optimization target vector comprising aerodynamic efficiency, cooling efficiency and structural strength indexes; constructing a multi-disciplinary unified proxy model with the design parameter vector as input and simultaneously outputting multiple performance prediction values, and performing supervised training by using an intelligent agent constructed based on a large model; taking the trained proxy model as an evaluator, calling a multi-objective optimization algorithm for iterative search, and outputting a Pareto optimal solution set; performing manufacturability filtering and visual screening on schemes in the solution set, performing simulation verification on an optimal scheme, and generating a blade design data package. The technical scheme provided by the application significantly shortens the conceptual design cycle and realizes global optimal balance of multi-disciplinary performance.

Description

Technical Field

[0001] This application relates to the field of gas turbine technology, and in particular to a collaborative optimization method and system for the conceptual design of gas turbine blades. Background Technology

[0002] Heavy-duty gas turbines play a central role in low-carbon power systems, and their turbine blades, as key components that withstand the impact of high-temperature, high-pressure gas combustion, undergo extremely complex design processes. Turbine blade design typically involves multiple disciplines, including aerodynamic design, cooling design, and structural integrity analysis, with close interrelationships among these disciplines. According to the design cycle, the design work can be divided into three stages: conceptual design, preliminary design, and detailed design. The conceptual design stage primarily involves rapid iteration across multiple disciplines based on performance models, experience, or tools to form a conceptual design scheme acceptable to all disciplines. The quality of this scheme has a decisive impact on the success or failure of subsequent design stages.

[0003] Traditional turbine blade conceptual design generally adopts a sequential iterative approach, with a typical process as follows: first, aerodynamic design is performed to form multi-stage turbine blade airfoil schemes; then, cooling design is carried out to construct cooling structures for different stages of the blade; finally, structural integrity analysis is performed to evaluate blade strength and lifespan. Only after all design indicators are qualified can the casting and machining stages begin. This sequential design method has many drawbacks. Because the design goals of different disciplines are mutually constrained, failure to meet requirements at any stage may lead to design rework. For example, in the aerodynamic design stage, if the flow design results do not meet efficiency requirements, a redesign is necessary; if the airfoil design does not meet casting process requirements, the airfoil must also be readjusted. In the cooling design stage, if the blade leading edge curvature does not meet cooling requirements, the airfoil must be modified; if the cooling volume exceeds the limit, the cooling scheme must be redesigned; if the cooling structure cannot be achieved through casting, the cooling structure must be readjusted. In the structural integrity analysis stage, if the blade lifespan does not meet requirements, the cooling design stage must be returned to for redesign. As can be seen, the traditional sequential design method involves a large amount of iterative rework within and between different disciplines, and each return consumes significant human and computational resources. Meanwhile, due to the numerous factors involved, obtaining a feasible design that satisfies all constraints requires repeated coordination and tracking by multiple engineers, resulting in significant labor costs. Furthermore, the serial iterative approach, limited by disciplinary barriers and local optimization thinking, struggles to perform a global search within the design space, often only yielding feasible solutions that satisfy basic constraints, rather than truly multidisciplinary globally optimal solutions. Therefore, there is an urgent need to propose a design scheme that can overcome the bottleneck of multidisciplinary coupling in the turbine blade conceptual design stage, improve design efficiency, and achieve global optimization. Summary of the Invention

[0004] This application provides a collaborative optimization method and system for the conceptual design of gas turbine blades, which at least solves the technical problems of low design efficiency, high manual coordination costs, and difficulty in obtaining the global optimal solution of multiple disciplines caused by the serial iteration of aerodynamics, cooling, and structure in traditional turbine blade conceptual design.

[0005] A first aspect of this application proposes a collaborative optimization method for the conceptual design of gas turbine blades, the method comprising:

[0006] A fully parameterized master model and a multidisciplinary optimization objective vector are constructed to define the blade through a design parameter vector. The design parameter vector includes a subset of aerodynamic design parameters, a subset of cooling design parameters, and a subset of structural design parameters. Each subset has preset parameter association and constraint rules. The multidisciplinary optimization objective vector includes aerodynamic efficiency index, cooling efficiency index, and structural strength index. A multidisciplinary unified agent model is constructed with the design parameter vector as input and the predicted values ​​of each index in the multidisciplinary optimization target vector as output. Then, the multidisciplinary unified agent model is trained by an agent based on a large model. The agent is used to control the simulation platform to generate training and validation data, and decide whether to supplement data and retrain based on the validation results. Using the trained multidisciplinary unified surrogate model as the evaluator and the design space of the fully parameterized master model as the search space, a multi-objective optimization algorithm is invoked for iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted value output by the multidisciplinary unified surrogate model and drives the iterative search to obtain a Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamic, cooling, and structural performance. The schemes in the Pareto optimal solution set are subjected to manufacturability filtering and visualization screening to obtain the preferred scheme. Then, the preferred scheme is verified by simulation to generate a blade design data package. The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index.

[0007] Preferably, the parameter association and constraint rules include: The correlation rules between blade thickness distribution and cooling channel layout; The relationship between the leading and trailing edge diameter of the blades and the dimensions of the cooling structure; The correlation between the average temperature of the blade cross section and the creep strength; The correlation rules between blade profile pressure and blade section aerodynamic bending moment.

[0008] Preferably, the subset of aerodynamic design parameters includes: blade inlet and outlet geometry angles and throat width; The subset of cooling design parameters includes: cooling channel layout and leading and trailing edge diameter; The subset of structural design parameters includes: blade wall thickness distribution.

