High-speed aircraft aerodynamic configuration agile optimization method based on artificial intelligence model
By combining a multi-gated neural network and a group algorithm based on a teaching-learning strategy with an intelligent autonomous decision-making model, the high cost and low efficiency of traditional aerodynamic optimization methods are solved, achieving high-precision and high-efficiency aerodynamic shape design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING LINJIN SPACE AIRCRAFT SYST ENG INST
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional aerodynamic optimization methods are computationally expensive and have long design cycles. Existing surrogate models lack accuracy, and artificial intelligence models have biased prediction results, making it difficult to meet the requirements for high-precision aerodynamic characteristics and rapid research and development.
By employing a neural network modeling technique with multi-gating mechanisms and a swarm algorithm based on a teaching-learning strategy, combined with an intelligent autonomous decision-making model, and constructing a multi-objective adaptive weight allocation model, we achieve high efficiency, robustness, and adaptability in aerodynamic optimization, forming a closed-loop optimization system.
It significantly improves aerodynamic optimization efficiency, reduces computational costs and design cycle, enhances aerodynamic optimization accuracy and stability, avoids local optimum traps, and achieves efficient aerodynamic shape design.
Smart Images
Figure CN122046544A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent algorithm technology and is used for the optimization design of aircraft aerodynamic layout. Background Technology
[0002] Aerodynamic optimization design is one of the core technologies for improving flight performance, reducing energy consumption, and ensuring flight safety. However, the nonlinear characteristics of complex flows, strong coupling effects, and the interaction of multi-dimensional design variables make aerodynamic optimization problems highly complex and challenging. The main problems currently existing in aerodynamic optimization are as follows:
[0003] 1) Traditional optimization methods suffer from high computational costs and long design cycles, making it difficult to meet the needs of rapid aircraft development. Traditional optimization algorithms often require a large amount of computational resources and time to complete the optimization process, and are difficult to escape the limitations of local optima.
[0004] 2) Existing surrogate models are insufficient in terms of accuracy and generalization ability, making it difficult to meet the needs of high-precision aerodynamic characteristic prediction. Traditional surrogate models (such as multinomial regression or response surface models) often fail to accurately capture nonlinear relationships when faced with complex aerodynamic characteristics, resulting in significant deviations between predicted and actual values. This lack of accuracy severely limits the application of surrogate models in practical engineering, especially in the field of aerospace vehicle design where high-precision optimization is required.
[0005] 3) The application of artificial intelligence technology carries potential technical risks. Model predictions often deviate significantly from actual physical properties. This deviation may stem from model uncertainties, computational errors, or nonlinear relationships between design variables. If this AI model "illusion" problem cannot be effectively avoided, the optimization process may produce incorrect design decisions, thereby reducing optimization effectiveness or even leading to design failure. Summary of the Invention
[0006] (I) Technical problems to be solved To address the problems of difficulty in convergence and low efficiency in aerodynamic layout optimization, this invention aims to integrate artificial intelligence modeling and dynamic autonomous decision-making into the optimization process, solve the problem of the imbalance between stability and efficiency in aerodynamic optimization, and achieve improved aerodynamic optimization efficiency under intelligent empowerment.
[0007] (II) Technical Solution An agile optimization method for the aerodynamic shape of a high-speed aircraft based on an artificial intelligence model, comprising the following steps: A1 establishes an aerodynamic proxy model: by introducing a multi-gating mechanism to dynamically coordinate the outputs of multiple expert models, in order to solve the multimodal feature learning and multi-objective optimization problems in complex tasks; A2 Global Optimization Algorithm: A group algorithm based on a teaching-learning strategy, which finds the optimal solution to a problem by simulating the teaching process of teachers to students and the mutual learning process among students. The algorithm regards the optimization problem as a learning process, in which potential solutions are regarded as students who need to learn knowledge from teachers to continuously improve. A3 establishes an intelligent autonomous decision-making model; by constructing a multi-objective adaptive weight allocation model, it realizes the dynamic collaborative optimization of the weights of key elements such as the confidence of the proxy model, the constraint penalty function, and the objective function, solves the mathematical modeling problem of dynamic matching of weight parameters and optimization objectives during the iteration process, and develops a closed-loop optimization system with autonomous perception-decision-execution capabilities, thereby improving the convergence speed, avoiding local optima, and ensuring the robustness of the global optimal solution in complex engineering optimization scenarios. A4 aerodynamic optimization iteration: Based on a surrogate model that predicts indicators such as lift-to-drag ratio, an artificial intelligence optimization algorithm is used to globally optimize the design parameters, and the results are verified by CFD. After verification, the CFD calculation results are put into the training set of the model to retrain the model and continue to optimize until the optimized design converges.
