Reference documentation for deal.II version Git b33e53217d 20190521 07:35:10 0400

This tutorial depends on step4.
Table of contents  


This program is devoted to two aspects: the use of mixed finite elements – in particular RaviartThomas elements – and using block matrices to define solvers, preconditioners, and nested versions of those that use the substructure of the system matrix. The equation we are going to solve is again the Poisson equation, though with a matrixvalued coefficient:
\begin{eqnarray*} \nabla \cdot K({\mathbf x}) \nabla p &=& f \qquad {\textrm{in}\ } \Omega, \\ p &=& g \qquad {\textrm{on}\ }\partial\Omega. \end{eqnarray*}
\(K({\mathbf x})\) is assumed to be uniformly positive definite, i.e., there is \(\alpha>0\) such that the eigenvalues \(\lambda_i({\mathbf x})\) of \(K(x)\) satisfy \(\lambda_i({\mathbf x})\ge \alpha\). The use of the symbol \(p\) instead of the usual \(u\) for the solution variable will become clear in the next section.
After discussing the equation and the formulation we are going to use to solve it, this introduction will cover the use of block matrices and vectors, the definition of solvers and preconditioners, and finally the actual test case we are going to solve.
We are going to extend this tutorial program in step21 to solve not only the mixed Laplace equation, but add another equation that describes the transport of a mixture of two fluids.
The equations covered here fall into the class of vectorvalued problems. A toplevel overview of this topic can be found in the Handling vector valued problems module.
In the form above, the Poisson equation (i.e., the Laplace equation with a nonzero right hand side) is generally considered a good model equation for fluid flow in porous media. Of course, one typically models fluid flow through the NavierStokes equations or, if fluid velocities are slow or the viscosity is large, the Stokes equations. In the first of these two models, the forces that act are inertia and viscous friction, whereas in the second it is only viscous friction – i.e., forces that one fluid particle exerts on a nearby one. This is appropriate if you have free flow in a large domain, say a pipe, a river, or in the air. On the other hand, if the fluid is confined in pores, then friction forces exerted by the pore walls on the fluid become more and more important and internal viscous friction becomes less and less important. Modeling this then first leads to the Brinkman model if both effects are important, and in the limit of very small pores to the Darcy equations. The latter is just a different name for the Poisson or Laplace equation, connotating it with the area to which one wants to apply it: slow flow in a porous medium. In essence it says that the velocity is proportional to the pressure gradient that drives the fluid through the porous medium.
The Darcy equation models this pressure that drives the flow. (Because the solution variable is a pressure, we here use the name \(p\) instead of the name \(u\) more commonly used for the solution of partial differential equations.) Typical applications of this view of the Laplace equation are then modeling groundwater flow, or the flow of hydrocarbons in oil reservoirs. In these applications, \(K\) is the permeability tensor, i.e., a measure for how much resistance the soil or rock matrix asserts on the fluid flow. In the applications just named, a desirable feature is that the numerical scheme is locally conservative, i.e., that whatever flows into a cell also flows out of it (or the difference is equal to the integral over the source terms over each cell, if the sources are nonzero). However, as it turns out, the usual discretizations of the Laplace equation (such as those used in step3, step4, or step6) do not satisfy this property. But, one can achieve this by choosing a different formulation of the problem and a particular combination of finite element spaces.
To this end, one first introduces a second variable, called the flux, \({\mathbf u}=K\nabla p\). By its definition, the flux is a vector in the negative direction of the pressure gradient, multiplied by the permeability tensor. If the permeability tensor is proportional to the unit matrix, this equation is easy to understand and intuitive: the higher the permeability, the higher the flux; and the flux is proportional to the gradient of the pressure, going from areas of high pressure to areas of low pressure (thus the negative sign).
With this second variable, one then finds an alternative version of the Laplace equation, called the mixed formulation:
\begin{eqnarray*} K^{1} {\mathbf u} + \nabla p &=& 0 \qquad {\textrm{in}\ } \Omega, \\ {\textrm{div}}\ {\mathbf u} &=& f \qquad {\textrm{in}\ }\Omega, \\ p &=& g \qquad {\textrm{on}\ } \partial\Omega. \end{eqnarray*}
Here, we have multiplied the equation defining the velocity \({\mathbf u}\) by \(K^{1}\) because this makes the set of equations symmetric: one of the equations has the gradient, the second the negative divergence, and these two are of course adjoints of each other, resulting in a symmetric bilinear form and a consequently symmetric system matrix under the common assumption that \(K\) is a symmetric tensor.
The weak formulation of this problem is found by multiplying the two equations with test functions and integrating some terms by parts:
\begin{eqnarray*} A(\{{\mathbf u},p\},\{{\mathbf v},q\}) = F(\{{\mathbf v},q\}), \end{eqnarray*}
where
\begin{eqnarray*} A(\{{\mathbf u},p\},\{{\mathbf v},q\}) &=& ({\mathbf v}, K^{1}{\mathbf u})_\Omega  ({\textrm{div}}\ {\mathbf v}, p)_\Omega  (q,{\textrm{div}}\ {\mathbf u})_\Omega \\ F(\{{\mathbf v},q\}) &=& (g,{\mathbf v}\cdot {\mathbf n})_{\partial\Omega}  (f,q)_\Omega. \end{eqnarray*}
Here, \({\mathbf n}\) is the outward normal vector at the boundary. Note how in this formulation, Dirichlet boundary values of the original problem are incorporated in the weak form.