[0009] Preferably, the multidisciplinary unified agent model is constructed using a multi-task learning neural network architecture; The multi-task learning neural network architecture includes: a shared low-level feature extraction network and multiple parallel subject-specific output branch networks; The shared low-level feature extraction network is used to extract shared geometric and physical features from the input design parameter vector; The subject-specific output branch networks are used to predict aerodynamic efficiency, cooling efficiency, and structural strength indices, respectively.

[0010] Preferably, the step of employing agent-supervised training based on a large model to construct the multidisciplinary unified agent model includes: Step S1: Construct the agent based on the large language model. During the fine-tuning process, the agent learns the design rules, physical constraints and historical case knowledge of the turbine blade. Step S2: The intelligent agent control simulation platform generates an initial training dataset and performs initial training on the multidisciplinary unified agent model; Step S3: The intelligent agent control simulation platform generates a brand-new verification dataset to verify the trained multidisciplinary unified agent model; Step S4: When the verification accuracy is less than the preset verification accuracy threshold, the agent analyzes and predicts the error distribution, identifies the region where the error exceeds the threshold, calls high-fidelity simulation to generate a supplementary dataset in the region, and retrains the multidisciplinary unified agent model based on the expanded dataset. Step S5: Repeat steps S3 and S4 until the verification accuracy is greater than or equal to the preset verification accuracy threshold.

[0011] Preferably, before invoking the multi-objective optimization algorithm for iterative search, the method further includes: The initial population generation module of knowledge enhancement is invoked to generate an initial design scheme set, which is then used as the initial population for the optimization algorithm. The knowledge-enhanced initial population generation module includes a variational autoencoder trained on a database of historical excellent design cases; the variational autoencoder is used to learn the parameter distribution patterns of historical cases and generate new design schemes based on the learned distribution. When generating design schemes, the variational autoencoder controls the diversity of the generated schemes by performing Gaussian sampling in the latent space and setting temperature parameters.

[0012] Furthermore, the multi-objective optimization algorithm is the NSGA-III algorithm.

[0013] Preferably, the iterative search process further includes: an active learning strategy, the active learning strategy comprising: The Monte Carlo Dropout method is used to quantify the prediction uncertainty of the multidisciplinary unified surrogate model for the current candidate design parameters; When the prediction uncertainty is greater than a preset prediction uncertainty threshold, the candidate design parameters are determined to be located in a sparse data region. The high-fidelity simulation was used to calculate the parameters of the candidate design and obtain the high-fidelity simulation results. The high-fidelity simulation results are used as real performance data to correct the predictions of the multidisciplinary unified agent model for the region. The corrected prediction or the high-fidelity simulation result is returned to the optimization algorithm to guide the subsequent search direction.

[0014] Preferably, the process of performing manufacturability filtering and visualization screening on the solutions in the Pareto optimal solution set to obtain preferred solutions includes: The manufacturability knowledge rule base is invoked to evaluate the casting feasibility and machining complexity of each scheme in the Pareto optimal solution set, and the manufacturability risk level of each scheme is marked according to the evaluation results. It provides a multi-dimensional performance parallel coordinate graph, which visualizes the aerodynamic efficiency index, cooling efficiency index, structural strength index and manufacturability risk level of each scheme in the Pareto solution set in the same coordinate system. The system receives filtering instructions from experts via the parallel coordinate graph, filters out schemes with manufacturability risk levels exceeding a threshold from the Pareto solution set, and selects one or more preferred schemes.

[0015] Furthermore, the manufacturability knowledge rule base includes: minimum wall thickness casting limit rules, air film hole machining accessibility rules, and cooling channel complexity rating rules.

[0016] The technical solutions provided by the embodiments of this application bring at least the following beneficial effects: This application proposes a collaborative optimization method and system for the conceptual design of gas turbine blades. The method includes: constructing a fully parameterized master model and a multidisciplinary optimization objective vector that define the blade through a design parameter vector. The design parameter vector includes a subset of aerodynamic design parameters, a subset of cooling design parameters, and a subset of structural design parameters, with preset parameter association and constraint rules between each subset. The multidisciplinary optimization objective vector includes aerodynamic efficiency indicators, cooling efficiency indicators, and structural strength indicators. A multidisciplinary unified proxy model is constructed, taking the design parameter vector as input and the predicted values ​​of each indicator in the multidisciplinary optimization objective vector as output. Then, an agent based on a large model is used to supervise the training of the multidisciplinary unified proxy model. The agent controls the simulation platform to generate training and validation data and decides whether to supplement data and retrain based on the validation results. The trained multidisciplinary unified proxy model is used as... The evaluator uses the design space of the fully parameterized master model as the search space and calls a multi-objective optimization algorithm for iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted values ​​output by the multidisciplinary unified surrogate model and drives the iterative search to obtain a Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamics, cooling, and structure. The schemes in the Pareto optimal solution set are then subjected to manufacturability filtering and visualization screening to obtain the preferred scheme. Simulation verification is then performed on the preferred scheme to generate a blade design data package. The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index. The technical solution proposed in this application significantly shortens the conceptual design cycle and achieves a globally optimal balance of multidisciplinary performance.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0018] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a collaborative optimization method for conceptual design of gas turbine blades according to an embodiment of this application; Figure 2 This is a detailed schematic diagram of a collaborative optimization method for the conceptual design of gas turbine blades according to an embodiment of this application; Figure 3This is a structural diagram of a collaborative optimization system for the conceptual design of gas turbine blades according to an embodiment of this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0020] This application proposes a collaborative optimization method and system for the conceptual design of gas turbine blades. The method includes: constructing a fully parameterized master model and a multidisciplinary optimization objective vector that define the blade using design parameter vectors. The design parameter vectors include a subset of aerodynamic design parameters, a subset of cooling design parameters, and a subset of structural design parameters, with preset parameter association and constraint rules between each subset. The multidisciplinary optimization objective vectors include aerodynamic efficiency indicators, cooling efficiency indicators, and structural strength indicators. A multidisciplinary unified proxy model is constructed, taking the design parameter vectors as input and the predicted values ​​of each indicator in the multidisciplinary optimization objective vectors as output. Then, an agent based on a large model is used to supervise the training of the multidisciplinary unified proxy model. The agent controls the simulation platform to generate training and validation data and decides whether to supplement data and retrain based on the validation results. The trained multidisciplinary unified proxy model is used as... The evaluator uses the design space of the fully parameterized master model as the search space and calls a multi-objective optimization algorithm for iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted values ​​output by the multidisciplinary unified surrogate model and drives the iterative search to obtain a Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamics, cooling, and structure. The schemes in the Pareto optimal solution set are then subjected to manufacturability filtering and visualization screening to obtain the preferred scheme. Simulation verification is then performed on the preferred scheme to generate a blade design data package. The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index. The technical solution proposed in this application significantly shortens the conceptual design cycle and achieves a globally optimal balance of multidisciplinary performance.

