Quasi-site-specific prediction for foundation capacity based on load-test database
Journal
Computers and Geotechnics
Journal Volume
184
Start Page
107305
ISSN
0266352X
Date Issued
2025-08
Author(s)
Wang, Jiun-Shiang
Abstract
This study proposes a novel hierarchical Bayesian model (HBM)-based stacking ensemble learning framework for the purpose of foundation capacity prediction. Site uniqueness poses a significant challenge to geotechnical design prediction because data from other sites cannot be directly transferred to the target site. As a result, there are usually significant uncertainties in the prediction because the target-site data are limited. For the purpose of soil/rock property prediction, quasi-site-specific models (such as HBM) have been proposed in the literature to assist the prediction by a soil/rock property database. For the purpose of foundation capacity prediction, it is also possible to assist the prediction by a load-test database. However, the current study shows that the performance of the original HBM degrades significantly because a load-test database is usually very sparse, and the original HBM trained by the very sparse database may have exceedingly high variability. To overcome this limitation of the original HBM in foundation capacity prediction, this study proposes a HBM-based stacking ensemble learning framework that selects the best learner from three base learners (three models with various complexity) for bias-variance tradeoff. The proposed framework is illustrated by a real case, and its prediction performance is validated by extensive leave-one-site-out cross-validations. Validation results suggest that the proposed framework generally outperforms the original HBM in making quasi-site-specific prediction for foundation capacity.
Subjects
Data-centric geotechnics
Foundation capacity
Hierarchical Bayesian model
Load-test database
Quasi-site-specific prediction
SDGs
Publisher
Elsevier Ltd
Type
journal article