To be wellposed, we have to look for solutions and test functions in the space \(H({\textrm{div}})=\{{\mathbf w}\in L^2(\Omega)^d:\ {\textrm{div}}\ {\mathbf w}\in L^2\}\) for \(\mathbf u\), \(\mathbf v\), and \(L^2\) for \(p,q\). It is a wellknown fact stated in almost every book on finite element theory that if one chooses discrete finite element spaces for the approximation of \({\mathbf u},p\) inappropriately, then the resulting discrete saddlepoint problem is instable and the discrete solution will not converge to the exact solution.
To overcome this, a number of different finite element pairs for \({\mathbf u},p\) have been developed that lead to a stable discrete problem. One such pair is to use the RaviartThomas spaces \(RT(k)\) for the velocity \({\mathbf u}\) and discontinuous elements of class \(DQ(k)\) for the pressure \(p\). For details about these spaces, we refer in particular to the book on mixed finite element methods by Brezzi and Fortin, but many other books on the theory of finite elements, for example the classic book by Brenner and Scott, also state the relevant results.
The deal.II library (of course) implements RaviartThomas elements \(RT(k)\) of arbitrary order \(k\), as well as discontinuous elements \(DG(k)\). If we forget about their particular properties for a second, we then have to solve a discrete problem
\begin{eqnarray*} A(x_h,w_h) = F(w_h), \end{eqnarray*}
with the bilinear form and right hand side as stated above, and \(x_h=\{{\mathbf u}_h,p_h\}\), \(w_h=\{{\mathbf v}_h,q_h\}\). Both \(x_h\) and \(w_h\) are from the space \(X_h=RT(k)\times DQ(k)\), where \(RT(k)\) is itself a space of \(dim\)dimensional functions to accommodate for the fact that the flow velocity is vectorvalued. The necessary question then is: how do we do this in a program?
Vectorvalued elements have already been discussed in previous tutorial programs, the first time and in detail in step8. The main difference there was that the vectorvalued space \(V_h\) is uniform in all its components: the \(dim\) components of the displacement vector are all equal and from the same function space. What we could therefore do was to build \(V_h\) as the outer product of the \(dim\) times the usual \(Q(1)\) finite element space, and by this make sure that all our shape functions have only a single nonzero vector component. Instead of dealing with vectorvalued shape functions, all we did in step8 was therefore to look at the (scalar) only nonzero component and use the fe.system_to_component_index(i).first
call to figure out which component this actually is.
This doesn't work with RaviartThomas elements: following from their construction to satisfy certain regularity properties of the space \(H({\textrm{div}})\), the shape functions of \(RT(k)\) are usually nonzero in all their vector components at once. For this reason, were fe.system_to_component_index(i).first
applied to determine the only nonzero component of shape function \(i\), an exception would be generated. What we really need to do is to get at all vector components of a shape function. In deal.II diction, we call such finite elements nonprimitive, whereas finite elements that are either scalar or for which every vectorvalued shape function is nonzero only in a single vector component are called primitive.
So what do we have to do for nonprimitive elements? To figure this out, let us go back in the tutorial programs, almost to the very beginnings. There, we learned that we use the FEValues
class to determine the values and gradients of shape functions at quadrature points. For example, we would call fe_values.shape_value(i,q_point)
to obtain the value of the i
th shape function on the quadrature point with number q_point
. Later, in step8 and other tutorial programs, we learned that this function call also works for vectorvalued shape functions (of primitive finite elements), and that it returned the value of the only nonzero component of shape function i
at quadrature point q_point
.
For nonprimitive shape functions, this is clearly not going to work: there is no single nonzero vector component of shape function i
, and the call to fe_values.shape_value(i,q_point)
would consequently not make much sense. However, deal.II offers a second function call, fe_values.shape_value_component(i,q_point,comp)
that returns the value of the comp
th vector component of shape function i
at quadrature point q_point
, where comp
is an index between zero and the number of vector components of the present finite element; for example, the element we will use to describe velocities and pressures is going to have \(dim+1\) components. It is worth noting that this function call can also be used for primitive shape functions: it will simply return zero for all components except one; for nonprimitive shape functions, it will in general return a nonzero value for more than just one component.
We could now attempt to rewrite the bilinear form above in terms of vector components. For example, in 2d, the first term could be rewritten like this (note that \(u_0=x_0, u_1=x_1, p=x_2\)):
\begin{eqnarray*} ({\mathbf u}_h^i, K^{1}{\mathbf u}_h^j) = &\left((x_h^i)_0, K^{1}_{00} (x_h^j)_0\right) + \left((x_h^i)_0, K^{1}_{01} (x_h^j)_1\right) + \\ &\left((x_h^i)_1, K^{1}_{10} (x_h^j)_0\right) + \left((x_h^i)_1, K^{1}_{11} (x_h^j)_1\right). \end{eqnarray*}
If we implemented this, we would get code like this:
This is, at best, tedious, error prone, and not dimension independent. There are obvious ways to make things dimension independent, but in the end, the code is simply not pretty. What would be much nicer is if we could simply extract the \({\mathbf u}\) and \(p\) components of a shape function \(x_h^i\). In the program we do that in the following way:
This is, in fact, not only the first term of the bilinear form, but the whole thing (sans boundary contributions).
