2026 Volume 7 Issue 3 Pages 890-901
Among the items of the seismic safety verification of integrated structures composed of road tunnels and river levees, the one focusing on the river levee requires confirming through dynamic analysis, accounting for liquefaction, that the post-earthquake residual levee height exceeds the design external water level. However, such dynamic analyses are time-consuming and costly to perform. In this study, physically meaningful features were deterministically extracted as explanatory variables from previously conducted dynamic analysis data, and the settlement of the river levee crown, taken as the target variable, was predicted using the foundation model TabPFN. The prediction accuracy was evaluated through leave-one-out cross-validation, while the explanatory variables were narrowed down by sequential forward selection. As a result, a surrogate model capable of predicting levee settlement quickly and easily was constructed. The constructed model predicted the levee crown settlement with a coefficient of determination of R² = 0.784−0.969, even for the small sample of nine cross-sections per case. Furthermore, the reliability of the predictive distribution output by TabPFN was evaluated using the coverage probability (PICP), the probability integral transform (PIT), and the continuous ranked probability score (CRPS), indicating that the model can support the evaluation of the safety margin in the seismic safety verification.