Computational Geomechanics. Manuel Pastor
(ii) bilinear for p: (c) (i) linear for u; (ii) linear for p: (d) (i) linear (with cubic bubble) for u; (ii) linear for Element (c) is not fully acceptable at incompressible–undrained limits.
We shall return to this problem in Chapter 5 where a modification is introduced allowing the same interpolations to be used for both u and p.
In that chapter, we shall discuss a possible amendment to the code permitting the use of identical u ‐p interpolation even in incompressible cases.
We note that all computations start from known values of
Thus, for known Δ
Further, we note that if pw ≥ 0 (full saturation, described in Section 2.2 of the previous chapter), then we have
and the permeability remains at its saturated value
However, when negative pressures are reached, i.e. when pw < 0, the values of Sw, χw, and k have to be determined from appropriate formulae or graphs.
3.2.4 General Applicability of Transient Solution (Consolidation, Static Solution, Drained Uncoupled, and Undrained)
3.2.4.1 Time Step Length
As explained in the previous Section 3.2.3, the computation always proceeds in an incremental manner and in the u‐p form in general, the explicit time stepping is not used as its limitation is very serious. Invariably, the algorithm is applied here to the unconditionally stable, implicit form and the equation system given by the Jacobian of (3.46) with variables Δ
With unconditional stability of the implicit scheme, the only limitation on the length of the time step is the accuracy achievable. Clearly, in the dynamic earthquake problem, short time steps will generally be used to follow the time characteristic of the input motion. In the examples that we shall give later, we shall frequently use simply the time interval Δt = 0.02 s which is the interval used usually in earthquake records.
However, once the input motion has ceased and its record no longer has to be followed, a much longer length of time step could be adopted. Indeed, after the passage of the earthquake, the remaining motion is caused by something resembling a consolidation process which has a slower response allowing longer time steps to be used.
The length of the time step based on accuracy considerations was first discussed in Zienkiewicz et al. (1984), Zienkiewicz and Shiomi (1984), and, later, by Zienkiewicz and Xie (1991), and Bergan and Mollener (1985).
The simplest process is that which considers the expansion for such a variable as u given by (3.43) and its comparison with a Taylor series expansion.
Clearly, for a scalar variable u, the error term is given by the first omitted terms of the Taylor expansion, i.e., in scalar values
(3.48)
Using an approximation of this third derivative shown below
(3.49)
we have
(3.50)
For a vector variable u, we must consider its L2 norm, i.e.
(3.51)
and we can limit the error to
This limit was reestablished later by Zienkiewicz and Xie (1991) who replaced the leading coefficient of (3.52) as a result of a more detailed analysis by
(3.53)
Whatever the form of error estimator adopted, the essence of the procedure is identical and this is given by establishing a priori some limits or tolerance which must not be exceeded, and modifying the time steps accordingly.
In the above, we have considered only the error in one of the variables, i.e. u but, in general, this suffices for quite a reasonable error control.
The tolerance is conveniently chosen as some percentage η of the maximum value of norm ‖u‖2 recorded. Thus, we write
(3.54)
with some minimum specified.
The time step can always be adjusted during the process of computation noting, however, not to change the length of the time step by more than a factor of 2 or ½, otherwise unacceptable oscillations may arise.
3.2.4.2 Splitting or Partitioned Solution Procedures
The most common way of solving the coupled system of equations is the monolithic scheme where both the coupled equations are solved simultaneously at each time step as discussed above. However, also splitting solutions are possible where the two (or three) equations are solved separately at each time step. To conserve the coupled nature of the problem, iterations within each time step are necessary even in the linear case. The reason for applying partitioned solutions may be the fact that solvers for the solid