Optimization for Improvement of Ground Consolidation Settlement using Machine Learning

Abstract

We studied a new approach to optimize the improvement of ground consolidation settlement using machine learning. We estimate a ground settlement by using a FEM simulator called MuDIAN. The simulator calculates a two-dimensional settlement pattern from a two-dimensional improvement pattern. Firstly, we explore improvement patterns to minimize the settlement by random design of experiment and additional genetic algorithm optimization (Step-1). Secondly, we analyze the exploration result using machine learning (Step-2). The machine learning model consists of an encoder and two decoders using Convolutional Neural Networks (CNN). The CNN encoder can transform two-dimensional improvement patterns into a latent space and visualize the similarity of the patterns. The one of CNN decoders can predict a settlement pattern from an arbitrary point on the latent space. Another CNN decoder can recover an improvement pattern. We can predict other improvement patterns which are not included in the exploration result to use these CNN decoders.