Geotechnical data is constrained with uncertainties. Among them is the natural variability
of the soil which is a type of uncertainty that cannot be reduced. When designing a
structure, this unpredictability of the soil has to be taken into account. The traditional
way of doing this is based on a deterministic design, where partial safety factors are applied
to the characteristic value of the loads and strength parameters, leading to a design
based on increased loads and decreased strengths.
The purpose of the thesis is to model the random variations in soil properties using a
probabilistic approach, based on the random field theory, in which soil properties are
characterized in terms of their means and standard deviations, and through a realiability
analysis to calibrate the partial safety factor for the undrained shear strength. The partial
safety factor in Denmark is set to gcu = 1:8, which is quite high compared to the general
factor proposed in Eurocode 7, gcu = 1:4.
Two-dimensional cross-correlated random fields for the undrained shear strength and
Young’s modulus of the soil are generated based on the matrix decomposition method.
The probabilistic finite element model is created by using different packages; Abaqus,
Python scripts, Fortran subroutines and Matlab scripts. 1000 Monte Carlo simulations
are performed, resulting in 1000 bearing capacities of the structure. The data is then
fitted and combined with the probability density function of the load a number of realizations
are performed with the crude Monte Carlo simulation technique. Through this
the partial safety factor of the undrained shear strength is calibrated to a corresponding
target reliability index. This whole procedure is performed for two examples, one for a
strip footing and one for an embankment.
The reliability-based calibration shows that the partial safety factor can be reduced significantly,
from gcu = 1:8 to gcu = 1:5, in both cases, by which the costs are reduced
too.
In addition, the effects of the correlation length on the failure mechanism is investigated,
and it is found that the failure mechanism is highly dependent on the correlation length.