Seismic Soil Liquefaction Hazard Assessment Using Bayesian Belief Networks

Discernment of seismic soil liquefaction and the resulting lateral displacement is a complex and nonlinear procedure that is affected by diversified factors of uncertainties and complexity.However,with increasing availability of in-situ data,artificial intelligence(AI)techniques such as artificial neural network(ANN)and adaptive neuro-fuzzy inference system(ANFIS)have been successfully used for evaluation of liquefaction potential and liquefaction-induced lateral displacement with improved accuracy in comparison with available methods.However,most of the introduced AI techniques have certain limitations such as difficulty in concluding the assessment results due to limited use of prior knowledge.Furthermore,the accuracy of the predictive models is not well interpreted but afar from addressing the problem of seismic liquefaction potential and liquefaction-induced lateral displacement completely.Therefore,the issue of seismic liquefaction hazard still posed considerable challenge for geotechnical earthquake professionals.Hence,it is essential to advance the idea that making more systematic and in-depth research on predicting the liquefaction potential and the resulting lateral displacement.In the recent past,Bayesian belief network(BBN)is being used as an alternate artificial intelligence technique that permits a probabilistic relationship among the set of variables and present a suitable framework to handle cause-effect relationships and uncertainties.The aim of this research is to develop new probabilistic graphical models to evaluate the liquefaction potential of soil and lateral displacement using Bayesian belief networks based on reliable post liquefaction in-situ test database.The scope of this research study includes the following main stages:At first,eleven significant factors of seismic soil liquefaction such as earthquake magnitude,closest distance to rupture surface,soil behavior type index,equivalent clean sand penetration resistance etc.were identified by systematic literature review(SLR)approach bearing in mind the selection principle of influence factors.The multilevel hierarchy structure was developed using interpretive structural modeling(ISM)and the strength of the relationship was examined using Matrice d'impacts croises multiplication appliquee a un classment(MICMAC)methodologies.The results provide a more accurate way for further establishment of quantitative method,such as Bayesian belief network for seismic soil liquefaction hazard assessment models within probabilistic framework.Secondly,simple probabilistic graphical models using five factors of seismic soil liquefaction were developed based on domain knowledge(DK)employing interpretive structural modeling(ISM)method,K2 machine learning(ML)algorithm,and hybrid approach combining DK and ML algorithm to evaluate seismic soil liquefaction potential.Several metrics such as overall accuracy(OA),precision,recall and F-measure were used to evaluate the performance of the proposed models that develop a quantitative basis for comparison with other models.The BBN model developed by hybrid approach i.e.,ISM and K2 ML algorithm shows relatively better performance than BBN models developed by ISM and K2 ML techniques as the hybrid approach methodology integrates the strengths of DK and K2 ML algorithm which elude the shortcomings of utilizing one method(i.e.,DK or ML algorithm)to conclude BBN.The BBN-K2 and DK model overall performance evaluation for the entire dataset is compatible and promising in liquefaction potential evaluation with those compared with ANN model from literature and C4.5 DT model in present study.Further,to extend the application scope of previously simple probabilistic graphical model based on five factors',a multi-factor probabilistic graphical model was proposed using the updated and relatively large cone penetration test(CPT)dataset based on the 11 significant factors of seismic soil liquefaction.The accuracy and robustness of a multi-factor probabilistic graphical model is verified by comparing with the C4.5 DT,simplified procedure and an evolutionary-based approach in terms of performance evaluation metrics.The results show that a multi-factor probabilistic graphical model is preferred over the others.Owing to its overall performance,simplicity in practice,data-driven characteristics,and ability to map the interactions between variables,the use of a multi-factor probabilistic graphical model in assessing seismic soil liquefaction is quite promising.A multi-factor robust probabilistic graphical model can not only quantitatively predict seismic soil liquefaction potential probability under certain influence factors of seismic,soil,and site conditions,but also identify the main diagnostic reasons and fault-finding states'combinations presumed to support decisions on seismic soil liquefaction mitigation measures for sustainable development.Finally,the novel probabilistic framework for evaluating liquefaction-induced lateral displacement using Bayesian belief network(BBN)approach based on interpretive structural modeling technique was presented.The developed BBN models were trained and tested using a wide-ranged case history records database to predict lateral displacements for free face and sloping ground conditions.The predictive performance results of the proposed BBN models were compared with the frequently used multiple linear regression(MLR)and genetic programming models.The results reveal that the BBN models are able to learn the complex relationship among lateral displacement and its influencing factors by means of a cause-effect relationship with reasonable precision.In the developed models there is no need to add new parameters(such as R*in the MLR model)or use functional(such as logarithmic,etc.)values affecting parameters;all parameters can be used into the model as they are,without any normalization or calibration.Furthermore,the results of sensitivity analysis present that the"peak ground acceleration,amax" and "average mean grain size D5015(mm)in cumulative thickness of saturated layers with corrected SPT number(N1)60<15" are the two most significant factors in the both BBN models.