Interpretable early-stage advance rate prediction for slurry shield tunnelling using TabPFN-SHAP

Accurate prediction of advance rate (AR) during early-stage shield tunnelling remains challenging because monitoring data are limited while operational parameters must adapt to changing geological conditions. This study develops an interpretable TabPFN–SHAP framework for AR prediction and construction decision support under limited-data conditions. Using monitoring data from Rings 1–273 of the Jiangyin–Jingjiang Yangtze River Tunnel, 12 representative geological and operational variables were selected from 36 candidates through TabPFN-based forward feature selection. On the independent test set, TabPFN achieved an MAE of 0.609, MAPE of 2.03 %, RMSE of 0.906, and R 2 of 0.930, demonstrating high-accuracy AR prediction during the early excavation stage. SHAP and generalized additive models further identified dominant predictive factors, nonlinear parameter–response relationships, statistical transition ranges, and interactions among geological and operational variables. These project-specific relationships were translated into coordinated parameter-adjustment recommendations, enabling the framework to link accurate early-stage prediction with interpretable engineering decision support. The proposed framework therefore provides a practical basis for improving the safety and efficiency of early-stage shield excavation. Future work will focus on multi-project and prospective field validation under diverse geological and construction conditions.