CPT-based fully probabilistic seismic liquefaction potential assessment to reduce uncertainty: Integrating XGBoost algorithm with Bayesian theorem

The presence of model and parameter uncertainties significantly affects seismic liquefaction potential assessments and may lead to improper geotechnical design. This study develops a fully probabilistic framework for liquefaction potential evaluation to reduce these uncertainties. It contains two major components: (i) an extreme gradient boosting (XGBoost) algorithm-based model to predict the probability of liquefaction (P-L) directly to deal with the uncertain liquefaction/non-liquefaction boundary; (ii) the Bayesian theorem to integrate prior knowledge with site-specific cone penetration test (CPT) data to obtain the updated distributions of input parameters. Comprehensive up-sampling and model validation methods are adopted to develop a reliable model building procedure and select the optimal liquefaction threshold. Probability contour maps on the normalized soil behaviour type (SBTn) chart and simplified probabilistic model are also provided to improve the practical feasibility. The results show that the XGBoost model can effectively predict the P-L and reduce model uncertainty. By integrating the XGBoost model with the Bayesian theorem, the parameter uncertainty can be considered explicitly and rigorously, and the updated P(L )distribution, considering the parameter uncertainty, is obtained. The proposed framework delivers reliable prediction of P-L and can be treated as an alternative or a supplementary technique to deterministic assessments.

Automated summary

This study develops a fully probabilistic framework for liquefaction potential evaluation, which includes an XGBoost model and the Bayesian theorem. It effectively predicts the probability of liquefaction and obtains updated parameter distributions. With comprehensive up-sampling and model validation methods, a reliable model building procedure is established, and probability contour maps and a simplified probabilistic model are provided for improved practical feasibility.