Integrating Data-Centric Insights With Bayesian Approach For Robust Geotechnical Predictions
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Department of Civil, NITS
Abstract
Uncertainty is an unavoidable aspect of geotechnical engineering, driven by the inherent
variability of soils, the incompleteness of site investigations, and the practical limitations of
time, cost, and accessibility. Deterministic methods remain the dominant practice, but they
reduce uncertainty to fixed parameters and rely on conservative safety factors. While simple,
such approaches often either overestimate risk, leading to uneconomical designs, or
underestimate it. Probabilistic approaches provide a more rigorous way to manage
uncertainty, and among these, Bayesian methods are particularly compelling. They allow
engineers to incorporate prior knowledge, update predictions as new data becomes available,
and represent both parameters and predictions as probability distributions. Despite these
advantages, the adoption of Bayesian methods in practice has been limited by technical
complexity, lack of software support, and unfamiliarity among practitioners. This thesis
addresses this gap by demonstrating how Bayesian frameworks can be applied to key
geotechnical problems, improving prediction, interpretability, and decision-making under
uncertainty. The motivation for this research lies in the need to bridge the divide between
probabilistic theory and engineering practice. Four geotechnical problems were chosen
deliberately because they represent core aspects of design and analysis where uncertainty is
unavoidable, existing deterministic or empirical models are inadequate, and data are
available but often underutilised. The estimation of shear wave velocity (Vs) from SPT blow
count (N) is fundamental for seismic site characterisation, yet regional variability leads to
inconsistent predictions. Bearing capacity estimation from Plate Load Tests (PLTs) is central
to foundation design but is hindered by limited data and model generalisability. Groutability
prediction in permeation grouting is a safety-critical process where complex soil–grout
interactions defy simple modelling. Soil liquefaction prediction is one of the most important
challenges in earthquake engineering, where the consequences of misclassification are
severe. Together, these problems span different domains of geotechnical engineering while
sharing the unifying theme of decision-making under uncertainty, making them ideal for
testing and demonstrating Bayesian approaches.
Shear wave velocity (Vs) is a fundamental parameter for assessing soil stiffness and dynamic
response, playing a central role in earthquake engineering, site classification, and
liquefaction evaluation. Its empirical correlation with Standard Penetration Test (SPT) blow
count (N) provides a practical means for estimating Vs from commonly available field data.
However, this correlation is known to vary across regions due to differences in geological
history, depositional environment, and testing practices. In this objective, a two-level
Bayesian multilevel regression model is used to capture these regional differences
systematically. By transforming both variables into their natural logarithms (ln Vs and ln N),
the model assumes a common slope across regions and allows for region-specific intercepts.
This hierarchical structure captures both within-region variability and between-region
differences, enabling better predictions in data-scarce regions through partial pooling. The
Bayesian framework also quantifies uncertainty in both parameters and predictions,
facilitating the development of regional Vs maps for seismic hazard assessments. Compared
to traditional single-level regressions, the multilevel model significantly improves predictive
accuracy and interpretability by accounting for the hierarchical nature of the data.
The Plate Load Test (PLT) is widely used to evaluate the bearing capacity of shallow
foundations under both Ultimate Limit State (ULS) and Serviceability Limit State (SLS)
conditions. However, the test is expensive, time-consuming, and often site-specific, limiting
the availability of large-scale datasets. Existing machine learning (ML) models trained on
PLT data often suffer from overfitting and limited generalizability. To address these
challenges, the objective is to develop an ensemble learning framework using Adaptive
Boosting (AdaBoost) and Gradient Boosting (GBoost) models optimised through Bayesian
Optimisation (BO). A dataset of 576 PLT records from 58 sites was compiled, incorporating
input features such as settlement (S), undrained cohesion (Cu), footing width (B),
embedment depth (Df), groundwater depth (Dw), and unit weight (γt). The Bayesian
optimised GBoost model achieved superior performance compared to both AdaBoost and
conventional Artificial Neural Networks. Bayesian Optimisation enabled the efficient tuning
of hyperparameters to minimise generalisation error and avoid overfitting. The proposed
method provides a robust and interpretable approach for bearing capacity prediction and can
be integrated into preliminary design stages, improving efficiency and reliability.
Permeation grouting using cement-based suspensions is an essential ground improvement
technique, especially for controlling groundwater inflow during tunnelling and underground
construction. The effectiveness of grouting depends on several interacting factors, including
soil gradation, grout composition, water-cement ratio, and injection pressure. Predicting
whether a grouting attempt will succeed (groutability) is challenging due to these complex,
nonlinear interactions and the inherent variability in soil conditions. Existing approaches,
especially Discrete Bayesian Networks (DBNs), suffer from information loss due to the
discretisation of continuous variables, limiting their predictive capabilities. This objective
introduces a Hybrid Bayesian Network (HBN) that allows for the integration of both discrete
and continuous variables, thereby preserving the richness of continuous soil parameters
while maintaining the causal interpretability of BN structures. The HBN was constructed
using the bnlearn package in R and benchmarked against DBNs and machine learning
classifiers such as Support Vector Machines and Naïve Bayes. The HBN achieved high
accuracy, recall, and AUC, while remaining robust to class imbalance and missing data. Its
computational efficiency and interpretability make it a valuable tool for field practitioners
dealing with groutability assessment and project risk management.
Soil liquefaction is one of the most destructive phenomena that can occur during
earthquakes. This objective explores the impact of variable transformation and model
formulation on the performance of logistic regression for predicting earthquake-induced soil
liquefaction. Using 399 field case histories, the variables qc1N and CSR7.5 were
transformed into √qc1N and ln(CSR7.5) to better satisfy assumptions of linearity and
normality. Probability contour plots for sand and sand mixtures reveal distinct patterns: a
nonlinear interaction in sands. Bayesian logistic regression models were developed using
both original and transformed variables and a benchmarked empirical method. Comparisons
on an independent validation set of 101 cases show that the Bayesian model consistently
offers higher accuracy and precision, particularly at lower probability thresholds. Unlike
fixed-parameter models, the Bayesian framework captures uncertainty in model coefficients
and produces probabilistic boundaries that better reflect the transitional nature of
liquefaction behaviour. The model’s conservative classification at higher resistance levels
makes it advantageous for seismic risk mitigation.
Across these four studies, the thesis demonstrates the practical strengths of Bayesian
frameworks in geotechnical engineering. They provide explicit quantification of uncertainty,
enhance interpretability through distributions, feature importance, and causal structures, and
adapt to new data through iterative updating. The consistent use of published datasets
ensured reproducibility and transparency, showing that existing information can be
leveraged effectively without requiring costly new testing. In conclusion, this thesis argues
for a paradigm shift in geotechnical engineering from deterministic simplifications toward
probabilistic, transparent, and data-conscious methods. By demonstrating how Bayesian
approaches can be successfully applied to problems of central importance, it shows that
uncertainty need not be treated as a limitation but as an opportunity to make predictions that
are safer, more interpretable, and more adaptable. In doing so, it bridges the gap between
statistical rigour and engineering practice, offering tools that are not only theoretically sound
but also practical for real-world decision-making.
Keywords: Bayesian inference, multilevel modelling, ensemble learning, Bayesian
optimisation, hybrid Bayesian networks, liquefaction prediction, geotechnical uncertainty
Description
Keywords
Citation
Collections
Version History
You are currently viewing version 1 of the item.