借助相关机器学习技术的气动外形优化设计新方法

Aerodynamic shape optimization is the critical technology of aircraft design, which has been attracted much attention of relevant researchers. At present, global aerodynamic optimization methods mainly include direct optimization and surrogated-based optimization. Among them, the shortcomings of direct method is that their low optimization efficiency and poor optimization effect for high-dimensional design. Although surrogated-based optimization methods can greatly improve the optimization efficiency, it is difficult to metamodel and optimize for high-dimensional aerodynamic design..The project will make full use of the data generated in the optimization process by machine learning technologies to further improve the efficiency and enhance the design effect of the existing aerodynamic optimization design methods. The feasibility and superiority of these methods are verified by standard aerodynamic optimization examples. This project intends to achieve the deep integration of machine learning technology and aerodynamic shape design; and builds a software platform for aerodynamic design to provide reliable algorithm guarantee and technical support for aircraft shape design in China. The research contents are as follows:.(1) With the help of supervised learning, data mining and other machine learning technologies, an improved data-driven evolutionary algorithm is proposed to improve the optimization efficiency of direct methods and enhance the effect of high-dimensional design..(2) With the help of data mining technology, variable correlation is analyzed, hierarchical modeling and optimization strategy is proposed, then, the surrogate model method based on hierarchical strategy is developed reduce the difficulty of high-dimensional aerodynamic design.. (3) With the help of manifold learning technology, a space transformation method suitable for aerodynamic design is developed to reduce the difficulty of high-dimensional design from the angle of dimensionality reduction.

全局气动优化设计是相关科研人员的研究热点和追求目标。目前，全局气动优化方法主要有基于进化算法的直接优化和代理模型优化。其中，直接方法存在优化效率低、高维设计效果差等问题。代理模型方法虽然可大幅提升优化效率，但面对复杂高维设计，存在建模难、优化难等问题。.针对上述问题，本项目拟开展借助机器学习相关技术，从数据驱动的角度，发展气动优化设计新方法，实现机器学习技术与气动外形设计的深度融合，为我国飞行器外形设计提供可靠的算法保证和技术支持。研究内容如下：.（1）借助监督式学习、数据挖掘等机器学习技术，提出数据驱动的进化算法改进方法，提升直接方法的优化效率、并增强高维设计效果。.（2）借助数据挖掘技术分析变量相关性，提出分层建模与优化策略，发展基于分层策略的代理模型方法，实现高维建模与优化分解，降低高维设计难度。.（3）借助流形学习技术，发展适用于气动设计的空间降维变换方法，降低高维设计难度。

全局气动优化设计对提高飞机气动性能具有重要意义。目前，全局气动优化方法主要有基于进化算法的直接优化和代理模型优化。直接方法存在优化效率低、高维设计效果差等问题；代理模型方法虽然可大幅提升优化效率，但面对复杂高维设计，存在建模难、优化难等问题。为解决上述问题，本项目借助机器学习相关技术，从数据驱动的角度，发展气动优化设计新方法，实现机器学习技术与复杂高维气动外形设计深度融合。本项目开展的具体研究内容如下：.（1）利用机器学习技术改进启发式进化算法，融合数据驱动在线学习策略，深度挖掘历史气动数据信息，以指导和驱动进化算法解决高维昂贵工程问题的全局优化能力。改进后的进化算法应用于NACA0012翼型气动外形优化设计，大幅减少了进化算法调用CFD的评估次数，直接优化效率提高了5倍以上。.（2）借助数据挖掘技术分析变量相关性，提出了一种基于数据驱动的代理模型援助支持策略，通过构建历史气动数据代理模型并结合代理模型最优预测方法，辅助驱动优化算法进行气动外形优化设计。该研究成功应用于RAE2822翼型和机翼气动外形优化设计，气动性能分别提升40%和37%以上，优化效果显著，并克服了传统的基于代理模型方法过于依赖代理模型精度的问题。.（3）融合流形学习技术，发展了一种适用于气动设计的几何域分解技术，解决复杂高维气动设计建模难、优化难问题。该方法应用于高维NACA0012机翼阻力最小化设计，开发了一套独立于优化框架且适合于高维气动外形设计的气动参数化模块和几何域分解模块，显著降低了高维气动外形优化和建模难度，优化效果相较于传统的直接优化方法(DO)和基于代理模型的优化方法(SBO)分别提高20.8%和17.8%。.本项目实现基于机器学习技术的气动外形优化设计，显著提升气动优化效率和效果，解决了复杂非线性气动外形优化建模难、优化难等问题，具有重要的理论及工程应用价值。