What this piece of code does is, given an fe_values
object, to extract the values of the first \(dim\) components of shape function i
at quadrature points q
, that is the velocity components of that shape function. Put differently, if we write shape functions \(x_h^i\) as the tuple \(\{{\mathbf u}_h^i,p_h^i\}\), then the function returns the velocity part of this tuple. Note that the velocity is of course a dim
dimensional tensor, and that the function returns a corresponding object. Similarly, where we subscript with the pressure extractor, we extract the scalar pressure component. The whole mechanism is described in more detail in the Handling vector valued problems module.
In practice, it turns out that we can do a bit better if we evaluate the shape functions, their gradients and divergences only once per outermost loop, and store the result, as this saves us a few otherwise repeated computations (it is possible to save even more repeated operations by calculating all relevant quantities in advance and then only inserting the results in the actual loop, see step22 for a realization of that approach). The final result then looks like this, working in every space dimension:
This very closely resembles the form in which we have originally written down the bilinear form and right hand side.
There is one final term that we have to take care of: the right hand side contained the term \((g,{\mathbf v}\cdot {\mathbf n})_{\partial\Omega}\), constituting the weak enforcement of pressure boundary conditions. We have already seen in step7 how to deal with face integrals: essentially exactly the same as with domain integrals, except that we have to use the FEFaceValues class instead of FEValues
. To compute the boundary term we then simply have to loop over all boundary faces and integrate there. The mechanism works in the same way as above, i.e. the extractor classes also work on FEFaceValues objects:
You will find the exact same code as above in the sources for the present program. We will therefore not comment much on it below.
After assembling the linear system we are faced with the task of solving it. The problem here is: the matrix has a zero block at the bottom right (there is no term in the bilinear form that couples the pressure \(p\) with the pressure test function \(q\)), and it is indefinite. At least it is symmetric. In other words: the Conjugate Gradient method is not going to work. We would have to resort to other iterative solvers instead, such as MinRes, SymmLQ, or GMRES, that can deal with indefinite systems. However, then the next problem immediately surfaces: due to the zero block, there are zeros on the diagonal and none of the usual preconditioners (Jacobi, SSOR) will work as they require division by diagonal elements.
For the matrix sizes we expect to run with this program, the by far simplest approach would be to just use a direct solver (in particular, the SparseDirectUMFPACK class that is bundled with deal.II). step29 goes this route and shows that solving any linear system can be done in just 3 or 4 lines of code.
But then, this is a tutorial: we teach how to do things. Consequently, in the following, we will introduce some techniques that can be used in cases like these. Namely, we will consider the linear system as not consisting of one large matrix and vectors, but we will want to decompose matrices into blocks that correspond to the individual operators that appear in the system. We note that the resulting solver is not optimal – there are much better ways to efficiently compute the system, for example those explained in the results section of step22 or the one we use in step43 for a problem similar to the current one. Here, our goal shall be to introduce new solution techniques and how they can be implemented in deal.II.
In view of the difficulties using standard solvers and preconditioners mentioned above, let us take another look at the matrix. If we sort our degrees of freedom so that all velocity come before all pressure variables, then we can subdivide the linear system \(Ax=h\) into the following blocks:
\begin{eqnarray*} \left(\begin{array}{cc} M & B \\ B^T & 0 \end{array}\right) \left(\begin{array}{cc} U \\ P \end{array}\right) = \left(\begin{array}{cc} F \\ G \end{array}\right), \end{eqnarray*}
where \(U,P\) are the values of velocity and pressure degrees of freedom, respectively, \(M\) is the mass matrix on the velocity space, \(B^T\) corresponds to the negative divergence operator, and \(B\) is its transpose and corresponds to the gradient.
By block elimination, we can then reorder this system in the following way (multiply the first row of the system by \(B^TM^{1}\) and then subtract the second row from it):
\begin{eqnarray*} B^TM^{1}B P &=& B^TM^{1} F  G, \\ MU &=& F  BP. \end{eqnarray*}
Here, the matrix \(S=B^TM^{1}B\) (called the Schur complement of \(A\)) is obviously symmetric and, owing to the positive definiteness of \(M\) and the fact that \(B\) has full column rank, \(S\) is also positive definite.
