Influence of Uncertainties in Critical State to Soil’s Mechanical Responses

Abstract

Failure strengths of the soils are controlled by critical state conditions. Laboratory tests are generally conducted to obtain these. However, many infrastructures experience stresses conditions that cannot be achieved using typical laboratory equipment. As a result, it is crucial to characterize these uncertainties for design practice. The goal of this study is to create prediction models of critical state conditions with uncertainties for the wide stress ranges for various soils. This research develops a database of critical states of various soils through a literature review, and presents a prediction model of the critical state conditions using predictor variables from particle gradation, morphology and mineralogy. Four distinct machine learning methods— Boosted Tree, K-Nearest Neighbors (KNN), Neural Networks, and Bootstrap Forest—were meticulously applied and evaluated based on their predictive accuracy, as gauged by standard metrics such as R-square and Root Average Squared Error (RASE). The Boosted Tree model demonstrated exceptional predictive consistency across training, validation, and testing phases. The KNN model provided insights into the importance of selecting the appropriate number of neighbors, with an optimal balance achieved at K=3. Neural Networks, with a single hidden layer of 10 neurons, showcased their capacity to handle non-linearity in data, while the Bootstrap Forest model underlined the advantage of using ensemble methods to improve prediction robustness. A comparative analysis revealed that ensemble methods, particularly the Boosted Tree and Bootstrap Forest models, achieved slightly higher accuracy and generalizability, suggesting their suitability for complex soil behavior predictions. The study's findings underscore the potential of machine learning as a transformative tool in predicting soil behavior at critical state.

Date of Award	8 May 2024
Original language	American English
Supervisor	Tadahiro Kishida (Supervisor)