Stochastic Characterisation of in-situ Soil Strength Properties and their Application to Reliability-based Geotechnical Design

This thesis presents a practical workflow for reliability-based shallow foundation design that uses cone penetration test (CPT) measurements to quantify soil variability and to then inform the random finite element method (RFEM).
CPT data from a Norwegian clay site and also from a Saudi Arabian sand site are used, separately, to obtain point and spatial statistics of the relevant in-situ soil measurements, including mean, standard deviation, coefficient of variation, and vertical and horizontal correlation lengths.
These statistics are then transformed to the relevant geotechnical parameters. In the case of clays, the in-situ CPT data are converted to undrained shear strength () whereas for sands, cone tip resistances are transformed to the effective friction angle ().
This statistical information can be then used as input to a random field generator to compute a field that realistically captures the heterogeneity measured at the site.

The random fields in this thesis are generated by the local average subdivision (LAS) method.
Standard normal random fields are generated first and these are then transformed into the relevant geotechnical property ( or ) that matches the target distribution and prescribed statistical moments.
These random fields for the relevant geotechnical property are subsequently mapped onto a finite element mesh for random finite element method analyses.
The RFEM is used within a Monte Carlo simulation framework to quantify the bearing capacity of a rigid strip footing under short-term undrained conditions (expressed by the bearing capacity factor ) and under long-term drained conditions (expressed by the bearing capacity factor ).

RFEM stochastic parametric analyses are performed to quantify how variability, correlation length, and anisotropy affect the mean and dispersion of bearing capacity.
The computational outcomes from these RFEM parametric analyses are collected in the form of a database that is used to train surrogate models for fast prediction.
The prediction capabilities of the trained surrogate models are then tested against independent RFEM benchmarks using in-situ data, including non-stationary cases with depth-dependent mean trends.
In particular, the validity of the surrogate models is tested using the statistical information of the Saudi Arabian sand site including pre- and post-vibro-compaction CPT measurements.
Overall, the comparisons of the surrogate predictions against the RFEM benchmark solutions show an excellent performance.