[0008] Furthermore, the aerodynamic proxy model described in A1 includes multiple expert networks and a gating network. Each expert is responsible for feature extraction of a specific sub-task, and the gating network is used to assign weights to the outputs of different experts. This effectively enables high-precision modeling of multiple aerodynamic indicators simultaneously. Compared with modeling each component separately, this neural network can maintain or even improve the accuracy of aerodynamic prediction.
[0009] Furthermore, the global optimization algorithm described in A2 is divided into two stages: the teacher teaching stage and the student learning stage. In the teacher learning stage, the algorithm selects the optimal solution in the current population as the teacher, and other students learn knowledge by moving towards the teacher's position. The moving distance is adjusted by the difference between the teacher and the students.
[0010] Furthermore, A3 describes an intelligent decision-making model that integrates agent model confidence assessment, constraint penalty function feedback, and multi-objective trade-off mechanism. This model enables autonomous and dynamic adjustment of weight parameters during the optimization process, forming an intelligent optimization theoretical framework that is efficient, robust, and adaptive. This provides new methodological support for solving high-dimensional, multi-constraint, and nonlinear challenges in complex engineering optimization problems.
[0011] Furthermore, the A4 iteration process is divided into three parts: "preliminary preparation", "internal iteration", and "external iteration". First, a large sample training library with different shapes is calculated through CFD numerical simulation. In the "internal iteration", a high-precision aerodynamic surrogate model is established based on the sample library. Based on the surrogate model, the current optimal design is obtained through the particle swarm optimization method. In the "external iteration", the optimal design parameters are calculated and verified by CFD, and the calculation results are added to the sample library. The "internal iteration" is repeated until the optimal design parameters converge.
[0012] (III) Beneficial Technical Effects This invention successfully applies artificial intelligence technology to the optimization design of various engineering problems. By improving the AI modeling method and intelligent decision-making method, it realizes the application of AI technology in aerodynamic shape design, solves the problem of the "last mile" of AI technology implementation, and greatly improves the efficiency of aerodynamic optimization compared with traditional optimization methods. Attached Figure Description
[0013] Figure 1 Flowchart of Agile Optimization of Aerodynamic Shape for High-Speed Aircraft Based on Artificial Intelligence Model Figure 2 MMOE neural network model architecture; Figure 3 A global optimization algorithm framework for "teaching" and "learning"; Figure 4 Artificial intelligence optimization process based on dynamic modeling of agent model; Figure 5 Intelligent pneumatic optimization platform; Figure 6 : Renderings showing the optimization process of shape A. Detailed Implementation
[0014] Besides the embodiments described below, the present invention may also have other embodiments or be implemented in different ways. Therefore, it should be understood that the present invention is not limited to the details of the results described in the following specification or shown in the accompanying drawings. When only one embodiment is described herein, the claims are not limited to that embodiment.
[0015] like Figure 1 As shown, an agile optimization method for the aerodynamic shape of high-speed aircraft based on an artificial intelligence model is proposed. This method includes aerodynamic surrogate model modeling, efficient global optimization algorithm, and establishment of intelligent autonomous decision-making model. By integrating the above technologies, an efficient aerodynamic optimization integrated platform with fully automated parametric modeling, surrogate model training, and autonomous optimization decision-making processes has been developed. Through the combination of dynamic surrogate model modeling and artificial intelligence optimization algorithm, a closed-loop optimization process from global optimization of design parameters to CFD verification has been realized.
