利用遗传算法和神经网络响应面来实现复合材料结构优化设计

运用正交试验设计选择设计样本，建立神经网络响应面，以代替复合材料结构优化中的大量的有限元分析；将神经网络响应面作为目标函数或者约束条件，加上其他常规约束条件进行优化模型的建立，再应用遗传算法（GA）进行优化，这可以实现设计分析与设计优化的分离。以复合材料帽型加筋板的重量优化问题为例，建立了重量响应面目标函数、强度和翘曲稳定性响应面约束条件；并通过NASTRAN进行有限元计算，以获取用于响应面训练的样本点数据。研究表明，该方法能以较少的结构分析次数，取得高精度的响应面近似模型，从而使优化效率大为提高。神经网络响应面能够获得与传统响应面同等，甚至更好的精度。

Composite Structural Optimization by Genetic Algorithm and Neural Network Response Surface Modeling

Neural-Network Response Surfaces (NNRS) is applied to replace the actual expensive finite element analysis during the composite structural optimization process. The Orthotropic Experiment Method (OEM) is used to select the most appropriate design samples for network training. The trained response surfaces can either be objective function or constraint conditions. Together with other conventional constraints, an optimization model is then set up and can be solved by Genetic Algorithm (GA). This allows the separation between design analysis modeling and optimization searching. Through an example of a hat-stiffened composite plate design, the weight response surface is constructed to be objective function, and strength and buckling response surfaces as constraints; and all of them are trained through NASTRAN finite element analysis. The results of optimization study illustrate that the cycles of structural analysis can be remarkably reduced or even eliminated during the optimization, thus greatly raising the efficiency of optimization process. It also observed that NNRS approximation can achieve equal or even better accuracy than conventional functional response surfaces.