---
jupytext:
  text_representation:
    extension: .md
    format_name: myst
    format_version: 0.13
    jupytext_version: 1.16.4
kernelspec:
  display_name: Python 3 (ipykernel)
  language: python
  name: python3
---

# Biharmonic equation

Authors: Julius Herb, Ottar Hellan and Jørgen S. Dokken

```{admonition} Download sources
:class: download
* {download}`Python script <./demo_biharmonic.py>`
* {download}`Jupyter notebook <./demo_biharmonic.ipynb>`
```
This demo illustrates how to:

- Solve a linear partial differential equation
- Use a discontinuous Galerkin method
- Solve a fourth-order differential equation
- Use the method of manufactured solutions

## Equation and problem definition

### Strong formulation

The biharmonic equation is a fourth-order elliptic equation.
On the domain $\Omega \subset \mathbb{R}^{d}$, $1 \le d \le 3$, it reads

$$
\nabla^{4} u = f \quad {\rm in} \ \Omega,
$$

where $\nabla^{4} \equiv \nabla^{2} \nabla^{2}=\Delta\Delta$ is the
biharmonic operator and $f$ is a prescribed source term.
To formulate a complete boundary value problem, the biharmonic equation
must be complemented by suitable boundary conditions.

 ### Choice of boundary conditions
As we have a fourth order partial differential equation, we are required
to supply two boundary conditions.
There are two common sets of conditions that people use for the
biharmonic equation, namely the *clamped* condition and the
*simply supported* condition, see for instance {cite:t}`Gander2017bcs`.

$$
\begin{aligned}
u =0 &\qquad \frac{\partial u}{\partial n} = 0
\text{ on } \partial\Omega,\\
u =0 &\qquad \Delta u = 0 \text{ on } \partial \Omega.
\end{aligned}
$$

In this demo we will consider the clamped boundary conditions

$$
\begin{aligned}
u &= g_D \text{ on } \partial\Omega,\\
\frac{\partial u}{\partial n} &= g_N \text{ on } \partial\Omega,\\
\end{aligned}
$$

as the simply supported boundary conditions reduces the system into two
sequential Poisson problems, named the Ciarlet-Raviart formulation
{cite}`ciarlet1974mixed`.

+++

### Weak formulation

Multiplying the biharmonic equation by a test function and integrating
by parts twice leads to a problem of second-order derivatives, which
would require $H^{2}$ conforming (roughly $C^{1}$ continuous) basis
functions. To solve the biharmonic equation using Lagrange finite element
basis functions, the biharmonic equation can be split into two second-
order equations (see the [Mixed Poisson demo](./demo_mixed-poisson)
for an example of a mixed method), or a variational
formulation can be constructed that imposes weak continuity of normal
derivatives between finite element cells. This demo uses a discontinuous
Galerkin approach to impose continuity of the normal derivative weakly,
see for instance {cite:t}`babuska1973penalty` or
{cite:t}`Georgoulis2009biharmonic`.

In this demo, we consider equation 3.20-3.21 from
{cite:t}`Georgoulis2009biharmonic`, a $\mathcal{C}^0$-interior penalty
method. However, instead of using a broken (discontinuous)
finite element space for the unknown $u$, we use a continuous space,
which simplifies the formulation to a weak formulation of the
biharmonic problem reads: find $u \in V_{g_D}$ such that

$$
a(u,v)=L(v) \quad \forall \ v \in V,
$$

where the bilinear form is

$$
\begin{aligned}
a(u, v) &=
\sum_{K \in \mathcal{T}} \int_{K} \Delta u \Delta v ~{\rm d}x \\
&+\sum_{E \in \mathcal{E}_h^{\rm int}}\int_{E}\left(
- \left<\Delta u \right>[\!\![ \nabla v ]\!\!]
- [\!\![ \nabla u ]\!\!] \left<\nabla^{2} v \right>
+ \frac{\beta}{h_E} [\!\![ \nabla u ]\!\!] [\!\![ \nabla v ]\!\!]
\right)~{\rm d}s\\
&+\sum_{E \in \mathcal{E}_h^{\rm ext}}\int_{E}\left(
- \Delta u  \nabla v \cdot n - \Delta v \nabla u \cdot n
+ \frac{\beta}{h_E} \nabla u \cdot n \nabla v \cdot n
\right)~{\rm d}s
\end{aligned}
$$

and the linear form is

$$
L(v) = \int_{\Omega} fv ~{\rm d}x
+\sum_{E \in \mathcal{E}_h^{\rm ext}}\int_{E}\left(
-g_N \Delta v + \frac{\beta}{h_E} g_N \nabla v \cdot n
\right) ~{\rm d}s.
$$

Furthermore, $\left< u \right> = \frac{1}{2} (u_{+} + u_{-})$,
$[\!\![ w ]\!\!]  = w_{+} \cdot n_{+} + w_{-} \cdot n_{-}$,
$\beta \ge 0$ is a penalty parameter and
$h_E$ is a measure of the cell size. We use the penalty parameter
described in {cite:t}`Bringmann2024penalty`.

where $K$ is an element of the mesh, while $\mathcal{E}_h^{\rm int}$
is the collection of all interior facets, while
$\mathcal{E}_h^{\rm ext}$ is the set of exterior facets.

```{note}
Note that the Dirichlet condition for $u$ will be enforced strongly
in this example.
```

