Variable-fidelity Transfer Learning Model for Efficient Buckling Analysis of Variable Stiffness Composite Cylindrical Shells

LI Zengcong,TIAN Kuo,HUANG Lei,WANG Bo

DOI: https://doi.org/10.13801/j.cnki.fhclxb.20210604.001

2021-01-01

Abstract:Compared with the traditional design method of composite cylindrical shells with straight fiber laminate, variable stiffness composite cylindrical shells can greatly increase the design space of composite material and thus achieve higher buckling loads by means of the curved fiber laminate. To describe the curved fiber path precisely, it is necessary to establish high-fidelity FE models for variable stiffness composite cylindrical shells. Therefore, it brings great challenges to the efficiency of buckling analysis and optimization of variable stiffness composite cylindrical shells. In this paper, a variable-fidelity transfer learning model was proposed for the fast prediction of linear buckling load and post-buckling load of variable stiffness composite cylindrical shells. Firstly, the appropriate high-fidelity model and low-fidelity model of variable stiffness composite cylindrical shells were constructed. Then, the deep neural network was established and trained with a large number of low-fidelity samples as the source dataset, and the pre-trained model was obtained. Finally, the last layer was retained by fine-tuning with a small number of high-fidelity samples as the target dataset, and the variable-fidelity transfer learning model was constructed after the re-training on the pre-trained model. The example results of linear buckling and post-buckling load prediction of variable stiffness composite cylindrical shells indicate that, the computational cost of variable-fidelity transfer learning model can reduce by 47.7% and 62.3% than surrogates built by the high-fidelity samples directly when achieving similar prediction accuracy, showing the advantage of high prediction efficiency of the proposed method. Besides, compared with the variable-fidelity surrogate models built by the bridge function and Co-Kriging, the proposed method shows the best prediction accuracy with different combinations of high-fidelity and low-fidelity samples, which demonstrates the advantage of high prediction accuracy of the proposed method.