[0021] The following describes a collaborative optimization method for the conceptual design of gas turbine blades according to an embodiment of this application, with reference to the accompanying drawings.

[0022] Example 1 Figure 1This is a flowchart illustrating a collaborative optimization method for the conceptual design of gas turbine blades according to an embodiment of this application, such as... Figure 1 As shown, the method includes: Step 1: Construct a fully parameterized master model of the blade and a multidisciplinary optimization target vector by defining the design parameter vector. The design parameter vector includes a subset of aerodynamic design parameters, a subset of cooling design parameters, and a subset of structural design parameters. Each subset has preset parameter association and constraint rules. The multidisciplinary optimization target vector includes aerodynamic efficiency index, cooling efficiency index, and structural strength index. In this embodiment of the disclosure, the parameter association and constraint rules include: The correlation rules between blade thickness distribution and cooling channel layout; The relationship between the leading and trailing edge diameter of the blades and the dimensions of the cooling structure; The correlation between the average temperature of the blade cross section and the creep strength; The correlation rules between blade profile pressure and blade section aerodynamic bending moment.

[0023] In this embodiment of the disclosure, the subset of aerodynamic design parameters includes: blade inlet and outlet geometry angles and throat width; The subset of cooling design parameters includes: cooling channel layout and leading and trailing edge diameter; The subset of structural design parameters includes: blade wall thickness distribution.

[0024] It should be noted that a fully parameterized master model of the turbine blade is constructed. This model fully defines the blade through a unified design parameter vector X = (X1, X2, X3) of dimension N, where X1 is a subset of aerodynamic design parameters, X2 is a subset of cooling design parameters, and X3 is a subset of structural design parameters. There are parameter relationships and constraints between the subsets. The construction of a fully parameterized master model includes: parameter decoupling and association definition, specifically: Identify and define core driving parameters that simultaneously affect the performance of multiple disciplines; Establish association rules between parameters. When the core driving parameters change, adjust the subordinate parameters through the association rules to ensure geometric and physical rationality.

[0025] The core driving parameters include the main characteristic parameters of each discipline, such as aerodynamic blade parameters, including inlet and outlet geometry angles, throat width, etc., and cooling channel layout parameters of the cooling discipline; the correlation rules include: the correlation between the leading and trailing edge diameter of the blade and the local cooling structure size, the correlation between the blade thickness distribution and the cooling channel layout, etc.; the correlation between the average temperature of the blade section and the local creep strength, and the correlation between the blade profile pressure and the aerodynamic bending moment of the blade section, etc., of the structural integrity analysis discipline.

[0026] Define a multidisciplinary optimization objective vector F(X) = [f1(X), f2(X), f3(X), ...], which includes at least: aerodynamic efficiency index f1, cooling efficiency index f2, blade section stress f3, and blade root contact surface stress f4; The structural strength indicators include: blade section stress f3 and blade root contact surface stress f4.

[0027] It should be noted that the fully parameterized master model can define the driving parameters in relation to each other, ensuring that any parameter change is immediately reflected in a physically consistent blade geometry, laying the foundation for subsequent collaborative optimization.

[0028] Example: Define "internal cooling channel structure of the blade" as a driving parameter. When the optimization algorithm adjusts this parameter, the system not only moves the position of the baffle, but also adjusts the following in conjunction: a) the baffle thickness according to the blade wall thickness; b) the thickness distribution of the blade profile; c) the position of the centroid of the blade section.

[0029] Step 2: Construct a multidisciplinary unified agent model with the design parameter vector as input and the predicted values ​​of each index in the multidisciplinary optimization target vector as output. Then, use an agent based on the large model to supervise the training of the multidisciplinary unified agent model. The agent is used to control the simulation platform to generate training and verification data, and decide whether to supplement the data and retrain based on the verification results. In this embodiment of the disclosure, the multidisciplinary unified agent model is constructed using a multi-task learning neural network architecture; The multi-task learning neural network architecture includes: a shared low-level feature extraction network and multiple parallel subject-specific output branch networks; The shared low-level feature extraction network is used to extract shared geometric and physical features from the input design parameter vector; The subject-specific output branch networks are used to predict aerodynamic efficiency, cooling efficiency, and structural strength indices, respectively.

