Integration of Double-Weighted Bayesian and Simplified Methods for Predicting Seismic Liquefaction Based on Multiple Databases

The Bayesian method is a versatile data-driven machine learning method that performs well in predicting seismic-induced soil liquefaction, but it does not consider physical mechanisms and its performance is easily affected by class imbalance and attribute weights. In addition, simplified methods consider the mechanism, while simplified methods based on different in situ tests often produce conflicting results for the same site, leaving engineers unable to decide which result to choose. To overcome the aforementioned problems, this paper proposes a framework for combining multiple simplified methods based on the double-weighted Bayesian combination (DWBC) approach, considering the effects of combination mode, class imbalance, and contribution weights of the simplified methods on the performance of the DWBC model. Compared with the three simplified methods based on different in situ tests, the proposed DWBC model significantly improves the liquefaction prediction accuracy and converts the deterministic prediction result to probabilistic. Furthermore, when comparing different ensemble strategies (e.g., majority voting, simple average, and weighted average approaches), different Bayesian combination modes, and the random forest (RF) model based on 250 liquefaction multidatabases using various performance measures, the DWBC model performs the best, followed by the Bayesian combination model without weighting and the majority voting method, while the RF model performs the worst. The performance of the DWBC model depends on the number and mode of the basic classifiers and the performance of the basic classifiers. The sensitivity of the DWBC method with respect to the class imbalance is also discussed.Practical ApplicationsSeismic liquefaction is a form of earthquake-induced disaster phenomenon. This study constructs an ensemble model for predicting earthquake-induced liquefaction based on the double-weighted Bayesian method to improve the prediction accuracy. The ensemble model takes the prediction results of the widely used simplified methods in various in situ test databases such as standard penetration test, cone penetration test, and shear wave velocity as inputs and liquefaction or nonliquefaction as outputs, while considering the effects of combination mode, class imbalance, and contribution weights of the simplified methods on the performance of the ensemble model. Thus, the ensemble model can avoid the situation where simplified models predict conflicting results in different in situ test databases for the same site and convert the deterministic prediction results of simplified methods into a probabilistic result. In this study, the proposed ensemble model performs much better than the simplified models and other ensemble models such as the random forest.