Consequently, if we could compute \(S\), we could apply the Conjugate Gradient method to it. However, computing \(S\) is expensive, and \(S\) is in fact also a full matrix. On the other hand, the CG algorithm doesn't require us to actually have a representation of \(S\), it is sufficient to form matrixvector products with it. We can do so in steps: to compute \(Sv=B^TM^{1}Bv=B^T(M^{1}(Bv))\), we
Note how we evaluate the expression \(B^TM^{1}Bv\) right to left to avoid matrixmatrix products; this way, all we have to do is evaluate matrixvector products.
vmult()
member function that does the matrixvector product. How a class chooses to implement this function is not important to the solver. Consequently, classes can implement it by, for example, doing a sequence of products and linear solves as discussed above.deal.II includes support for describing such linear operations in a very general way. This is done with the LinearOperator class that, like the MatrixType concept, defines a minimal interface for applying a linear operation to a vector:
The key difference between a LinearOperator and an ordinary matrix is however that a LinearOperator does not allow any further access to the underlying object. All you can do with a LinearOperator is to apply its "action" to a vector! We take the opportunity to introduce the LinearOperator concept at this point because it is a very useful tool that allows you to construct complex solvers and preconditioners in a very intuitive manner.
As a first example let us construct a LinearOperator object that represents \(M^{1}\). This means that whenever the vmult()
function of this operator is called it has to solve a linear system. This requires us to specify a solver (and corresponding) preconditioner. Assuming that M
is a reference to the upper left block of the system matrix we can write:
Rather than using a SolverControl we use the ReductionControl class here that stops iterations when either an absolute tolerance is reached (for which we choose \(10^{18}\)) or when the residual is reduced by a certain factor (here, \(10^{10}\)). In contrast the SolverControl class only checks for absolute tolerances. We have to use ReductionControl in our case to work around a minor issue: The right hand sides that we will feed to op_M_inv
are essentially formed by residuals that naturally decrease vastly in norm as the outer iterations progress. This makes control by an absolute tolerance very error prone.
We now have a LinearOperator op_M_inv
that we can use to construct more complicated operators such as the Schur complement \(S\). Assuming that B
is a reference to the upper right block constructing a LinearOperator op_S
is a matter of two lines:
Here, the multiplication of three LinearOperator objects yields a composite object op_S
whose vmult()
function first applies \(B\), then \(M^{1}\) (i.e. solving an equation with \(M\)), and finally \(B^T\) to any given input vector. In that sense op_S.vmult()
is similar to the following code:
(tmp1
and tmp2
are two temporary vectors). The key point behind this approach is the fact that we never actually create an inner product of matrices. Instead, whenever we have to perform a matrix vector multiplication with op_S
we simply run all individual vmult
operations in above sequence.
SchurComplement
that provides a suitable vmult()
function. Skipping over some details this might have looked like the following: All that is left for us to do now is to form the right hand sides of the two equations defining \(P\) and \(U\), and then solve them with the Schur complement matrix and the mass matrix, respectively. For example the right hand side of the first equation reads \(B^TM^{1}FG\). This could be implemented as follows:
Again, this is a perfectly valid approach, but the fact that deal.II requires us to manually resize the final and temporary vector, and that every operation takes up a new line makes this hard to read. This is the point where a second class in the linear operator framework can will help us. Similarly in spirit to LinearOperator, a PackagedOperation stores a "computation":
The class allows lazy evaluation of expressions involving vectors and linear operators. This is done by storing the computational expression and only performing the computation when either the object is converted to a vector object, or PackagedOperation::apply() (or PackagedOperation::apply_add()) is invoked by hand. Assuming that F
and G
are the two vectors of the right hand side we can simply write:
Here, schur_rhs
is a PackagedOperation that records the computation we specified. It does not create a vector with the actual result immediately.
With these prerequisites at hand, solving for \(P\) and \(U\) is a matter of creating another solver and inverse:
One may ask whether it would help if we had a preconditioner for the Schur complement \(S=B^TM^{1}B\). The general answer, as usual, is: of course. The problem is only, we don't know anything about this Schur complement matrix. We do not know its entries, all we know is its action. On the other hand, we have to realize that our solver is expensive since in each iteration we have to do one matrixvector product with the Schur complement, which means that we have to do invert the mass matrix once in each iteration.
There are different approaches to preconditioning such a matrix. On the one extreme is to use something that is cheap to apply and therefore has no real impact on the work done in each iteration. The other extreme is a preconditioner that is itself very expensive, but in return really brings down the number of iterations required to solve with \(S\).
We will try something along the second approach, as much to improve the performance of the program as to demonstrate some techniques. To this end, let us recall that the ideal preconditioner is, of course, \(S^{1}\), but that is unattainable. However, how about
\begin{eqnarray*} \tilde S^{1} = [B^T ({\textrm{diag}\ }M)^{1}B]^{1} \end{eqnarray*}
as a preconditioner? That would mean that every time we have to do one preconditioning step, we actually have to solve with \(\tilde S\). At first, this looks almost as expensive as solving with \(S\) right away. However, note that in the inner iteration, we do not have to calculate \(M^{1}\), but only the inverse of its diagonal, which is cheap.
Thankfully, the LinearOperator framework makes this very easy to write out. We already used a Jacobi preconditioner (preconditioner_M
) for the \(M\) matrix earlier. So all that is left to do is to write out how the approximate Schur complement should look like:
Note how this operator differs in simply doing one Jacobi sweep (i.e. multiplying with the inverses of the diagonal) instead of multiplying with the full \(M^{1}\). (This is how a single Jacobi preconditioner step with \(M\) is defined: it is the multiplication with the inverse of the diagonal of \(M\); in other words, the operation \(({\textrm{diag}\ }M)^{1}x\) on a vector \(x\) is exactly what PreconditionJacobi does.)
