Bayesian adaptive LASSO seismic liquefaction discrimination model incorporating soil classification and model uncertainty

• A Bayesian logistic regression model with MCMC sampling method offers probabilistic soil classification using CPTu data. • Key liquefaction factors are automatically identified via Bayesian adaptive LASSO regression methods. • The soil classification index ( I c ) and model uncertainty has critical impacts on the accuracy of liquefaction prediction. • Superior generalization is demonstrated by the proposed model against benchmark methods or. Soil classification and model uncertainty significantly affect the accuracy of seismic liquefaction discrimination models. In this study, we developed a Bayesian Logistic Regression (BLR) model for soil classification and a Bayesian Adaptive Least Absolute Shrinkage and Selection Operator Logistic Regression (BALASSO-LR) model for liquefaction discrimination based on a cone penetration test with pore pressure measurement data. We aimed to evaluate the effects of two classification strategies, i.e., one-versus-one (OvO) and one-versus-rest (OvR), on the performance of a BLR model. Additionally, we evaluated the influence of liquefaction-related factors, including the soil behavior type index, I c , and performed model uncertainty analysis to enhance the predictive reliability of the BALASSO-LR model. These models were then compared with the conventional simplified soil behavior classification method and the existing logistic regression (LR) models for liquefaction prediction. The results indicated that the BLR soil classification model using the OvR scheme achieved a prediction accuracy of 81.2%, representing improvements of 16.2% and 1.9% over the conventional Soil Behavior Type (SBT) method and the OvO scheme-based BLR model, respectively. The BALASSO-LR model for liquefaction discrimination achieved an accuracy of 84.1%. Omitting the soil classification index I c decreased accuracy by 2.3%. When compared with the three simplified methods and existing LR models, BALASSO-LR exhibited an improvement of 5.7%–11.4% in the accuracy. In the uncertainty analysis, the No-U-Turn sampler (NUTS) algorithm with a prior distribution of N∼ (0,100) achieved the highest accuracy (84.1%), surpassing that of the Metropolis algorithm by 6%.