Transfer learning based variable-fidelity surrogate model for shell buckling prediction

Abstract

This paper provides a novel establishment method of the variable-fidelity surrogate model (VFSM) driven by transfer learning for shell buckling prediction problems such as buckling prediction of variable-stiffness composite shells, aiming at achieving higher prediction accuracy than the traditional VFSM constructed by the bridge function. The transfer learning based variable-fidelity surrogate model (TL-VFSM) includes two steps. In the first step, low-fidelity model and high-fidelity model points are calculated and collected into the source dataset and target dataset for transfer learning, respectively. Based on the source dataset, a Deep Neural Network (DNN) model is initialized. Then, the hyperparameter tuning of the DNN model is carried out based on Bayesian Optimization (BO), which uses the target dataset as the validation set. After the hyperparameter pre-tuning, the pre-trained DNN model is obtained. In the second step, early layers of the pre-trained DNN model are remained, while the last layer is replaced. Then, another hyperparameter tuning is carried out to fine-tune the pre-trained DNN model using BO. After that, the transfer learning is finished and the re-trained DNN model is obtained. Through above two steps, the TL-VFSM is constructed. The effectiveness and efficiency of the proposed TL-VFSM are verified by three test functions and two engineering examples of variable-stiffness composite shells and hierarchical stiffened shells. Example results indicate that, using the same computational cost, the proposed TL-VFSM achieves higher prediction accuracy than traditional VFSMs constructed by the bridge function, and it can be a potential method for time-consuming buckling prediction problems with multi-source data or variable fidelities.