基于POD-BPNN模型的热启动策略及其在气动代理优化中的应用

Hot-start Strategy Based on POD-BPNN Model and Its Application in Aerodynamic Surrogate-based Optimization

传统气动优化设计需要大量CFD分析,而代理优化(SBO)方法能够有效降低CFD分析次数,但该方法并没有改变单次CFD分析时间。提出一种基于本征正交分解—反向传播神经网络(POD-BPNN)模型的热启动策略,并应用于气动代理优化。使用POD-BPNN模型对SBO中的初始样本建立从几何设计变量到流场数据的预测模型;在SBO迭代过程中,使用该模型预测新样本的流场,并将其作为CFD分析的初场,进行热启动计算,获得新样本的数据;添加新样本数据到POD-BPNN建模样本并更新模型,直到优化结束,通过案例对该策略进行对比验证。结果表明:在跨声速翼型减阻优化设计中,基于POD-BPNN的热启动策略使得单次CFD计算时间降低68%,SBO的效率整体提升37%。

The traditional aerodynamic optimization design requires computational fluid dynamics(CFD)analysis,and the surrogate-based optimization(SBO)method can effectively reduce the number of CFD analyse,but it cannot speed up a single CFD analysis time.A hot-start strategy using the proper orthogonal decomposition-back propagation based neural network(POD-BPNN)model is proposed,and applied in surrogate-based aerodynamic optimization.The prediction model from geometric design variables to flowfield data through the initial samples is built with SBO POD-BPNN model.During iteration of the SBO,the flowfield of a new sample is predicted by the built model,and the flowfield is used as the initial flowfield of the hot-start CFD analysis for the new sample.The new sample is used to update the POD-BPNN model until the end of the optimization.The comparison verification of proposed strategy is performed with instance.The results show that,in the aerodynamic optimization design of the transonic airfoil,the hot-start strategy based on POD-BPNN model can reduce the time of a single CFD analysis by 68%,and improve the efficiency of the SBO by 37%.