[0016] 1) Aerodynamic surrogate modeling To address the problem of pneumatic surrogate modeling, a surrogate modeling technique based on deep neural networks has been developed. For example... Figure 2 As shown, the Multi-Gate Mixture of Experts (MMoE) is an improved architecture based on the Mixture of Experts (MoE) model. It dynamically coordinates the outputs of multiple expert models by introducing a multi-gating mechanism to solve multimodal feature learning and multi-objective optimization problems in complex tasks. Its core structure includes multiple expert networks (each expert responsible for feature extraction of a specific sub-task) and a gating network (used to assign weights to the outputs of different experts). Using MMoE, high-precision modeling of multiple aerodynamic parameters can be effectively achieved simultaneously. Compared to modeling each component individually, this neural network can maintain or even improve the accuracy of aerodynamic predictions.
[0017] In MMOE, the system architecture comprises a gating network and multiple parallel expert networks. The gating network receives input data and generates a set of weight distributions to dynamically and selectively activate one or more of the most relevant expert networks. Each expert network, as an independent sub-network, focuses on processing a specific pattern or domain sub-task in the input data space. The output of the gating network is integrated with the outputs of the selected expert networks through a weighted averaging mechanism to form the final system output. This design allows the model to maintain a large number of parameters while significantly reducing computational overhead through a sparse activation mechanism, and effectively improving the prediction accuracy and generalization ability of the surrogate model. There are significant coupling relationships among the six components of aerodynamic forces of high-speed aircraft, and these six components can be predicted using the MMOE model.
[0018] Comparison of prediction accuracy between MMOE and traditional neural networks
[0019] 1400 shapes were sampled using optimal Latin hypercube sampling, with five flight states calculated for each shape, resulting in 7000 parameter combinations. CFD calculations were used to obtain six-component aerodynamic characteristic parameters, and an aerodynamic database containing 7000 samples was established for model training and testing. The training set had 5000 samples, and the test set had 2000 samples. Table 1 shows a comparison of the prediction accuracy of MMOE and traditional neural networks for the test set. MMOE significantly improved the prediction accuracy for aerodynamic forces of unknown shapes, especially for yaw moment, reducing the error by more than half. Furthermore, MMOE only requires 500 training steps and 1000 seconds to train the neural network. Compared to the Kriging model, MMOE has a similar prediction error but a significantly shorter training time. The Kriging model takes more than an hour to train, while the MMOE model only takes less than twenty minutes, resulting in a significant improvement in efficiency.
[0020] The surrogate modeling technique reduces the CFD computation requirements from thousands to less than half while maintaining accuracy. Furthermore, sensitivity analysis of various design variables was conducted based on the surrogate model, successfully extracting key design variables and providing valuable insights for designers to further improve flight performance.
[0021] 2) High-efficiency intelligent global optimization algorithm Based on the MMOE deep learning agent model, an intelligent decision-making particle swarm optimization algorithm—Teaching-Learning-Based Optimization (TLBO)—is applied. It is a swarm algorithm based on a teaching-learning strategy, attracting attention due to its simplicity and the fact that it does not require adjusting too many parameters. The core idea of the TLBO algorithm is to find the optimal solution to the problem by simulating the teaching process between teachers and students and the mutual learning process among students. Compared with other swarm algorithms, such as genetic algorithms or particle swarm optimization, TLBO does not require setting additional parameters such as crossover rate, mutation rate, or learning factor. The algorithm treats the optimization problem as a learning process, where potential solutions are seen as students who need to learn knowledge from the teacher (i.e., the current optimal solution) to continuously improve. The algorithm framework is as follows: Figure 3 As shown.
[0022] The TLBO algorithm mainly consists of two phases: the teacher instruction phase and the student learning phase. In the teacher learning phase, the algorithm selects the best solution in the current population as the teacher, and other students learn knowledge by moving towards the teacher's position. The moving distance is adjusted by the difference between the teacher and the students.
