Regionalized Geospatial Liquefaction Model Using Bayesian Logistic Regression, with Case Studies in Türkiye and Alaska

ABSTRACT Geospatial liquefaction models are scalable predictive tools with global applications, but they can exhibit regional bias due to differing local characteristics or varying input data quality. This study proposes a regionalized geospatial liquefaction model framework based on Bayesian logistic regression (BLR), which leverages globally trained models as informative priors and incorporates regional data to update model parameters within a principled Bayesian framework. Two regional case studies—Türkiye and Alaska—are used to evaluate the proposed method. Results show that BLR models significantly improve upon global models in predictive accuracy, reducing bias and uncertainty while maintaining generalizability and mitigating overfitting risks associated with purely local models. A covariance-scaling strategy controls the trade-off between global priors and regional likelihoods, enabling flexible model tuning. Using data from the 2023 Kahramanmaraş earthquake for validation and the BLR model updated with data from Europe and Türkiye, accuracy improves from 0.77 to 0.84 and the F1 score from 0.74 to 0.85, with model coefficients remaining broadly consistent with the global model in terms of contributing variables (peak ground velocity [PGV], VS30, precipitation, water table depth, and distance to water). The resulting model reduces model bias for the region. Using data from the 2018 Anchorage earthquake for validation and the BLR model updated with data from that earthquake, accuracy increases from 0.62 to 0.71 and the F1 score from 0.46 to 0.73. The updated model reduces contributions from water table depth and precipitation to near zero while increasing those from PGV and distance to water. The proposed BLR framework offers a robust and adaptable solution for developing region-specific geospatial liquefaction models, particularly in regions where global model predictions exhibit significant bias.