Data-Driven Approach to Inversion Analysis of Three-Dimensional Inner Soil Structure via Wave Propagation Analysis

Various approaches based on both computational science and data science/machine learning have been proposed with the development of observation systems and network technologies. Computation cost associated with computational science can be reduced by introducing the methods based on data science/machine learning. In the present paper, we focus on a method to estimate inner soil structure via wave propagation analysis. It is regarded as one of the parameter optimization approaches using observation data on the surface. This application is in great demand to ensure better reliability in numerical simulations. Typical optimization requires many forward analyses; thus, massive computation cost is required. We propose an approach to substitute evaluation using neural networks for most cases of forward analyses and to reduce the number of forward analyses. Forward analyses in the proposed method are used for producing the training data for a neural network; thereby they can be computed independently, and the actual elapsed time can be reduced by using a large-scale supercomputer. We demonstrated that the inner soil structure was estimated with the sufficient accuracy for practical damage evaluation. We also confirmed that the proposed method achieved estimating parameters within a shorter timeframe compared to a typical approach based on simulated annealing.