With all this we almost have the preconditioner completed: it should be the inverse of the approximate Schur complement. We implement this again by creating a linear operator with inverse_operator() function. This time however we would like to choose a relatively modest tolerance for the CG solver (that inverts op_aS
). The reasoning is that op_aS
is only coarse approximation to op_S
, so we actually do not need to invert it exactly. This, however creates a subtle problem: preconditioner_S
will be used in the final outer CG iteration to create an orthogonal basis. But for this to work, it must be precisely the same linear operation for every invocation. We ensure this by using an IterationNumberControl that allows us to fix the number of CG iterations that are performed to a fixed small number (in our case 30):
That's all!
Obviously, applying this inverse of the approximate Schur complement is a very expensive preconditioner, almost as expensive as inverting the Schur complement itself. We can expect it to significantly reduce the number of outer iterations required for the Schur complement. In fact it does: in a typical run on 7 times refined meshes using elements of order 0, the number of outer iterations drops from 592 to 39. On the other hand, we now have to apply a very expensive preconditioner 25 times. A better measure is therefore simply the runtime of the program: on a current laptop (as of January 2019), it drops from 3.57 to 2.05 seconds for this test case. That doesn't seem too impressive, but the savings become more pronounced on finer meshes and with elements of higher order. For example, an seven times refined mesh and using elements of order 2 (which amounts to about 0.4 million degrees of freedom) yields an improvement of 1134 to 83 outer iterations, at a runtime of 168 seconds to 40 seconds. Not earth shattering, but significant.
In this tutorial program, we will solve the Laplace equation in mixed formulation as stated above. Since we want to monitor convergence of the solution inside the program, we choose right hand side, boundary conditions, and the coefficient so that we recover a solution function known to us. In particular, we choose the pressure solution
\begin{eqnarray*} p = \left(\frac \alpha 2 xy^2 + \beta x  \frac \alpha 6 x^3\right), \end{eqnarray*}
and for the coefficient we choose the unit matrix \(K_{ij}=\delta_{ij}\) for simplicity. Consequently, the exact velocity satisfies
\begin{eqnarray*} {\mathbf u} = \left(\begin{array}{cc} \frac \alpha 2 y^2 + \beta  \frac \alpha 2 x^2 \\ \alpha xy \end{array}\right). \end{eqnarray*}
This solution was chosen since it is exactly divergence free, making it a realistic test case for incompressible fluid flow. By consequence, the right hand side equals \(f=0\), and as boundary values we have to choose \(g=p_{\partial\Omega}\).
For the computations in this program, we choose \(\alpha=0.3,\beta=1\). You can find the resulting solution in the results section below, after the commented program.
Since this program is only an adaptation of step4, there is not much new stuff in terms of header files. In deal.II, we usually list include files in the order baselacgriddofsfenumerics, followed by C++ standard include files:
The only two new header files that deserve some attention are those for the LinearOperator and PackagedOperation classes:
This is the only significant new header, namely the one in which the RaviartThomas finite element is declared:
Finally, as a bonus in this program, we will use a tensorial coefficient. Since it may have a spatial dependence, we consider it a tensorvalued function. The following include file provides the TensorFunction
class that offers such functionality:
The last step is as in all previous programs:
MixedLaplaceProblem
class templateAgain, since this is an adaptation of step6, the main class is almost the same as the one in that tutorial program. In terms of member functions, the main differences are that the constructor takes the degree of the RaviartThomas element as an argument (and that there is a corresponding member variable to store this value) and the addition of the compute_error
function in which, no surprise, we will compute the difference between the exact and the numerical solution to determine convergence of our computations:
The second difference is that the sparsity pattern, the system matrix, and solution and right hand side vectors are now blocked. What this means and what one can do with such objects is explained in the introduction to this program as well as further down below when we explain the linear solvers and preconditioners for this problem:
Our next task is to define the right hand side of our problem (i.e., the scalar right hand side for the pressure in the original Laplace equation), boundary values for the pressure, as well as a function that describes both the pressure and the velocity of the exact solution for later computations of the error. Note that these functions have one, one, and dim+1
components, respectively, and that we pass the number of components down to the Function<dim>
base class. For the exact solution, we only declare the function that actually returns the entire solution vector (i.e. all components of it) at once. Here are the respective declarations:
And then we also have to define these respective functions, of course. Given our discussion in the introduction of how the solution should look like, the following computations should be straightforward:
In addition to the other equation data, we also want to use a permeability tensor, or better – because this is all that appears in the weak form – the inverse of the permeability tensor, KInverse
. For the purpose of verifying the exactness of the solution and determining convergence orders, this tensor is more in the way than helpful. We will therefore simply set it to the identity matrix.
However, a spatially varying permeability tensor is indispensable in reallife porous media flow simulations, and we would like to use the opportunity to demonstrate the technique to use tensor valued functions.
