| 1 |
|
| 2 |
.. _demo_mixed_poisson: |
| 3 |
|
| 4 |
Mixed formulation for Poisson equation |
| 5 |
====================================== |
| 6 |
|
| 7 |
This demo is implemented in a single Python file, |
| 8 |
:download:`demo_mixed-poisson.py`, which contains both the variational |
| 9 |
forms and the solver. |
| 10 |
|
| 11 |
This demo illustrates how to solve Poisson equation using a mixed |
| 12 |
(two-field) formulation. In particular, it illustrates how to |
| 13 |
|
| 14 |
* Use mixed and non-continuous finite element spaces |
| 15 |
* Set essential boundary conditions for subspaces and H(div) spaces |
| 16 |
* Define a (vector-valued) expression using additional geometry information |
| 17 |
|
| 18 |
|
| 19 |
Equation and problem definition |
| 20 |
------------------------------- |
| 21 |
|
| 22 |
An alternative formulation of Poisson equation can be formulated by |
| 23 |
introducing an additional (vector) variable, namely the (negative) |
| 24 |
flux: :math:`\sigma = \nabla u`. The partial differential equations |
| 25 |
then read |
| 26 |
|
| 27 |
.. math:: |
| 28 |
\sigma - \nabla u &= 0 \quad {\rm in} \ \Omega, \\ |
| 29 |
\nabla \cdot \sigma &= - f \quad {\rm in} \ \Omega, |
| 30 |
|
| 31 |
with boundary conditions |
| 32 |
|
| 33 |
.. math:: |
| 34 |
u = u_0 \quad {\rm on} \ \Gamma_{D}, \\ |
| 35 |
\sigma \cdot n = g \quad {\rm on} \ \Gamma_{N}. |
| 36 |
|
| 37 |
The same equations arise in connection with flow in porous media, and |
| 38 |
are also referred to as Darcy flow. |
| 39 |
|
| 40 |
After multiplying by test functions :math:`\tau` and :math:`v`, |
| 41 |
integrating over the domain, and integrating the gradient term by |
| 42 |
parts, one obtains the following variational formulation: find |
| 43 |
:math:`\sigma \in \Sigma` and :math:`v \in V` satisfying |
| 44 |
|
| 45 |
.. math:: |
| 46 |
\int_{\Omega} (\sigma \cdot \tau + \nabla \cdot \tau \ u) \ {\rm d} x |
| 47 |
&= \int_{\Gamma} \tau \cdot n \ u \ {\rm d} s |
| 48 |
\quad \forall \ \tau \in \Sigma, \\ |
| 49 |
|
| 50 |
\int_{\Omega} \nabla \cdot \sigma v \ {\rm d} x |
| 51 |
&= - \int_{\Omega} f \ v \ {\rm d} x |
| 52 |
\quad \forall \ v \in V. |
| 53 |
|
| 54 |
Here :math:`n` denotes the outward pointing normal vector on the |
| 55 |
boundary. Looking at the variational form, we see that the boundary |
| 56 |
condition for the flux (:math:`\sigma \cdot n = g`) is now an |
| 57 |
essential boundary condition (which should be enforced in the function |
| 58 |
space), while the other boundary condition (:math:`u = u_0`) is a |
| 59 |
natural boundary condition (which should be applied to the variational |
| 60 |
form). Inserting the boundary conditions, this variational problem can |
| 61 |
be phrased in the general form: find :math:`(\sigma, u) \in \Sigma_g |
| 62 |
\times V` such that |
| 63 |
|
| 64 |
.. math:: |
| 65 |
|
| 66 |
a((\sigma, u), (\tau, v)) = L((\tau, v)) |
| 67 |
\quad \forall \ (\tau, v) \in \Sigma_0 \times V |
| 68 |
|
| 69 |
where the variational forms :math:`a` and :math:`L` are defined as |
| 70 |
|
| 71 |
.. math:: |
| 72 |
|
| 73 |
a((\sigma, u), (\tau, v)) &= |
| 74 |
\int_{\Omega} \sigma \cdot \tau + \nabla \cdot \tau \ u |
| 75 |
+ \nabla \cdot \sigma \ v \ {\rm d} x \\ |
| 76 |
L((\tau, v)) &= - \int_{\Omega} f v \ {\rm d} x |
| 77 |
+ \int_{\Gamma_D} u_0 \tau \cdot n \ {\rm d} s |
| 78 |
|
| 79 |
and :math:`\Sigma_g = \{ \tau \in H({\rm div}) \text{ such that } \tau \cdot n|_{\Gamma_N} = g \}` |
| 80 |
and :math:`V = L^2(\Omega)`. |
| 81 |
|
| 82 |
To discretize the above formulation, two discrete function spaces |
| 83 |
:math:`\Sigma_h \subset \Sigma` and :math:`V_h \subset V` are needed |
| 84 |
to form a mixed function space :math:`\Sigma_h \times V_h`. A stable |
| 85 |
