Integration of Bayesian optimisation with a Bayesian neural network for spatial prediction of cone penetrometer test data in geotechnical site characterisation

Geotechnical site characterisation, which is essential for reliable foundation design, faces challenges due to inherent soil spatial variability and typically relies on comprehensive site investigation (SI) data. In situ SI tests like cone penetrometer tests (CPT), while valuable for assessing soil properties, suffer from horizontal sparsity due to cost constraints, restricting accurate subsurface characterisation. Therefore, the present study introduces a novel framework that integrates a Bayesian neural network with Bayesian optimisation, termed “BNN-BO”, with the dual objective of predicting CPT cone tip resistance profiles at unsampled locations and quantifying prediction uncertainties. In the framework, the Bayesian neural network (BNN) is employed to address both epistemic and aleatoric uncertainties, while Bayesian optimisation (BO) is used to automate hyperparameter tuning to optimise model performance. The BNN-BO framework is demonstrated through comparative studies on two benchmark datasets from the literature and through application to a project-specific CPT dataset from offshore Australia. Comparative studies demonstrate that the BNN-BO model performed well compared to two recently published methods (i.e. GLasso and ASSD-BCS), particularly in terms of predictive accuracy. These findings underscore the opportunity that advanced machine learning techniques with uncertainty quantification bring to improve the accuracy and reliability of geotechnical site characterisation.