A new method of interval Bayesian penalized network for gravelly soil seismic liquefaction prediction considering parameter confidence and model flaws uncertainties

Seismic liquefaction prediction of gravelly soils is a complex systematic problem involving multiple uncertainties. Existing studies ignore the parameter confidence uncertainty introduced during the simplification of liquefaction field test data and the model flaws in the model uncertainty. This study proposes a new Interval Bayesian Penalty Network (IBPN) method. The IBPN characterizes, employing interval probabilities, the parameter uncertainty introduced by using the mean value to represent the whole critical liquefiable soil layer when the data are simplified, and subsequently dynamically optimize false negative and false positive errors in liquefaction predictions by introducing a risk-sensitive penalty function. By comparing with five existing methods, including those that consider the uncertainties, the results show that the IBPN method significantly outperforms the other algorithms in terms of prediction accuracy after simultaneously resolving the uncertainties caused by data simplification and prediction errors. The discussion revealed that considering parameter uncertainty is more important than consideration of model flaws for improving prediction accuracy. In addition, the validation of new historical seismic liquefaction data demonstrates the effectiveness and generalization ability of the IBPN method. This work not only provides a more accurate tool for gravelly soil liquefaction risk assessment but also suggests new research ideas for dealing with complex uncertain systems.