Possibly unsurprising, deal.II also has a base class not only for scalar and generally vectorvalued functions (the Function
base class) but also for functions that return tensors of fixed dimension and rank, the TensorFunction
template. Here, the function under consideration returns a dimbydim matrix, i.e. a tensor of rank 2 and dimension dim
. We then choose the template arguments of the base class appropriately.
The interface that the TensorFunction
class provides is essentially equivalent to the Function
class. In particular, there exists a value_list
function that takes a list of points at which to evaluate the function, and returns the values of the function in the second argument, a list of tensors:
The implementation is less interesting. As in previous examples, we add a check to the beginning of the class to make sure that the sizes of input and output parameters are the same (see step5 for a discussion of this technique). Then we loop over all evaluation points, and for each one first clear the output tensor and then set all its diagonal elements to one (i.e. fill the tensor with the identity matrix):
The value we are going to store for a given point does not depend on its coordinates and we use the points
object only for checking that the values
object has the correct size. In release mode, AssertDimension
is defined empty and the compiler will complain that the points
object is unused. The following line silences this warning.
In the constructor of this class, we first store the value that was passed in concerning the degree of the finite elements we shall use (a degree of zero, for example, means to use RT(0) and DG(0)), and then construct the vector valued element belonging to the space \(X_h\) described in the introduction. The rest of the constructor is as in the early tutorial programs.
The only thing worth describing here is the constructor call of the fe
variable. The FESystem
class to which this variable belongs has a number of different constructors that all refer to binding simpler elements together into one larger element. In the present case, we want to couple a single RT(degree) element with a single DQ(degree) element. The constructor to FESystem
that does this requires us to specify first the first base element (the FE_RaviartThomas
object of given degree) and then the number of copies for this base element, and then similarly the kind and number of FE_DGQ
elements. Note that the RaviartThomas element already has dim
vector components, so that the coupled element will have dim+1
vector components, the first dim
of which correspond to the velocity variable whereas the last one corresponds to the pressure.
It is also worth comparing the way we constructed this element from its base elements, with the way we have done so in step8: there, we have built it as fe (FE_Q<dim>(1), dim)
, i.e. we have simply used dim
copies of the FE_Q(1)
element, one copy for the displacement in each coordinate direction.
This next function starts out with wellknown functions calls that create and refine a mesh, and then associate degrees of freedom with it:
However, then things become different. As mentioned in the introduction, we want to subdivide the matrix into blocks corresponding to the two different kinds of variables, velocity and pressure. To this end, we first have to make sure that the indices corresponding to velocities and pressures are not intermingled: First all velocity degrees of freedom, then all pressure DoFs. This way, the global matrix separates nicely into a \(2 \times 2\) system. To achieve this, we have to renumber degrees of freedom base on their vector component, an operation that conveniently is already implemented:
The next thing is that we want to figure out the sizes of these blocks so that we can allocate an appropriate amount of space. To this end, we call the DoFTools::count_dofs_per_component() function that counts how many shape functions are nonzero for a particular vector component. We have dim+1
vector components, and DoFTools::count_dofs_per_component() will count how many shape functions belong to each of these components.
There is one problem here. As described in the documentation of that function, it wants to put the number of \(x\)velocity shape functions into dofs_per_component[0]
, the number of \(y\)velocity shape functions into dofs_per_component[1]
(and similar in 3d), and the number of pressure shape functions into dofs_per_component[dim]
. But, the RaviartThomas element is special in that it is nonprimitive, i.e., for RaviartThomas elements all velocity shape functions are nonzero in all components. In other words, the function cannot distinguish between \(x\) and \(y\) velocity functions because there is no such distinction. It therefore puts the overall number of velocity into each of dofs_per_component[c]
, \(0\le c\le \text{dim}\). On the other hand, the number of pressure variables equals the number of shape functions that are nonzero in the dimth component.
Using this knowledge, we can get the number of velocity shape functions from any of the first dim
elements of dofs_per_component
, and then use this below to initialize the vector and matrix block sizes, as well as create output.
The next task is to allocate a sparsity pattern for the matrix that we will create. We use a compressed sparsity pattern like in the previous steps, but as system_matrix
is a block matrix we use the class BlockDynamicSparsityPattern
instead of just DynamicSparsityPattern
. This block sparsity pattern has four blocks in a \(2 \times 2\) pattern. The blocks' sizes depend on n_u
and n_p
, which hold the number of velocity and pressure variables. In the second step we have to instruct the block system to update its knowledge about the sizes of the blocks it manages; this happens with the dsp.collect_sizes ()
call.
We use the compressed block sparsity pattern in the same way as the nonblock version to create the sparsity pattern and then the system matrix:
Then we have to resize the solution and right hand side vectors in exactly the same way as the block compressed sparsity pattern:
Similarly, the function that assembles the linear system has mostly been discussed already in the introduction to this example. At its top, what happens are all the usual steps, with the addition that we do not only allocate quadrature and FEValues
objects for the cell terms, but also for face terms. After that, we define the usual abbreviations for variables, and the allocate space for the local matrix and right hand side contributions, and the array that holds the global numbers of the degrees of freedom local to the present cell.
