Determination of Sample Size on Surrogate Model-Based Parameter Inverse Analysis of a Super-High Arch Dam

This paper investigates the impact of the sample size on a surrogate model in the context of parameter inverse analysis for high arch dams. A deep learning-based surrogate model is developed and integrated with Jaya optimization algorithm to enhance the computational efficiency and accuracy of the inverse analysis. The input variables for the training set of the surrogate model are generated by Latin Hypercube Sampling (LHS). The output variables are obtained based on a high-precision finite element model calculation. By comparing the model accuracy and computation time across different sample sizes (ranges from 20 to 200 times the number of input variables), the optimal sample size is identified. The study was conducted for the case study of an actual high arch dam in China for which measured data are available. The results indicate that a sample size of 100 times the number of input variables achieves a favorable balance between accuracy and computation time.