choice of finite element spaces is to let :math:`\Sigma_h` be the |
| 86 |
Brezzi-Douglas-Marini elements of polynomial order :math:`k` and let |
| 87 |
:math:`V_h` be discontinuous elements of polynomial order :math:`k-1`. |
| 88 |
|
| 89 |
We will use the same definitions of functions and boundaries as in the |
| 90 |
demo for Poisson's equation. These are: |
| 91 |
|
| 92 |
* :math:`\Omega = [0,1] \times [0,1]` (a unit square) |
| 93 |
* :math:`\Gamma_{D} = \{(0, y) \cup (1, y) \in \partial \Omega\}` |
| 94 |
* :math:`\Gamma_{N} = \{(x, 0) \cup (x, 1) \in \partial \Omega\}` |
| 95 |
* :math:`u_0 = 0` |
| 96 |
* :math:`g = \sin(5x)` (flux) |
| 97 |
* :math:`f = 10\exp(-((x - 0.5)^2 + (y - 0.5)^2) / 0.02)` (source term) |
| 98 |
|
| 99 |
With the above input the solution for :math:`u` and :math:`\sigma` will look as |
| 100 |
follows: |
| 101 |
|
| 102 |
.. image:: mixed-poisson_u.png |
| 103 |
:scale: 75 % |
| 104 |
:align: center |
| 105 |
|
| 106 |
.. image:: mixed-poisson_sigma.png |
| 107 |
:scale: 75 % |
| 108 |
:align: center |
| 109 |
|
| 110 |
|
| 111 |
Implementation |
| 112 |
-------------- |
| 113 |
|
| 114 |
This demo is implemented in the :download:`demo_mixed-poisson.py` |
| 115 |
file. |
| 116 |
|
| 117 |
First, the required modules are imported:: |
| 118 |
|
| 119 |
from dolfin import * |
| 120 |
import matplotlib.pyplot as plt |
| 121 |
|
| 122 |
Then, we need to create a :py:class:`Mesh <dolfin.cpp.Mesh>` covering |
| 123 |
the unit square. In this example, we will let the mesh consist of 32 x |
| 124 |
32 squares with each square divided into two triangles:: |
| 125 |
|
| 126 |
# Create mesh |
| 127 |
mesh = UnitSquareMesh(32, 32) |
| 128 |
|
| 129 |
.. index:: |
| 130 |
pair: FunctionSpace; Brezzi-Douglas-Marini |
| 131 |
pair: FunctionSpace; Discontinous Lagrange |
| 132 |
|
| 133 |
Next, we need to build the function space. :: |
| 134 |
|
| 135 |
# Define finite elements spaces and build mixed space |
| 136 |
BDM = FiniteElement("BDM", mesh.ufl_cell(), 1) |
| 137 |
DG = FiniteElement("DG", mesh.ufl_cell(), 0) |
| 138 |
W = FunctionSpace(mesh, BDM * DG) |
| 139 |
|
| 140 |
The second argument to :py:class:`FunctionSpace |
| 141 |
<dolfin.functions.functionspace.FunctionSpace>` specifies underlying |
| 142 |
finite element, here mixed element obtained by ``*`` operator. |
| 143 |
|
| 144 |
.. math:: |
| 145 |
|
| 146 |
W = \{ (\tau, v) \ \text{such that} \ \tau \in BDM, v \in DG \}. |
| 147 |
|
| 148 |
Next, we need to specify the trial functions (the unknowns) and the |
| 149 |
test functions on this space. This can be done as follows:: |
| 150 |
|
| 151 |
# Define trial and test functions |
| 152 |
(sigma, u) = TrialFunctions(W) |
| 153 |
(tau, v) = TestFunctions(W) |
| 154 |
|
| 155 |
In order to define the variational form, it only remains to define the |
| 156 |
source function :math:`f`. This is done just as for the :ref:`Poisson |
| 157 |
demo <demo_poisson_equation>`:: |
| 158 |
|
| 159 |
# Define source function |
| 160 |
f = Expression("10*exp(-(pow(x[0] - 0.5, 2) + pow(x[1] - 0.5, 2)) / 0.02)", degree=2) |
| 161 |
|
| 162 |
We are now ready to define the variational forms a and L. Since, |
| 163 |
:math:`u_0 = 0` in this example, the boundary term on the right-hand |
| 164 |
side vanishes. :: |
| 165 |
|
| 166 |
# Define variational form |
| 167 |
a = (dot(sigma, tau) + div(tau)*u + div(sigma)*v)*dx |
| 168 |
L = - f*v*dx |
| 169 |
|
| 170 |
It only remains to prescribe the boundary condition for the |
| 171 |
flux. Essential boundary conditions are specified through the class |
| 172 |
:py:class:`DirichletBC <dolfin.fem.bcs.DirichletBC>` which takes three |
| 173 |
arguments: the function space the boundary condition is supposed to be |
| 174 |
applied to, the data for the boundary condition, and the relevant part |