[0023] 3) Intelligent autonomous decision-making model In the field of AI-driven intelligent decision-making optimization algorithms, this research breaks through the limitations of traditional fixed multi-objective weight mechanisms in optimization, developing a dynamic optimization algorithm framework based on intelligent decision-making technology. By constructing a multi-objective adaptive weight allocation model, it achieves dynamic collaborative optimization of the weights of key elements such as surrogate model confidence, constraint penalty function, and objective function. The research focuses on solving the mathematical modeling problem of dynamically matching weight parameters with optimization objectives during iteration, developing a closed-loop optimization system with autonomous perception-decision-execution capabilities. This improves convergence speed, avoids local optima, and ensures the robustness of the global optimum in complex engineering optimization scenarios. By constructing an intelligent decision-making model that integrates surrogate model confidence evaluation, constraint penalty function feedback, and multi-objective trade-off mechanisms, it enables autonomous dynamic adjustment of weight parameters during optimization, forming an intelligent optimization theoretical framework that combines efficiency, robustness, and adaptability. This provides new methodological support for solving the challenges of high-dimensionality, multi-constraint, and nonlinearity in complex engineering optimization problems.
[0024] In the intelligent autonomous decision-making algorithm, a novel concept of "conservative factor" was creatively invented and introduced into the dynamic decision-making model to achieve real-time online feedback adjustment of the conservative factor α. A bimodal control strategy was designed: when α>0, the optimization convergence ability is enhanced, and when α<0, global optimization is activated; the conservative factor is integrated into the improved particle swarm optimization algorithm, which improves the population diversity index by 28% and reduces the number of convergence iterations by 40%.
[0025] The objective function of the original aerodynamic optimization problem is as follows:
[0026] The currently developed method improves the objective function as follows:
[0027] in To comprehensively consider the loss function under constraints, The original objective function is... This is the penalty factor parameter, used to adjust the intensity of the penalty for constraint violation; To constrain the calculation of violation items; It is a penalty function term that integrates prediction uncertainty. Its function construction comprehensively considers multiple dimensions of information, including model prediction variance, constraint violation degree, and objective function value. Conservative factor. The sign and magnitude of the value determine the orientation of the optimization strategy: when When the value is positive, the penalty for constraint violation increases, prompting the algorithm to perform a refined search within the current high-confidence region, exhibiting local optimization characteristics; when When negative values are taken, the penalty mechanism transforms into an exploration incentive, guiding the algorithm to expand into potential optimization regions outside the constraint boundaries, thus achieving a global search function. In the actual optimization process, a model is constructed based on characteristic parameters such as the optimization iteration algebra number and the rate of change of the mean variance of the model prediction. The adaptive update strategy dynamically adjusts parameter values as the confidence of the surrogate model increases, achieving a synergistic improvement in optimization efficiency and quality finding while ensuring constraint satisfaction.
[0028] 4) Aerodynamic optimization iteration like Figure 4 As shown, the above technologies are integrated into the software environment to develop an intelligent optimization platform that supports full-process automation of parametric modeling, surrogate model training, and autonomous optimization decision-making. An AI optimization method based on dynamic modeling using a surrogate model is employed. The AI optimization process based on dynamic modeling using a surrogate model involves: using a surrogate model that predicts indicators such as lift-to-drag ratio, using AI optimization algorithms to globally optimize design parameters, and verifying the results through CFD. After verification, the CFD calculation results are fed into the model's training set to retrain the model and continue optimization until the optimized design converges.
[0029] like Figure 5 As shown, a platform was established based on an iterative process divided into three parts: "preliminary preparation," "internal iteration," and "external iteration." First, a large-sample training database of different shapes was calculated using CFD numerical simulation. In the "internal iteration," a high-precision aerodynamic surrogate model was built based on the database, and the current optimal design was obtained using a particle swarm optimization method. In the "external iteration," the optimal design parameters were calculated and verified using CFD, and the results were added to the database. The "internal iteration" was then repeated until the optimal design parameters converged. The entire platform includes an expert intervention layer, an adaptive control layer, and an automatic optimization layer, covering key aspects such as design variable initialization, surrogate model construction, optimization algorithm execution, constraint evaluation, and optimization result output. First, the range of design variables is determined based on engineering requirements. An initial sample point set is generated using experimental design methods such as Latin hypercube sampling. Aerodynamic performance data corresponding to the sample points is obtained based on high-precision CFD numerical simulation. Then, a Bayesian neural network surrogate model is constructed using this dataset. This model can not only achieve efficient prediction of aerodynamic performance, but also quantify the prediction confidence of the model by outputting the prediction variance. During the optimization iteration process, the prediction uncertainty is used as a penalty function to guide aerodynamic optimization. The TLBO global optimization algorithm is used to search for the optimal solution in the design space that meets multiple constraints. By introducing a "conservative factor" to dynamically adjust the weight coefficient of the penalty function, an adaptive balance between global exploration and local optimization is achieved, and finally, the optimal aerodynamic shape scheme that meets the engineering design requirements is output.
