EVALUATION OF LIQUEFACTION POTENTIAL OF SOILS BY USING PROBABILISTIC AND MACHINE LEARNING METHODS

Abstract

Soil liquefaction is a substantial seismic hazard that endangers both human life and infrastructure. This research specifically examines the occurrence of soil liquefaction events in past earthquakes, with a special emphasis on the 1964 Niigata, Japan and 1964 Alaska, USA earthquakes. These occurrences were important achievements in the comprehension of harm caused by liquefaction. Geotechnical engineers often use in-situ experiments, such as the standard penetration test (SPT) to evaluate the likelihood of liquefaction. The attraction for this option arises from the difficulties connected in acquiring undisturbed samples of superior quality, as well as the related expenses. Although shear wave velocity tests and Baker penetration tests are alternative in-situ testing, they are less often used in this assessment procedure. Geotechnical engineering specialists choose the deterministic framework for liquefaction assessment because of its clear mathematical approach and low needs for data, time, and effort. This work emphasises the need of integrating probabilistic and reliability methodologies into the design process of crucial life line structures to enable well-informed risk-based decision-making. This research investigates the several methodologies and protocols used by scholars to construct prediction models for evaluating the likelihood of liquefaction. Recently, many models like as Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forests (RF), Genetic Programming (GP), Ensemble models, and SVM-Grey Wolf Optimisation (SVM-GWO) have been extensively used to assess the likelihood of liquefaction. Furthermore, the use of probability-based models that include reliability analysis has shown its effectiveness in this matter. This research investigates the constraints of certain models, such as their sluggish convergence rate, vulnerability to model overfitting, and dependence on the inclusion of a single variable. The primary objective of this work is to investigate the nonlinearity of liquefaction and improve existing techniques for assessing liquefaction susceptibility. This initiative adds to the area of geotechnical engineering by helping to reduce the dangers associated with liquefaction. The objective of this project is to create machine learning models that use deterministic, probabilistic, and reliability-based methods to evaluate the likelihood of soil liquefaction. The work presents a new equation that combines Bayes conditional probability with Genetic Programming (GP). In addition, a novel soil liquefaction prediction model is presented, which improves the correlation features, chi-square, relief characteristics, and technical indicators. This research examines the efficacy of ensemble classifiers, including Deep Belief Networks (DBN), Long Short-Term Memory (LSTM), and Support Vector Machines (SVM), when vii integrated with an optimised Bidirectional Gated Recurrent Unit (Bi-GRU), to improve the accuracy of predictions. This research suggests using a new technique called Average Cat and Salp swarm algorithm (AC-SSO) and an Opposition-based self-adaptive shark smell optimizer (OSA-SSO) model to find the best weights in the Bi-GRU model. This research aims to analyse post-liquefaction and borehole data obtained from the National Capital Region (NCR), with a special emphasis on Delhi. The data used in this study is obtained from the Standard Penetration Test (SPT) database. This research aims to assess the potential for liquefaction and provide performance metrics for liquefaction in the field. The created models are used to construct SPT CRR models. The suggested deterministic approaches use genetic programming (GP) to create CRR models in combination with the commonly employed CSR7.5 model. This research evaluates the effectiveness of deterministic models specifically designed for SPT by comparing them to established statistical approaches utilising separate datasets. This research aims to assess the likelihood of liquefaction occurring by using probabilistic assessment methods. It especially focuses on determining the probability of liquefaction (PL) and measuring the level of caution inherent in deterministic models when it comes to PL. This thesis investigates the correlation between the Fs and PL by using mapping functions derived from the Bayesian theory of conditional probability. This work evaluates the predictive accuracy of created SPT-based probabilistic models over different PL limitations, in comparison to current probabilistic models. The study presents a sophisticated soil liquefaction prediction model that integrates improved correlation features, chi square analysis, relief characteristics, and technical indicators. This thesis investigates the incorporation of ensemble classifiers, such as Deep Belief Networks (DBN), Long Short-Term Memory (LSTM), and Support Vector Machines (SVM), with an optimised Bidirectional Gated Recurrent Unit (Bi-GRU) to get dependable prediction results. This study presents a new method for calculating the best weights in Bi-GRU by using an innovative AC-SSO and OSA-SSO model. The suggested models are then assessed and contrasted with pre-existing models, including both augmented correlation properties and those without such enhancements. This study presents a strategy that efficiently decreases the percentage of false negatives, which is a crucial part of evaluating the model. By acknowledging the possibility of failure and integrating safety measures, significant improvements have been noticed.