Predicting seismic-induced liquefaction potential of gravelly soils using dynamic penetration case histories

As demonstrated by numerous environmental disasters worldwide, many sites have suffered from seismically induced liquefaction, resulting in substantial economic losses. Consequently, there is an urgent need for reliable prediction methods to assess vulnerability to liquefaction. In this study, the liquefaction potential (LP) of gravelly soil sites is predicted using available seismological parameters (Mw, R, t, PGA), soil parameters (G, F, D50), and site profile parameters (N’120, σ’v, Dw, Hn, Dn) through AI-based symbolic regression techniques, namely response surface methodology (RSM), genetic programming (GP), evolutionary polynomial regression (EPR), and the group method of data handling neural network (GMDH-NN). A total of 234 data records were compiled from earthquake case histories reported in the literature and divided into 80% for training and 20% for validation. RSM was employed to model the database, whereas GP, EPR, and GMDH-NN were used for liquefaction potential classification in the investigated area. Comparative evaluation of model performance indicates that the RSM yielded a statistically significant parametric LP equation, operating with a degree of variation of 0.61 and a p-value of 0.0001. For the classification models, GP, EPR, and GMDH indicate that liquefaction is expected when the predicted LP ≥ 0.5, while liquefaction is not triggered when LP < 0.5. The total misclassification cases were categorized into positive errors, where liquefaction occurred but was not predicted, and negative errors, where liquefaction was predicted but not observed, with the latter being more conservative. In practice, negative errors are not entirely definitive, as liquefaction may locally occur beneath surface layers without being visibly manifested. Although all four predictive models achieved comparable accuracy levels ranging from 88% to 90%, further analysis revealed that the GP model produced 12% positive errors, whereas both EPR and GMDH models resulted in only 6% positive errors, indicating that they are more conservative and safer than the GP model. In addition, the GMDH model is considerably more complex than the EPR model, providing EPR with a notable practical advantage. Correlation analysis further demonstrated that the vertical effective overburden stress (σ’v) and the dynamic penetration test blow count (N’120) are the most influential parameters, each exhibiting a correlation coefficient greater than 0.3g.