BRIDGE: Bridging Routine Inputs and Digital modelling for Geotechnical Engineering

Abstract

The mechanical properties needed for geotechnical design require laboratory tests and are thus available for far fewer field samples than routinely measured basic (index) properties which are available for nearly every sample. We term this difference as the characterization gap. Purely data-driven models do not explicitly enforce physical laws and may be unreliable beyond their training domain, while direct constitutive calibration preserves physical consistency but requires advanced laboratory data and expert judgment. We therefore propose BRIDGE (Bridging Routine Inputs and Digital modelling for Geotechnical Engineering) a modular framework that predicts mechanical properties from routinely available basic properties. BRIDGE screens the dataset for outliers and non-physical responses, bins the retained samples into Cells, and selects a small calibration dataset from each Cell, while the remaining samples are used for validation. A mechanical model is then calibrated for each calibration sample, and the converged calibrations collectively define the mechanical surrogate of the Cell together with its parameter ranges and the corresponding predictive envelope in the mechanical property space. A Cell is validated when most validation samples fall within this envelope, whereupon the surrogate can predict the mechanical property ranges of new samples whose basic properties lie within the Cell. In this work, we adopt a finite element implementation of the Extended Cam-Clay (ECC) constitutive model as the mechanical model, whose four parameters are calibrated in a multi-stage process. Applied to the Tokyo-CLAY/14/67760 database, BRIDGE predicts the mechanical properties of an entire Cell from calibrations of just 10 to 20% of its samples. Finally, since the mechanical model, the binning criteria, and the sampling method are components of BRIDGE, the framework is easily extensible.