[0030] Furthermore, the method of employing agent-supervised training based on a large model to train the multidisciplinary unified agent model includes: Step S1: Construct the agent based on the large language model. During the fine-tuning process, the agent learns the design rules, physical constraints and historical case knowledge of the turbine blade. Step S2: The intelligent agent control simulation platform generates an initial training dataset and performs initial training on the multidisciplinary unified agent model; Step S3: The intelligent agent control simulation platform generates a brand-new verification dataset to verify the trained multidisciplinary unified agent model; Step S4: When the verification accuracy is less than the preset verification accuracy threshold, the agent analyzes and predicts the error distribution, identifies the region where the error exceeds the threshold, calls high-fidelity simulation to generate a supplementary dataset in the region, and retrains the multidisciplinary unified agent model based on the expanded dataset. Step S5: Repeat steps S3 and S4 until the verification accuracy is greater than or equal to the preset verification accuracy threshold.

[0031] It should be noted that a multidisciplinary unified agent model M_unified is constructed and trained. This model takes the design parameter vector X as input and outputs the calculated values ​​F'(X) of each index in the multidisciplinary optimization target vector F(X) in one synchronous operation. The calculated values ​​can be obtained by fusing historical simulation data, experimental data and reduced-order physical models. The multidisciplinary unified agent model M_unified is built using a multi-task learning neural network architecture, which includes a shared low-level feature extraction network for learning shared geometric and physical features from the input parameters X, and multiple parallel discipline-specific output branch networks for predicting aerodynamic, cooling, and structural performance indicators, respectively.

[0032] A supervised optimization agent is constructed using a large model. This agent serves as the core to guide the construction of the proxy model and clarifies its specific construction direction. During the training phase, data generated by a multi-disciplinary coupled design simulation platform (referred to as the "integrated design platform") under the agent's control is used as the source dataset for training the proxy model. In the validation phase, the agent calls the integrated design platform to generate a new dataset, which is used as the validation set to verify the effectiveness of the trained proxy model. If the model's validation accuracy does not meet the preset standard, the agent will invoke a high-fidelity simulation calculation process to generate a supplementary dataset and complete data expansion. The proxy model will then be retrained based on the expanded dataset.

[0033] By leveraging the agent's driving mechanism, the size of the dataset required for training the agent model can be significantly reduced. Compared to traditional model building methods without agent-driven approaches, this solution can complete model training with less data, effectively reducing the model's dependence on training data.

[0034] It should be noted that the construction process of the intelligent agent includes: 1. Base Model Selection First, select a general-purpose language model as a base, such as the GPT series, Llama series, or GLM series models. This base model possesses fundamental capabilities such as natural language understanding and generation, context learning, and reasoning, providing support for subsequent domain-specific fine-tuning.

[0035] 2. Domain Knowledge Base Construction A dedicated knowledge base for turbine blade design is constructed, which includes the following three categories of knowledge.

[0036] The first category is the design rule library, which includes blade geometric constraints, such as minimum wall thickness limits, the matching relationship between leading and trailing edge diameters and cooling channel dimensions, reasonable ranges for aerodynamic parameters, and design specifications for cooling structures. These rules are mainly derived from design manuals and internal company standards.

[0037] The second category is a physical constraint knowledge base, including the physical laws of gas-thermal-solid coupling, the allowable temperature and stress limit of materials, and empirical correlations between cooling efficiency and cold air flow. This knowledge mainly comes from physical models and empirical formulas.

[0038] The third category is the historical case database, which includes the design parameter vector X of existing turbine blades and the corresponding simulation performance data F. This database contains both successful and unsuccessful design cases. This data primarily originates from historical simulation and experimental data.

[0039] 3. Fine-tuning training Using the constructed domain knowledge base, the base large language model is fine-tuned and trained to enable it to have reasoning capabilities in the field of turbine blade design. The fine-tuning process includes the following aspects.

[0040] First, fine-tune the instructions. Construct "question-answer" pairs for the training data. For example, the question could be, "Given the validation error distribution of the current agent model, which regions have the highest prediction uncertainty?", and the answer could be, "The leading edge region and the leaf root transition region; it is recommended to supplement sampling within the following parameter range." Through training with a large number of similar task instructions, the agent learns how to respond to design-related decision requests.

[0041] Second, retrieval-enhanced generation. The content in the knowledge base is vectorized and indexed. When the agent needs to make a decision, it first retrieves the most relevant knowledge from the knowledge base and then generates decision instructions based on the retrieval results. This approach ensures that the agent's decision-making is always accurate and reliable.

[0042] Third, training reasoning ability. The agent is trained to have multi-step reasoning ability, such as being able to complete the following reasoning chain in sequence: analyzing and verifying the error distribution, locating high error regions, determining the cause of data sparsity, deciding on the specific strategy for supplementary sampling, and finally generating executable simulation task instructions.

[0043] 4. Encapsulation of decision-making capabilities The trained agent is encapsulated into a callable module, providing the following decision-making interfaces to the outside world.

[0044] The first interface is the training strategy generation interface. It takes the current agent model architecture and the existing data scale as input and outputs the sampling strategy for the initial training data, which is used to guide the generation of the initial training dataset.

[0045] The second interface is the evaluation and validation result interface. It takes the prediction error distribution of the surrogate model on the validation set as input and outputs whether the accuracy meets the standard and the location information of the unacceptable areas, which is used to determine whether retraining is needed.

