A new approach for constructing two Bayesian network models for predicting the liquefaction of gravelly soil

Many studies have indicated that the triggering conditions for gravelly soil liquefaction are different from those for sandy soils. However, the existing prediction methods and models do not consider the differences. Moreover, most approaches of constructing Bayesian network (BN) models for predicting seismic liquefaction based on domain knowledge or data-driven are either too subjective or too objective, resulting in suboptimal structures. Therefore, to solve these shortcomings, two new BN models for predicting gravelly soil liquefaction are constructed using a new hybrid approach combining the maximal information coefficient and domain knowledge based on the dynamic penetration test and shear wave velocity test databases. The performance of the proposed hybrid approach is validated by comparing other existing modeling approaches, and two new BN models performed much better than other models in both two databases compared with the existing models or methods for predicting gravelly soil liquefaction in the training, validation, and testing sets. Furthermore, the differences and advantages of all methods or models mentioned in this paper are discussed, and factor sensitivity analysis in the BN models illustrates that those triggering conditions different from sandy liquefaction are worth considering in the prediction of gravelly soil liquefaction.