Data-driven non-intrusive shape-topology optimization framework for curved shells

Aiming at obtaining an optimal configuration of curved shells in the compact design space, it is an effective way to perform the shape and topology optimization simultaneously. Considering the complexity in solving the coupled shape and topology optimization problem, this paper presents an easy-to-implement data-driven non-intrusive shape-topology optimization framework for curved shells, including two steps. In the off-line step, the shape equation-driven mesh deformation method is firstly proposed, contributing to using a small number of design variables for fast shape change and optimization of the skin of curved shell without re-modeling and re-meshing. Latin hypercube sampling is used to generate sample points as the input data in the design space, and the corresponding skin shape is obtained. In each design domain determined by different skin shapes, the topology optimizations are performed in parallel or by multiple optimization solvers, after that above multi-source optimization results are collected into the output dataset. The mapping relationship is trained between the input data and output data via surrogate modeling technique, and this non-intrusive concept contributes to integrating the independent shape and topology optimization into a single-loop data-driven optimization. The cross validation-Voronoi adaptive model updating method is used to improve the prediction accuracy. In the on-line step, the single-loop data-driven optimization is carried out using covariance matrix adaptive evolution strategy and adaptive optimization updating method, which can obtain optimized results more efficiently than direct topology optimization or shape optimization method. In order to verify the effectiveness of the proposed data-driven non-intrusive shape-topology optimization framework, three optimization examples are carried out. Example results show that compared with the results by the single topology optimization, the strain energy results by the proposed framework can be reduced by 20.55%, 15% and 51.6%, respectively, highlighting the outstanding optimization ability of the proposed framework.