[0046] The third interface is the supplementary strategy generation interface. It takes as input information about regions where the prediction error exceeds a threshold and outputs the parameter range and sampling density for supplementary sampling, which is used to generate a supplementary dataset.

[0047] The fourth interface is for calling the simulation platform interface, which takes as input a list of parameter vectors for supplementary sampling and outputs simulation task instructions for performing high-fidelity simulation calculations.

[0048] The intelligent agent is located at the core scheduling layer of the entire optimization method, below which is the integrated design simulation platform, and above which is the multidisciplinary unified agent model.

[0049] In terms of workflow, the agent first calls the integrated design simulation platform to generate an initial training dataset, which is then delivered to the agent model for initial training. After the agent model is trained, the agent calls the integrated design simulation platform again to generate a new validation dataset to verify the performance of the trained agent model.

[0050] When the verification accuracy fails to meet the preset standard, the agent analyzes the prediction error distribution, identifies regions where the error exceeds a threshold, and determines these regions as sparse data areas. Then, the agent decides on a supplementary sampling strategy, including determining the parameter range, sampling density, and number of sampling points that need to be supplemented, and translates these strategies into specific design parameter vectors. Subsequently, the agent invokes an integrated design simulation platform to perform high-fidelity simulation calculations on these design parameter vectors to obtain a supplementary dataset. Finally, the agent model is retrained based on the expanded dataset.

[0051] This process is repeated until the validation accuracy of the agent model reaches the preset standard. Through the above integration method, the agent plays the role of strategist and scheduler throughout the entire training process of the agent model, ensuring that the training process is efficient and directed.

[0052] Through the above construction process, the intelligent agent achieves the following technical effects.

[0053] First, it is domain knowledge-driven. The agent's decision-making does not rely on blind sampling or random search, but rather on intelligent decision-making based on design rules, physical constraints, and historical experience, making the generation of training data more targeted.

[0054] Second, dynamic adaptability. The agent can dynamically adjust the data supplementation strategy based on the current validation results of the agent model, focusing on supplementing data in areas where the agent model performs poorly and reducing sampling in areas where it performs well, thereby optimizing the allocation of training resources.

[0055] Third, closed-loop processing is achieved. From the generation of initial training strategies, evaluation of verification results, decision-making based on supplementary data to the invocation of simulation tasks, the entire process is completed by the intelligent agent without human intervention, significantly reducing the coordination workload of engineers.

[0056] Fourth, it offers high data efficiency. Compared to traditional random sampling or Latin hypercube sampling methods, this approach reduces the training data requirement by approximately 30% to 50% while achieving the same prediction accuracy as the surrogate model, thereby significantly reducing the computational cost of high-fidelity simulation.

[0057] It should be noted that this embodiment uses agent-supervised training to create a unified model M_unified that can simultaneously output multi-disciplinary performance. Its advantages are: A supervised optimization agent is constructed to guide the construction of the agent model and provide specific directions for its development. Data generated by an ensemble design platform is used as source data for training. The agent then generates a new set of data through the same platform, serving as a new validation set. If the validation accuracy is low, the agent generates a new dataset to supplement the data before retraining. This agent-driven approach significantly reduces the amount of training data required. Without agent-driven training, a much larger amount of data would be needed.

[0058] Step 3: Using the trained multidisciplinary unified surrogate model as the evaluator and the design space of the fully parameterized master model as the search space, a multi-objective optimization algorithm is invoked for iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted value output by the multidisciplinary unified surrogate model and drives the iterative search to obtain the Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamic, cooling, and structural performance. The design space of the fully parameterized master model refers to the multidimensional continuous space formed by the possible value range of each parameter in the design parameter vector.

[0059] In this embodiment of the disclosure, before invoking the multi-objective optimization algorithm for iterative search, the method further includes: The initial population generation module of knowledge enhancement is invoked to generate an initial design scheme set, which is then used as the initial population for the optimization algorithm. The knowledge-enhanced initial population generation module includes a variational autoencoder trained on a database of historical excellent design cases; the variational autoencoder is used to learn the parameter distribution patterns of historical cases and generate new design schemes based on the learned distribution. When generating design schemes, the variational autoencoder controls the diversity of the generated schemes by performing Gaussian sampling in the latent space and setting temperature parameters.

[0060] It should be noted that the multi-objective optimization algorithm is the NSGA-III algorithm.

[0061] In this embodiment of the disclosure, the iterative search process further includes: an active learning strategy, the active learning strategy comprising: The Monte Carlo Dropout method is used to quantify the prediction uncertainty of the multidisciplinary unified surrogate model for the current candidate design parameters; When the prediction uncertainty is greater than a preset prediction uncertainty threshold, the candidate design parameters are determined to be located in a sparse data region. The high-fidelity simulation was used to calculate the parameters of the candidate design and obtain the high-fidelity simulation results. The high-fidelity simulation results are used as real performance data to correct the predictions of the multidisciplinary unified agent model for the region. The corrected prediction or the high-fidelity simulation result is returned to the optimization algorithm to guide the subsequent search direction.

[0062] It should be noted that the full-process collaborative optimization engine is established by using the multi-disciplinary unified agent model M_unified as the objective function evaluator, the design space of the fully parameterized master model as the search space, running a multi-objective optimization algorithm, performing parallel optimization on the optimization objective vector F(X), and directly outputting the global Pareto optimal solution set that achieves a balance in aerodynamics, cooling, and structural performance. Specifically, before optimization begins, the knowledge-enhanced initial population generation module is invoked. This module generates a set of initial design schemes that are both diverse and have high-performance potential based on the historical case library and the design rule knowledge graph, which serves as the initial population for the optimization algorithm. During the optimization iteration process, an active learning strategy with feasibility constraints is introduced: when the optimization algorithm explores a region with high uncertainty in the surrogate model prediction, a small number of high-fidelity simulation calculations are triggered, and the results are used to correct the surrogate model prediction and guide the optimization direction, balancing exploration and utilization.

