Multi-Fidelity deep learning for predicting the nonlinear buckling behaviour of concrete thin-shells

This research introduces a novel approach to rapidly estimate the nonlinear buckling behaviour of concrete thin-shells, using Multi-Fidelity deep learning. The prediction speed of these models could potentially be used to improve design space exploration during the structural shape optimisation phase. Indeed, the use of nonlinear Finite Element (FE) analysis for estimating the buckling factor is unpractical in such settings because of its high computational cost. This research considers the use of Multi-Fidelity models to mitigate the computational cost required to constitute a sufficiently large dataset for training deep learning models. Two datasets - a low-fidelity dataset and a high-fidelity dataset – that contain concrete thin-shells with various shapes and material properties were therefore generated. The buckling factor of the 20,000 thin-shells in the low-fidelity dataset were obtained through linear eigenvalue FE analysis, which has a low computational cost. Additionally, the buckling factors of the 5,000 thin-shells in the high-fidelity dataset were obtained using computationally expensive nonlinear FE analyses. These datasets were used to train Multi-Fidelity Multilayer Perceptrons in two different approaches: using several sequentially connected models, and using Transfer Learning. These two approaches were also compared to a Single-Fidelity baseline. It was found that the models were able to make highly accurate predictions of the nonlinear buckling factor (Mean Absolute Errors are consistently below 0.65%), while being more than 97,000 times quicker than the average time required for a nonlinear FE analysis. Additionally, the Multi-Fidelity approaches were found to be beneficial when the amount of high-fidelity data is limited.