2019 CSCE Annual Conference - Laval (Greater Montreal)

2019 CSCE Annual Conference - Laval (Greater Montreal) Conference


Title
On developing proper single and ensemble machine learning frameworks for predicting seismic-based soil liquefaction


Author(s)
Dr. Mohammad Alobaidi, McGill University (Presenter)
Dr. Mohamed Meguid, McGill University
Dr. Fateh Chebana, Eau Terre Environnement, Institut National de la Recherche Scientifique
Abstract

Seismic-induced liquefaction of soils is one of the major ground failure consequences of earthquakes. This phenomenon leads to catastrophic loss of lives and irreversible damage to the critical infrastructure. Prediction of liquefaction potential is, hence, considered as the major research frontier in geotechnical earthquake engineering. Conventional techniques used to determine the level of liquefaction susceptibility of soil mainly relied on informed-determination of liquefaction threshold from observed soil behavior records against key-parameters such as Standard Penetration Test and Shear Wave Velocity. Statistical procedures were also commonly utilized in producing useful inferences about the likelihood of liquefaction given some soil-related as well as earthquake-related information. Nevertheless, with the ever-lasting need of more robust techniques that can provide better generalization ability over a wide collection of liquefaction observations from different databases, rather than local thresholding of the phenomenon, more complex methods are required. The availability of computational recourse nowadays further motivated the creation of more complex techniques that have provider far better performance than their predecessor. Recently, evolutionary Machine Learning techniques have been proposed in the broad literature and provided superior performance in learning complex relationships, while maintaining a reliable generalization ability. A more recent advancement in Machine Learning is Ensemble Learning. This new framework is generally defined as the generation-learning-fusion of individual Machine Learning models in a predefined ensemble architecture that not only produces a far stable global model, but also guarantees higher performance and diminishing uncertainty. To this extent, in liquefaction prediction studies, limited adherence to proper utilization of Machine Learning have been observed over the considered literature. Also, little attention has been paid to the recent development in supervised learning techniques and exploiting them toward enhanced liquefaction prediction. This work presents different Ensemble Learning frameworks and examines their capability in enhancing liquefaction prediction. Different Machine Learning techniques in Ensemble Learning frameworks are investigated over a seismic-based liquefaction database. The performance of the resulted models is compared with single models used in the literature, and the effect of the training data-availability on the ensemble models’ generalization ability is also examined.

Keywords: Machine learning; ensemble learning; classification; liquefaction prediction.