[0063] It should be noted that the human-machine collaborative decision-making and verification of the solutions in the global Pareto optimal solution set includes rule-based manufacturability filtering and high-fidelity simulation verification. The final optimized solution and the corresponding complete data pair (X, F_true) are fed back to the training database to update the multidisciplinary unified agent model M_unified.

[0064] Human-machine collaborative decision-making and verification include: It provides visualization tools such as multidimensional performance parallel coordinate plots to help experts intuitively compare various schemes on the Pareto frontier; Integrate a manufacturability knowledge rule base to assess and mark the casting feasibility and machining complexity of the proposed solutions. For the final scheme selected by experts, a high-fidelity gas-thermal-solid coupling analysis is performed using a driven simulation process to obtain the true performance F_true and evaluate the prediction accuracy of the surrogate model.

[0065] It should be noted that optimization algorithms (such as NSGA-III) directly search within the N-dimensional space defined by the fully parameterized master model. For each evaluation, M_unified is called to obtain the estimated aerodynamic efficiency, cooling efficiency, blade section stress, and blade root contact surface stress for that set of parameters. Optimization objectives can be set as Max(f1), Max(f2), Min(f3), Min(f4), etc.

[0066] Knowledge-enhanced initial population: Train a generative model (such as VAE) using historical data to generate new samples with similar well-designed historical samples and reasonable parameter combinations, and use them as the initial population to accelerate convergence.

[0067] Active learning strategy: When the optimization search encounters a sparse data region (where the surrogate model's prediction uncertainty is high), an intelligent agent can be invoked to insert 1-2 high-fidelity simulations. The results are used both for validation and to update the surrogate model in real time, ensuring the optimization direction remains reliable and forming an inner loop of "optimization-learning".

[0068] Step 4: Perform manufacturability filtering and visualization screening on the solutions in the Pareto optimal solution set to obtain the preferred solution, and then perform simulation verification on the preferred solution to generate a blade design data package; The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index.

[0069] In this embodiment of the disclosure, the process of performing manufacturability filtering and visualization screening on the solutions in the Pareto optimal solution set to obtain preferred solutions includes: The manufacturability knowledge rule base is invoked to evaluate the casting feasibility and machining complexity of each scheme in the Pareto optimal solution set, and the manufacturability risk level of each scheme is marked according to the evaluation results. It provides a multi-dimensional performance parallel coordinate graph, which visualizes the aerodynamic efficiency index, cooling efficiency index, structural strength index and manufacturability risk level of each scheme in the Pareto solution set in the same coordinate system. The system receives filtering instructions from experts via the parallel coordinate graph, filters out schemes with manufacturability risk levels exceeding a threshold from the Pareto solution set, and selects one or more preferred schemes.

[0070] The manufacturability knowledge rule base includes: minimum wall thickness casting limit rules, air film hole machining accessibility rules, and cooling channel complexity rating rules.

[0071] It should be noted that the output is a Pareto front that considers multiple performance trade-offs. Through visualization tools, designers can: This allows us to intuitively see "how much increased stress is required to improve efficiency by 0.1%".

[0072] By applying a manufacturability filter, solutions that may have too low a casting yield can be eliminated with a single click.

[0073] For the final 2-3 candidate solutions, high-fidelity verification will be initiated. The verification results will be fed back to the database to enhance the M_unified model and enable the system to self-evolve.

[0074] The detailed process of the collaborative optimization method for the conceptual design of gas turbine blades proposed in this embodiment is as follows: Figure 2 As shown, it will not be elaborated further here.

[0075] The collaborative optimization method for the conceptual design of gas turbine blades proposed in this embodiment breaks the traditional serial and fragmented design process, realizing a paradigm shift from "multi-round iteration and step-by-step optimization" to "integrated modeling and collaborative optimization", which greatly improves the innovation efficiency and global optimality of the conceptual design.

[0076] In summary, the collaborative optimization method for the conceptual design of gas turbine blades proposed in this embodiment significantly shortens the conceptual design cycle and achieves a global optimal balance of multidisciplinary performance.

[0077] Example 2 Figure 3 This is a structural diagram of a collaborative optimization system for the conceptual design of gas turbine blades according to an embodiment of this application, as shown below. Figure 3 As shown, the system includes: The first construction module 100 is used to construct a fully parameterized master model of the blade defined by the design parameter vector and a multi-disciplinary optimization target vector. The design parameter vector includes a subset of aerodynamic design parameters, a subset of cooling design parameters and a subset of structural design parameters, and there are preset parameter association and constraint rules between each subset. The multi-disciplinary optimization target vector includes aerodynamic efficiency index, cooling efficiency index and structural strength index. The parameter association and constraint rules include: The correlation rules between blade thickness distribution and cooling channel layout; The relationship between the leading and trailing edge diameter of the blades and the dimensions of the cooling structure; The correlation between the average temperature of the blade cross section and the creep strength; The correlation rules between blade profile pressure and blade section aerodynamic bending moment.

[0078] The subset of aerodynamic design parameters includes: blade inlet and outlet geometry angles and throat width; The subset of cooling design parameters includes: cooling channel layout and leading and trailing edge diameter; The subset of structural design parameters includes: blade wall thickness distribution.

