Deep residual U-net with input of static structural responses for efficient U* load transfer path analysis

U* index theory is widely used to illustrate the load transfer paths inside an engineering structure. However, the conventional U* load transfer path analysis based on the finite element method is computationally demanding especially for large-scale structures. In this research, a convolutional neural network based on the architecture of residual U-Net is introduced to realize high-efficiency U* estimation of plate-type structures with arbitrary dimensions, boundary conditions, and loading conditions for the first time. Besides the geometrical information of the structures, the static structural responses including the feature maps of nodal displacement and stress are involved in the network input. Different input data combinations are experimented to study how they contribute to the model training. It is noticed that the stress and displacement data can significantly lower the output errors in U* prediction, and the geometrical information helps in noise reduction in U* contour graphs. The proposed method is tested with homogeneous plates and functionally graded plates respectively indicating its remarkable performance in load transfer path prediction. Moreover, this method shortens the U* calculation time by over 95% compared to the conventional finite element method. The improved efficiency of load transfer path analysis greatly facilitates the implementation of structural analysis, design, and optimization.