[0030] After the initial aerodynamic shape is generated, a deep learning model is used as the aerodynamic predictor surrogate model, which is then connected to the aerodynamic optimization module. An AI-based intelligent decision-making optimization method is employed to complete the short-cycle agile optimization of the aircraft's aerodynamic shape, achieving a generalized integrated intelligent design for aircraft aerodynamic shape generation and optimization. Further comparative analysis of the performance of the developed optimization platform and traditional optimization methods during the optimization processes of shape A and shape B is conducted to verify the practicality and efficiency of the framework in engineering applications. Given that engineering aerodynamic optimization relies on computationally expensive high-precision numerical simulations, this study focuses on comparing the practical value of the new method with traditional methods in reducing computational costs and improving optimization efficiency.
[0031] Figure 6 The optimization process of shape A is shown. Shape A was modeled parametrically using 22 design parameters. The optimization goal was to improve the lift-to-drag ratio and improve the stability of the focal position. Figure 5 (Optimization Process A) illustrates the changes in lift-to-drag ratio and focal position with the number of iterations for the intelligent global optimization framework and the traditional local optimization framework. Based on the data in the table, the intelligent global optimization framework achieved a 2.5% improvement in lift-to-drag ratio after only 95 iterations, with the focal position remaining stable within the engineering tolerance range; while the traditional local optimization algorithm required 420 iterations, with only a 2% improvement in lift-to-drag ratio. The iteration curves show that the intelligent global optimization framework exhibits a clear and stable optimization direction from the initial iteration stage, with a continuously improving lift-to-drag ratio and a rapidly converging focal position to the target range; conversely, the traditional local optimization algorithm experiences multiple fluctuations in its optimization process, lacks directional stability, and leads to a significant increase in the number of iterations.
[0032] Table 2 further confirms the superiority of the intelligent global optimization framework. For shape B, the intelligent optimization platform only requires 83 iterations to achieve an improvement of 0.045 in the aerodynamic optimization index; the traditional local optimization algorithm requires 240 iterations to achieve the same improvement of 0.04. Although the final optimization indices of the two are similar, the intelligent global optimization framework requires only one-third the number of iterations of the traditional method, significantly reducing the aerodynamic optimization cycle.
[0033] Table 2 Aerodynamic Optimization Process of Engineering Shape
[0034] Based on the optimization results of both shape A and shape B, the intelligent global optimization framework demonstrates three core advantages in engineering aerodynamic optimization: it inherently supports more efficient parallel computing, accelerates the optimization process through multi-sample parallel evaluation, and overcomes the dependence of traditional local optimization on serial computing; the framework has low dependence on the initial shape quality, and can still find an effective optimization path through global exploration even if the initial design performance is poor; and the efficiency of single optimization operation is significantly improved, with the optimization time shortened by more than three times compared to traditional local optimization algorithms, which can greatly reduce the engineering research and development cycle and cost.
[0035] 1) AI proxy modeling method incorporating aerodynamic principles Based on the knowledge of aerodynamic laws of near-space vehicles, an innovative AI model architecture suitable for aerodynamic optimization problems is proposed, forming a high-precision aerodynamic modeling based on multi-gated multi-expert neural networks. This reduces the aerodynamic prediction error for different parameterized shapes by more than half, further improving the efficiency of aerodynamic optimization.
[0036] 2) High-performance intelligent global optimization algorithm for teaching and learning Develop and apply the "Teaching and Learning" TLBO intelligent global optimization algorithm. By mimicking human "teacher-led instruction and student-interaction," it forms a cluster collaborative optimization mechanism for high-dimensional aerodynamic design space. Compared with traditional gradient optimization algorithms, it can effectively avoid the "trap" of local optima in the optimization process, achieving breakthroughs in aircraft performance indicators. 3) Intelligent online autonomous decision-making algorithm oriented towards optimization objectives.