[0079] The second construction module 200 is used to construct a multidisciplinary unified agent model with the design parameter vector as input and the predicted values ​​of each index in the multidisciplinary optimization target vector as output. Then, the multidisciplinary unified agent model is trained by an intelligent agent based on the large model. The intelligent agent is used to control the simulation platform to generate training and verification data, and decide whether to supplement the data and retrain based on the verification results. The iterative search module 300 is used to use the trained multidisciplinary unified surrogate model as the evaluator and the design space of the fully parameterized master model as the search space, and to call a multi-objective optimization algorithm to perform iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted value output by the multidisciplinary unified surrogate model and drives the iterative search to obtain a Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamic, cooling, and structural performance. It should be noted that the aforementioned multidisciplinary unified agent model is constructed using a multi-task learning neural network architecture; The multi-task learning neural network architecture includes: a shared low-level feature extraction network and multiple parallel subject-specific output branch networks; The shared low-level feature extraction network is used to extract shared geometric and physical features from the input design parameter vector; The subject-specific output branch networks are used to predict aerodynamic efficiency, cooling efficiency, and structural strength indices, respectively.

[0080] The multi-objective optimization algorithm is the NSGA-III algorithm.

[0081] The filtering module 400 is used to perform manufacturability filtering and visual filtering on the schemes in the Pareto optimal solution set to obtain the preferred scheme, and then perform simulation verification on the preferred scheme to generate a blade design data package. The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index.

[0082] In this embodiment of the disclosure, the second construction module 200 is further configured to: Step E1: Construct the agent based on the large language model. During the fine-tuning process, the agent learns the design rules, physical constraints and historical case knowledge of the turbine blade. Step E2: The intelligent agent control simulation platform generates an initial training dataset and performs initial training on the multidisciplinary unified agent model; Step E3: The intelligent agent control simulation platform generates a brand-new verification dataset to verify the trained multidisciplinary unified agent model; Step E4: When the verification accuracy is less than the preset verification accuracy threshold, the agent analyzes and predicts the error distribution, identifies the region where the error exceeds the threshold, calls high-fidelity simulation to generate a supplementary dataset in the region, and retrains the multidisciplinary unified agent model based on the expanded dataset. Step E5: Repeat steps E3 and E4 until the verification accuracy is greater than or equal to the preset verification accuracy threshold.

[0083] In this embodiment of the disclosure, the iterative search module 300 is further configured to: The initial population generation module of knowledge enhancement is invoked to generate an initial design scheme set, which is then used as the initial population for the optimization algorithm. The knowledge-enhanced initial population generation module includes a variational autoencoder trained on a database of historical excellent design cases; the variational autoencoder is used to learn the parameter distribution patterns of historical cases and generate new design schemes based on the learned distribution. When generating design schemes, the variational autoencoder controls the diversity of the generated schemes by performing Gaussian sampling in the latent space and setting temperature parameters.

[0084] Furthermore, the iterative search module 300 is also used for: The Monte Carlo Dropout method is used to quantify the prediction uncertainty of the multidisciplinary unified surrogate model for the current candidate design parameters; When the prediction uncertainty is greater than a preset prediction uncertainty threshold, the candidate design parameters are determined to be located in a sparse data region. The high-fidelity simulation was used to calculate the parameters of the candidate design and obtain the high-fidelity simulation results. The high-fidelity simulation results are used as real performance data to correct the predictions of the multidisciplinary unified agent model for the region. The corrected prediction or the high-fidelity simulation result is returned to the optimization algorithm to guide the subsequent search direction.

[0085] In this embodiment of the disclosure, the filtering module 400 is further configured to: The manufacturability knowledge rule base is invoked to evaluate the casting feasibility and machining complexity of each scheme in the Pareto optimal solution set, and the manufacturability risk level of each scheme is marked according to the evaluation results. It provides a multi-dimensional performance parallel coordinate graph, which visualizes the aerodynamic efficiency index, cooling efficiency index, structural strength index and manufacturability risk level of each scheme in the Pareto solution set in the same coordinate system. The system receives filtering instructions from experts via the parallel coordinate graph, filters out schemes with manufacturability risk levels exceeding a threshold from the Pareto solution set, and selects one or more preferred schemes.

[0086] The manufacturability knowledge rule base includes: minimum wall thickness casting limit rules, air film hole machining accessibility rules, and cooling channel complexity rating rules.

[0087] In summary, the collaborative optimization system proposed in this embodiment for the conceptual design of gas turbine blades significantly shortens the conceptual design cycle and achieves a global optimal balance of multidisciplinary performance.