| 175 |
of the boundary. |
| 176 |
|
| 177 |
We want to apply the boundary condition to the first subspace of the |
| 178 |
mixed space. Subspaces of a mixed :py:class:`FunctionSpace |
| 179 |
<dolfin.functions.functionspace.FunctionSpace>` can be accessed |
| 180 |
by the method :py:func:`sub |
| 181 |
<dolfin.functions.functionspace.FunctionSpace.sub>`. In our case, |
| 182 |
this reads ``W.sub(0)``. (Do *not* use the separate space ``BDM`` as |
| 183 |
this would mess up the numbering.) |
| 184 |
|
| 185 |
Next, we need to construct the data for the boundary condition. An |
| 186 |
essential boundary condition is handled by replacing degrees of |
| 187 |
freedom by the degrees of freedom evaluated at the given data. The |
| 188 |
:math:`BDM` finite element spaces are vector-valued spaces and hence |
| 189 |
the degrees of freedom act on vector-valued objects. The effect is |
| 190 |
that the user is required to construct a :math:`G` such that :math:`G |
| 191 |
\cdot n = g`. Such a :math:`G` can be constructed by letting :math:`G |
| 192 |
= g n`. In particular, it can be created by subclassing the |
| 193 |
:py:class:`Expression <dolfin.functions.expression.Expression>` |
| 194 |
class. Overloading the ``eval_cell`` method (instead of the usual |
| 195 |
``eval``) allows us to extract more geometry information such as the |
| 196 |
facet normals. Since this is a vector-valued expression, we also need |
| 197 |
to overload the ``value_shape`` method. |
| 198 |
|
| 199 |
.. index:: |
| 200 |
single: Expression; (in Mixed Poisson demo) |
| 201 |
|
| 202 |
:: |
| 203 |
|
| 204 |
# Define function G such that G \cdot n = g |
| 205 |
class BoundarySource(UserExpression): |
| 206 |
def __init__(self, mesh, **kwargs): |
| 207 |
self.mesh = mesh |
| 208 |
super().__init__(**kwargs) |
| 209 |
def eval_cell(self, values, x, ufc_cell): |
| 210 |
cell = Cell(self.mesh, ufc_cell.index) |
| 211 |
n = cell.normal(ufc_cell.local_facet) |
| 212 |
g = sin(5*x[0]) |
| 213 |
values[0] = g*n[0] |
| 214 |
values[1] = g*n[1] |
| 215 |
def value_shape(self): |
| 216 |
return (2,) |
| 217 |
|
| 218 |
G = BoundarySource(mesh, degree=2) |
| 219 |
|
| 220 |
Specifying the relevant part of the boundary can be done as for the |
| 221 |
Poisson demo (but now the top and bottom of the unit square is the |
| 222 |
essential boundary): :: |
| 223 |
|
| 224 |
# Define essential boundary |
| 225 |
def boundary(x): |
| 226 |
return x[1] < DOLFIN_EPS or x[1] > 1.0 - DOLFIN_EPS |
| 227 |
|
| 228 |
Now, all the pieces are in place for the construction of the essential |
| 229 |
boundary condition: :: |
| 230 |
|
| 231 |
bc = DirichletBC(W.sub(0), G, boundary) |
| 232 |
|
| 233 |
To compute the solution we use the bilinear and linear forms, and the |
| 234 |
boundary condition, but we also need to create a :py:class:`Function |
| 235 |
<dolfin.functions.function.Function>` to store the solution(s). The |
| 236 |
(full) solution will be stored in the ``w``, which we initialise using |
| 237 |
the :py:class:`FunctionSpace |
| 238 |
<dolfin.functions.functionspace.FunctionSpace>` ``W``. The actual |
| 239 |
computation is performed by calling :py:func:`solve |
| 240 |
<dolfin.fem.solving.solve>`. The separate components ``sigma`` and |
| 241 |
``u`` of the solution can be extracted by calling the :py:func:`split |
| 242 |
<dolfin.functions.function.Function.split>` function. Finally, we plot |
| 243 |
the solutions to examine the result. :: |
| 244 |
|
| 245 |
# Compute solution |
| 246 |
w = Function(W) |
| 247 |
solve(a == L, w, bc) |
| 248 |
(sigma, u) = w.split() |
| 249 |
|
| 250 |
# Plot sigma and u |
| 251 |
plt.figure() |
| 252 |
plot(sigma) |
| 253 |
|
| 254 |
plt.figure() |
| 255 |
plot(u) |
| 256 |
|
| 257 |
plt.show() |