Application of various machine learning approches to estimate liquefaction risk

Abstract

In this thesis, the possibility of using quadratic discriminant analysis (QDA), artificial neural networks (ANN), random forest and support vector machine (SVM), which are four famous machine learning approaches, to model the complex relationship between liquefaction risk and soil seismic features has been investigated. Nowadays with the development of computational speed, such approaches can give engineers faster and economical results and in many cases there is no need to take extreme assumptions about the structure of a problem in order to simplify and make it solvable. Machine learning techniques use data to extract information. For this thesis, a liquefaction database with 415 case histories has been used. Three soil parameters (depth to critical layer, σv, VS1) and two seismic parameters (Mw, PGA) are considered as the models inputs and the liquefaction potential of soil is the output. It has been shown that all of the mentioned models can reasonably predict whether a soil is liquefiable or not, however, random forest outperformed the other methods and showed the most accuracy amongst the models. Finally random forest performance has been compared to the performance of the simplified approach, which is a traditional solution, to assess whether a soil is liquefiable or not.