[0088] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0089] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0090] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A collaborative optimization method for the conceptual design of gas turbine blades, characterized in that, Includes the following steps: A fully parameterized master model and a multidisciplinary optimization objective vector are constructed to define the blade through a design parameter vector. The design parameter vector includes a subset of aerodynamic design parameters, a subset of cooling design parameters, and a subset of structural design parameters. Each subset has preset parameter association and constraint rules. The multidisciplinary optimization objective vector includes aerodynamic efficiency index, cooling efficiency index, and structural strength index. The aerodynamic design parameter subset includes the blade inlet and outlet geometric angles and throat width. The subset of cooling design parameters includes: cooling channel layout and leading and trailing edge diameter; The subset of structural design parameters includes: blade wall thickness distribution; The parameter association and constraint rules include: The correlation rules between blade thickness distribution and cooling channel layout; The relationship between the leading and trailing edge diameter of the blades and the dimensions of the cooling structure; The correlation between the average temperature of the blade cross section and the creep strength; The correlation rules between blade profile pressure and blade section aerodynamic bending moment; The fully parameterized master model uses a unified design parameter vector of dimension N, X = (X1, X2, X3), where X1 is a subset of aerodynamic design parameters, X2 is a subset of cooling design parameters, and X3 is a subset of structural design parameters. The construction of the fully parameterized master model includes parameter decoupling and association definition, specifically including: Identify and define core driving parameters that simultaneously affect the performance of multiple disciplines; Establish association rules between parameters. When the core driving parameters change, adjust the subordinate parameters through the association rules to ensure geometric and physical rationality. A multidisciplinary unified agent model is constructed with the design parameter vector as input and the predicted values ​​of each index in the multidisciplinary optimization target vector as output. Then, the multidisciplinary unified agent model is trained in a supervised manner using an intelligent agent built based on a large model. The intelligent agent is used to control the simulation platform to generate training and verification data, and decide whether to supplement data and retrain based on the verification results. The multidisciplinary unified agent model is constructed using a multi-task learning neural network architecture. The multi-task learning neural network architecture includes: a shared low-level feature extraction network and multiple parallel subject-specific output branch networks; The shared low-level feature extraction network is used to extract shared geometric and physical features from the input design parameter vector; The subject-specific output branch networks are used to predict aerodynamic efficiency, cooling efficiency, and structural strength indices, respectively. The method of employing agent-supervised training based on a large model to train the multidisciplinary unified agent model includes: Step S1: Construct the agent based on the large language model. During the fine-tuning process, the agent learns the design rules, physical constraints and historical case knowledge of the turbine blade. Step S2: The intelligent agent control simulation platform generates an initial training dataset and performs initial training on the multidisciplinary unified agent model; Step S3: The intelligent agent control simulation platform generates a brand-new verification dataset to verify the trained multidisciplinary unified agent model; Step S4: When the verification accuracy is less than the preset verification accuracy threshold, the agent analyzes and predicts the error distribution, identifies the region where the error exceeds the threshold, calls high-fidelity simulation to generate a supplementary dataset in the region, and retrains the multidisciplinary unified agent model based on the expanded dataset. Step S5: Repeat steps S3 and S4 until the verification accuracy is greater than or equal to the preset verification accuracy threshold. Using the trained multidisciplinary unified surrogate model as the evaluator and the design space of the fully parameterized master model as the search space, a multi-objective optimization algorithm is invoked for iterative search. The optimization algorithm calculates the fitness of each candidate solution based on the predicted value output by the multidisciplinary unified surrogate model and drives the iterative search to obtain a Pareto optimal solution set that achieves Pareto global equilibrium in aerodynamic, cooling, and structural performance. The schemes in the Pareto optimal solution set are subjected to manufacturability filtering and visual screening to obtain the preferred schemes. Then, simulation verification is performed on the preferred schemes to generate a blade design data package. The process of manufacturability filtering and visual screening of the schemes in the Pareto optimal solution set to obtain the preferred schemes includes: The manufacturability knowledge rule base is invoked to evaluate the casting feasibility and machining complexity of each scheme in the Pareto optimal solution set, and the manufacturability risk level of each scheme is marked according to the evaluation results. It provides a multi-dimensional performance parallel coordinate graph, which visualizes the aerodynamic efficiency index, cooling efficiency index, structural strength index and manufacturability risk level of each scheme in the Pareto solution set in the same coordinate system. Receive the filtering instructions input by the experts through the parallel coordinate graph, filter out the schemes whose manufacturability risk level exceeds the threshold from the Pareto solution set, and select one or more preferred schemes; The manufacturability knowledge rule base includes: minimum wall thickness casting limit rules, air film hole machining accessibility rules, and cooling channel complexity rating rules; The blade design data package includes: a design parameter vector of the preferred scheme, actual performance data, and manufacturability evaluation results. The actual performance data includes: measured aerodynamic efficiency values ​​for verifying the aerodynamic efficiency index, blade cross-sectional temperature distribution and cooling air consumption for verifying the cooling efficiency index, and blade equivalent stress distribution for verifying the structural strength index.

2. The method as described in claim 1, characterized in that, Before invoking the multi-objective optimization algorithm for iterative search, the following steps are also included: The initial population generation module of knowledge enhancement is invoked to generate an initial design scheme set, which is then used as the initial population for the optimization algorithm. The knowledge-enhanced initial population generation module includes a variational autoencoder trained on a database of historical excellent design cases; the variational autoencoder is used to learn the parameter distribution patterns of historical cases and generate new design schemes based on the learned distribution. When generating design schemes, the variational autoencoder controls the diversity of the generated schemes by performing Gaussian sampling in the latent space and setting temperature parameters.

3. The method as described in claim 1, characterized in that, The multi-objective optimization algorithm is the NSGA-III algorithm.

4. The method as described in claim 1, characterized in that, The iterative search process also includes an active learning strategy, which includes: The Monte Carlo Dropout method is used to quantify the prediction uncertainty of the multidisciplinary unified surrogate model for the current candidate design parameters; When the prediction uncertainty is greater than a preset prediction uncertainty threshold, the candidate design parameters are determined to be located in a sparse data region. The high-fidelity simulation was used to calculate the parameters of the candidate design and obtain the high-fidelity simulation results. The high-fidelity simulation results are used as real performance data to correct the predictions of the multidisciplinary unified agent model for the region. The corrected prediction or the high-fidelity simulation result is returned to the optimization algorithm to guide the subsequent search direction.

Images