[0037] This innovative approach introduces the concept of "conservative factors" for the first time, effectively overcoming the "illusion" problem often encountered by AI models and initially forming an "intelligent agent" for optimization decision-making. Through the negative feedback control mechanism of "conservative factors," AI can adjust the robustness of the design in real time during aerodynamic optimization, effectively mitigating the technical risks of artificial intelligence technology in application and significantly improving optimization efficiency.
[0038] 4) Intelligent aerodynamic shape optimization platform integrating multiple advanced algorithms An intelligent aerodynamic shape optimization platform has been developed, which effectively integrates various technologies and improves aerodynamic optimization efficiency by an order of magnitude compared to traditional optimization methods.
[0039] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A method for agile optimization of the aerodynamic shape of a high-speed aircraft based on an artificial intelligence model, characterized in that, The steps include the following: A1 establishes an aerodynamic proxy model: by introducing a multi-gating mechanism to dynamically coordinate the outputs of multiple expert models, in order to solve the multimodal feature learning and multi-objective optimization problems in complex tasks; A2 Global Optimization Algorithm: A group algorithm based on a teaching-learning strategy, which finds the optimal solution to a problem by simulating the teaching process of teachers to students and the mutual learning process among students. The algorithm regards the optimization problem as a learning process, in which potential solutions are regarded as students who need to learn knowledge from teachers to continuously improve. A3 establishes an intelligent autonomous decision-making model; by constructing a multi-objective adaptive weight allocation model, it realizes the dynamic collaborative optimization of the weights of key elements such as the confidence of the proxy model, the constraint penalty function, and the objective function, solves the mathematical modeling problem of dynamic matching of weight parameters and optimization objectives during the iteration process, and develops a closed-loop optimization system with autonomous perception-decision-execution capabilities, thereby improving the convergence speed, avoiding local optima, and ensuring the robustness of the global optimal solution in complex engineering optimization scenarios. A4 aerodynamic optimization iteration: Based on a surrogate model that predicts indicators such as lift-to-drag ratio, an artificial intelligence optimization algorithm is used to globally optimize the design parameters, and the results are verified by CFD. After verification, the CFD calculation results are put into the training set of the model to retrain the model and continue to optimize until the optimized design converges.
2. The method as described in claim 1, characterized in that, The aerodynamic proxy model described in A1 includes multiple expert networks and a gating network. Each expert is responsible for feature extraction of a specific sub-task, and the gating network is used to assign weights to the outputs of different experts. This effectively enables high-precision modeling of multiple aerodynamic indicators simultaneously. Compared with modeling each component separately, this neural network can maintain or even improve the accuracy of aerodynamic prediction.
3. The method as described in claim 1, characterized in that, The global optimization algorithm described in A2 consists of two phases: the teacher teaching phase and the student learning phase. In the teacher learning phase, the algorithm selects the optimal solution in the current population as the teacher, and other students learn knowledge by moving towards the teacher's position. The moving distance is adjusted by the difference between the teacher and the students.
4. The method as described in claim 1, characterized in that, As described in A3, by constructing an intelligent decision-making model that integrates surrogate model confidence assessment, constraint penalty function feedback, and multi-objective trade-off mechanism, the weight parameters are autonomously and dynamically adjusted during the optimization process. This forms an intelligent optimization theoretical framework that is efficient, robust, and adaptive, providing new methodological support for solving high-dimensional, multi-constraint, and nonlinear challenges in complex engineering optimization problems.
5. The method as described in claim 1, characterized in that, The A4 iteration process is divided into three parts: "preliminary preparation", "internal iteration", and "external iteration". First, a large sample training library with different shapes is calculated through CFD numerical simulation. In the "internal iteration", a high-precision aerodynamic surrogate model is established based on the sample library. Based on the surrogate model, the current optimal design is obtained through particle swarm optimization. In the "external iteration", the optimal design parameters are calculated and verified by CFD, and the calculation results are added to the sample library. The "internal iteration" is repeated until the optimal design parameters converge.