The next step is to declare objects that represent the source term, pressure boundary value, and coefficient in the equation. In addition to these objects that represent continuous functions, we also need arrays to hold their values at the quadrature points of individual cells (or faces, for the boundary values). Note that in the case of the coefficient, the array has to be one of matrices.
Finally, we need a couple of extractors that we will use to get at the velocity and pressure components of vectorvalued shape functions. Their function and use is described in detail in the Handling vector valued problems report. Essentially, we will use them as subscripts on the FEValues objects below: the FEValues object describes all vector components of shape functions, while after subscription, it will only refer to the velocities (a set of dim
components starting at component zero) or the pressure (a scalar component located at position dim
):
With all this in place, we can go on with the loop over all cells. The body of this loop has been discussed in the introduction, and will not be commented any further here:
The final step in the loop over all cells is to transfer local contributions into the global matrix and right hand side vector. Note that we use exactly the same interface as in previous examples, although we now use block matrices and vectors instead of the regular ones. In other words, to the outside world, block objects have the same interface as matrices and vectors, but they additionally allow to access individual blocks.
The linear solvers and preconditioners we use in this example have been discussed in significant detail already in the introduction. We will therefore not discuss the rationale for our approach here any more, but rather only comment on some remaining implementational aspects.
As already outlined in the introduction, the solve function consists essentially of two steps. First, we have to form the first equation involving the Schur complement and solve for the pressure (component 1 of the solution). Then, we can reconstruct the velocities from the second equation (component 0 of the solution).
As a first step we declare references to all block components of the matrix, the right hand side and the solution vector that we will need.
Then, we will create corresponding LinearOperator objects and create the op_M_inv
operator:
This allows us to declare the Schur complement op_S
and the approximate Schur complement op_aS
:
We now create a preconditioner out of op_aS
that applies a fixed number of 30 (inexpensive) CG iterations:
Now on to the first equation. The right hand side of it is \(B^TM^{1}FG\), which is what we compute in the first few lines. We then solve the first equation with a CG solver and the preconditioner we just declared.
After we have the pressure, we can compute the velocity. The equation reads \(MU=BP+F\), and we solve it by first computing the right hand side, and then multiplying it with the object that represents the inverse of the mass matrix:
After we have dealt with the linear solver and preconditioners, we continue with the implementation of our main class. In particular, the next task is to compute the errors in our numerical solution, in both the pressures as well as velocities.
To compute errors in the solution, we have already introduced the VectorTools::integrate_difference
function in step7 and step11. However, there we only dealt with scalar solutions, whereas here we have a vectorvalued solution with components that even denote different quantities and may have different orders of convergence (this isn't the case here, by choice of the used finite elements, but is frequently the case in mixed finite element applications). What we therefore have to do is to `mask' the components that we are interested in. This is easily done: the VectorTools::integrate_difference
function takes as one of its arguments a pointer to a weight function (the parameter defaults to the null pointer, meaning unit weights). What we have to do is to pass a function object that equals one in the components we are interested in, and zero in the other ones. For example, to compute the pressure error, we should pass a function that represents the constant vector with a unit value in component dim
, whereas for the velocity the constant vector should be one in the first dim
components, and zero in the location of the pressure.
In deal.II, the ComponentSelectFunction
does exactly this: it wants to know how many vector components the function it is to represent should have (in our case this would be dim+1
, for the joint velocitypressure space) and which individual or range of components should be equal to one. We therefore define two such masks at the beginning of the function, following by an object representing the exact solution and a vector in which we will store the cellwise errors as computed by integrate_difference
:
As already discussed in step7, we have to realize that it is impossible to integrate the errors exactly. All we can do is approximate this integral using quadrature. This actually presents a slight twist here: if we naively chose an object of type QGauss<dim>(degree+1)
as one may be inclined to do (this is what we used for integrating the linear system), one realizes that the error is very small and does not follow the expected convergence curves at all. What is happening is that for the mixed finite elements used here, the Gauss points happen to be superconvergence points in which the pointwise error is much smaller (and converges with higher order) than anywhere else. These are therefore not particularly good points for integration. To avoid this problem, we simply use a trapezoidal rule and iterate it degree+2
times in each coordinate direction (again as explained in step7):
With this, we can then let the library compute the errors and output them to the screen:
The last interesting function is the one in which we generate graphical output. Note that all velocity components get the same solution name "u". Together with using DataComponentInterpretation::::component_is_part_of_vector this will cause DataOut<dim>::write_vtu() to generate a vector representation of the individual velocity components, see step22 or the Generating graphical output section of the Handling vector valued problems module for more information. Finally, it seems inappropriate for higher order elements to only show a single bilinear quadrilateral per cell in the graphical output. We therefore generate patches of size (degree+1)x(degree+1) to capture the full information content of the solution. See the step7 tutorial program for more information on this.
This is the final function of our main class. It's only job is to call the other functions in their natural order:
main
functionThe main function we stole from step6 instead of step4. It is almost equal to the one in step6 (apart from the changed class names, of course), the only exception is that we pass the degree of the finite element space to the constructor of the mixed Laplace problem (here, we use zeroth order elements).
