Prediction of pore pressure and deviatoric stress generation for liquefiable soils under cyclic loading using machine learning

Abstract

Earthquakes are prominent causes in building failures and loss of lives. Along with the structural reasons, soil related reasons have also substantial effects on earthquake performance of structures. Especially, soil liquefaction stands out with the instant occurence and devastating outcomes. Specifically, the prediction of seismic-induced liquefaction potential in granular soils has arisen as a notable concern. To predict the cyclic behavior of liquefiable soils, many theoretical constitutive models are developed based on many assumptions and complex mathematical background. However, theoretical models either do not adequately explain the observed behavior or requires numerous soil parameters. Since industry demand is fast, simple and precise models rather than complex models having vast amount of parameters and necessitates engineers having higher education, utility of complex models are disadvantaged. Therefore, it is necessity to propose a novel approach with considering accuracy and efficiency trade-off, employing cutting edge technologies. In this thesis, Machine Learning (ML) methods are employed to predict cyclic behavior of saturated liqufiable soils and first steps of the smart constitutive model is taken. Databases of stress controlled Cyclic Simple Shear Tests (CSST) and strain controlled Cyclic Triaxial Tests (CTX) conducted on clean sands are collected. Data gathering are performed with both collecting raw spreadsheets and digitizing the images with available softwares. CSST database contains 263 laboratory tests compiled from 12 sources. CTX database contains 56 laboratory tests compiled from 3 sources. These extensive databases are important to encompass wide input feature range and hence, reducing the risk of overfitting and increasing generalization capacity. With this amount of samples and wide input range, this study distinguishes from similar studies in the literature. Scope of this thesis are predicting entire excess pore water pressure ratio (ru) buildup in CSST database and predicting entire deviatoric stress (σd) buildup in CTX database. Inputs for the CSST database are variation of the number of cycles as a time series (N), variation of the cyclic stress ratio as a time series (CSR), initial relative density (Dr) and the initial effective vertical stress (σ'v). Inputs for the CTX database are the variation of time in seconds as a time series (s), variation of the cyclic strain as a time series (ε), initial relative density (Dr), initial effective confining stress (σ'3) and initial void ratio (e). Capturing both general trend of the graphs and finding the behavior near to the liquefaction point are aimed in predictions. The study initially focused on the CSST database with creating different ML based models, seeking the best model option in both prediction accuracy and robustness. Namely, Random Forest Method (RF), XGBoost, Stacked Long Short-Term Memory (LSTM) network, One Dimensional Convolutional (Conv1D) network and Encoder-Decoder network are created respectively. RF and XGBoost models are built-in methods that are not directly produced for time series problems. It is understood that despite RF and XGBoost models are competitive with deep learning (DL) based models in validation scores and robustness, their prediction graphs deviate especially near to the liquefaction zone. Stacked LSTM network, Conv1D network and Encoder-Decoder network have promising results in CSST database considering difficulty level of problem with having distinct inputs and patterns. DL based models have prediction capability with roughly 10% error. Tested samples are predicted with smaller mean absolute error (MAE) than 0.15 in the 70% of the samples, in the outputs that normalized between 0 and 1. MAE for average K-Fold Cross Validation scores from the stacked LSTM model, Conv1D model, and Encoder-Decoder model altered to 0.1346, 0.1323, and 0.1252, respectively. However, it is inferred that CSST database has still robustness issues. But, after trying numerous models, it can be said that this problem occurs mainly because of database itself having distinct samples and wide input feature range. Then, the know-how gained in the CSST database is applied on CTX database. Stacked LSTM network and Encoder-Decoder networks are created in the CTX database. Conv1D network is not applied in CTX database because there was not meaningful increase in predictive capacity and robustness observed in CSST database. In CTX database, models reached prediction capacity with 1% error for predicting σd and nearly 70% of the samples are predicted with MAE smaller than 0.02 for outputs that normalized between 0 and 1. Robustness problem is not observed in CTX database. This is substantial success with considering nature of the problem. Since CTX database comprised of strain-controlled tests, using this model in Finite Element Method (FEM) based applications is also convenient. Overall, substantial performance is reached in both CSST database and CTX database. Since CTX database contain samples that are more relatable than CSST database, performance is higher in CTX database. This thesis offers a novel smart model to predict cyclic behavior of liquefiable soils with predicting entire behavior through the cyclic loading using comprehensive actual laboratory test results. Further usage of this models in the geotechnical applications is possible.