If we run the program as is, we get this output for the \(32\times 32\) mesh we use (for a total of 1024 cells with 1024 pressure degrees of freedom since we use piecewise constants, and 2112 velocities because the RaviartThomas element defines one degree per freedom per face and there are \(1024 + 32 = 1056\) faces parallel to the \(x\)axis and the same number parallel to the \(y\)axis):
$ make run [ 66%] Built target step20 Scanning dependencies of target run [100%] Run step20 with Release configuration Number of active cells: 1024 Total number of cells: 1365 Number of degrees of freedom: 3136 (2112+1024) 24 CG Schur complement iterations to obtain convergence. Errors: e_p_L2 = 0.0445032, e_u_L2 = 0.010826 [100%] Built target run
The fact that the number of iterations is so small, of course, is due to the good (but expensive!) preconditioner we have developed. To get confidence in the solution, let us take a look at it. The following three images show (from left to right) the xvelocity, the yvelocity, and the pressure:
Let us start with the pressure: it is highest at the left and lowest at the right, so flow will be from left to right. In addition, though hardly visible in the graph, we have chosen the pressure field such that the flow leftright flow first channels towards the center and then outward again. Consequently, the xvelocity has to increase to get the flow through the narrow part, something that can easily be seen in the left image. The middle image represents inward flow in ydirection at the left end of the domain, and outward flow in ydirection at the right end of the domain.
As an additional remark, note how the xvelocity in the left image is only continuous in xdirection, whereas the yvelocity is continuous in ydirection. The flow fields are discontinuous in the other directions. This very obviously reflects the continuity properties of the RaviartThomas elements, which are, in fact, only in the space H(div) and not in the space \(H^1\). Finally, the pressure field is completely discontinuous, but that should not surprise given that we have chosen FE_DGQ(0)
as the finite element for that solution component.
The program offers two obvious places where playing and observing convergence is in order: the degree of the finite elements used (passed to the constructor of the MixedLaplaceProblem
class from main()
), and the refinement level (determined in MixedLaplaceProblem::make_grid_and_dofs
). What one can do is to change these values and observe the errors computed later on in the course of the program run.
If one does this, one finds the following pattern for the \(L_2\) error in the pressure variable:
Finite element order  

Refinement level  0  1  2 
0  1.45344  0.0831743  0.0235186 
1  0.715099  0.0245341  0.00293983 
2  0.356383  0.0063458  0.000367478 
3  0.178055  0.00159944  4.59349e05 
4  0.0890105  0.000400669  5.74184e06 
5  0.0445032  0.000100218  7.17799e07 
6  0.0222513  2.50576e05  9.0164e08 
\(O(h)\)  \(O(h^2)\)  \(O(h^3)\) 
The theoretically expected convergence orders are very nicely reflected by the experimentally observed ones indicated in the last row of the table.
One can make the same experiment with the \(L_2\) error in the velocity variables:
Finite element order  

Refinement level  0  1  2 
0  0.367423  0.127657  5.10388e14 
1  0.175891  0.0319142  9.04414e15 
2  0.0869402  0.00797856  1.23723e14 
3  0.0433435  0.00199464  1.86345e07 
4  0.0216559  0.00049866  2.72566e07 
5  0.010826  0.000124664  3.57141e07 
6  0.00541274  3.1166e05  4.46124e07 
\(O(h)\)  \(O(h^2)\)  \(O(h^3)\) 
The result concerning the convergence order is the same here.
Realistic flow computations for ground water or oil reservoir simulations will not use a constant permeability. Here's a first, rather simple way to change this situation: we use a permeability that decays very rapidly away from a central flowline until it hits a background value of 0.001. This is to mimic the behavior of fluids in sandstone: in most of the domain, the sandstone is homogeneous and, while permeable to fluids, not overly so; on the other stone, the stone has cracked, or faulted, along one line, and the fluids flow much easier along this large crack. Here is how we could implement something like this:
Remember that the function returns the inverse of the permeability tensor.
With a significantly higher mesh resolution, we can visualize this, here with x and yvelocity:
It is obvious how fluids flow essentially only along the middle line, and not anywhere else.
Another possibility would be to use a random permeability field. A simple way to achieve this would be to scatter a number of centers around the domain and then use a permeability field that is the sum of (negative) exponentials for each of these centers. Flow would then try to hop from one center of high permeability to the next one. This is an entirely unscientific attempt at describing a random medium, but one possibility to implement this behavior would look like this:
A piecewise constant interpolation of the diagonal elements of the inverse of this tensor (i.e., of normalized_permeability
) looks as follows:
With a permeability field like this, we would get xvelocities and pressures as follows:
We will use these permeability fields again in step21 and step43.
As mentioned in the introduction, the Schur complement solver used here is not the best one conceivable (nor is it intended to be a particularly good one). Better ones can be found in the literature and can be built using the same block matrix techniques that were introduced here. We pick up on this theme again in step22, where we first build a Schur complement solver for the Stokes equation as we did here, and then in the Improved Solvers section discuss better ways based on solving the system as a whole but preconditioning based on individual blocks. We will also come back to this in step43.