## Implementation

We follow the example of {cite:t}`Georgoulis2009biharmonic` and use the
method of manufactured solutions to construct a $f$, $g_D$, $g_N$ that
satisfies

$$
u(x, y) = \sin (2\pi x) \sin (2 \pi y) \text{ in } [0, 1] \times [0, 1].
$$

We first import the modules and functions that the program uses:

```python
from pathlib import Path

from mpi4py import MPI

import numpy as np

import ufl
from dolfinx import default_real_type, default_scalar_type, fem, has_adios2, io, mesh, plot
from dolfinx.fem.petsc import LinearProblem
from dolfinx.mesh import CellType, GhostMode

if np.issubdtype(default_real_type, np.float32):  # type: ignore
    print("float32 not yet supported for this demo.")
    exit(0)
```

We begin by using {py:func}`create_rectangle
<dolfinx.mesh.create_rectangle>` to create a rectangular
{py:class}`Mesh <dolfinx.mesh.Mesh>` of the domain, and creating a
finite element {py:class}`FunctionSpace <dolfinx.fem.FunctionSpace>`
$V$ on the mesh.

```python
N = 32
msh = mesh.create_unit_square(
    MPI.COMM_WORLD,
    N,
    N,
    cell_type=CellType.triangle,
    ghost_mode=GhostMode.shared_facet,
)
```

As noted in {cite:t}`Georgoulis2009biharmonic` second order Lagrange
elements yield sub-optimal convergence, and therefore we choose to use
third order elements

```python
degree = 3
V = fem.functionspace(msh, ("Lagrange", degree))
```

The second argument to {py:func}`functionspace
<dolfinx.fem.functionspace>` is a tuple consisting of `(family,
degree)`, where `family` is the finite element family, and `degree`
specifies the polynomial degree. in this case `V` consists of
third-order, continuous Lagrange finite element functions.
For further details of how one can specify
finite elements as tuples, see {py:class}`ElementMetaData
<dolfinx.fem.ElementMetaData>`.

Next, we locate the mesh facets that lie on the boundary
$\Gamma_D = \partial\Omega$.
We do this using using {py:func}`exterior_facet_indices
<dolfinx.mesh.exterior_facet_indices>` which returns all mesh boundary
facets (Note: if we are only interested in a subset of those, consider
{py:func}`locate_entities_boundary
<dolfinx.mesh.locate_entities_boundary>`).

```python
tdim = msh.topology.dim
msh.topology.create_connectivity(tdim - 1, tdim)
facets = mesh.exterior_facet_indices(msh.topology)
```

We now find the degrees-of-freedom that are associated with the
boundary facets using {py:func}`locate_dofs_topological
<dolfinx.fem.locate_dofs_topological>`

```python
fdim = tdim - 1
dofs = fem.locate_dofs_topological(V=V, entity_dim=fdim, entities=facets)
```

and use {py:func}`dirichletbc <dolfinx.fem.dirichletbc>` to create a
{py:class}`DirichletBC <dolfinx.fem.DirichletBC>`
class that represents the boundary condition.

+++

We define the manufactured solution and interpolate it into the function
space of our unknown to apply it as a strong boundary condition

```python
x = ufl.SpatialCoordinate(msh)
u_ex = ufl.sin(2 * ufl.pi * x[0]) ** 2 * ufl.sin(2 * ufl.pi * x[1]) ** 2
g_D_expr = fem.Expression(u_ex, V.element.interpolation_points)
g_D = fem.Function(V)
g_D.interpolate(g_D_expr)
bc = fem.dirichletbc(value=g_D, dofs=dofs)
```

Next, we express the variational problem using UFL.

First, the penalty parameter $\beta$ is defined. In addition, we define
a variable `h` for the cell diameter $h_E$, a variable `n`for the
outward-facing normal vector $n$ and a variable `h_avg` for the
average size of cells sharing a facet
$\left< h \right> = \frac{1}{2} (h_{+} + h_{-})$. Here, the UFL syntax
`('+')` and `('-')` restricts a function to the `('+')` and `('-')`
sides of a facet.

```python
a = fem.Constant(msh, default_scalar_type(4.0))
vol = ufl.CellVolume(msh)
h = ufl.CellDiameter(msh)
beta = 3.0 * a * degree * (degree - 1) / 8.0 * h("+") ** 2 * ufl.avg(1 / vol)
beta_boundary = 3.0 * a * degree * (degree - 1) * h**2 / vol
n = ufl.FacetNormal(msh)
h_avg = (h("+") + h("-")) / 2.0
```

After that, we can define the variational problem consisting of the
bilinear form $a$ and the linear form $L$. We use {py:mod}`ufl` to derive
the manufactured $f$ and $g_N$.
Note that with `dS`, integration is carried out over all the interior
facets $\mathcal{E}_h^{\rm int}$, whereas with `ds` it would be only
the facets on the boundary of the domain, i.e. $\partial\Omega$.
The jump operator
$[\!\![ w ]\!\!] = w_{+} \cdot n_{+} + w_{-} \cdot n_{-}$ w.r.t. the
outward-facing normal vector $n$ is in UFL available as `jump(w, n)`.

```python
# Define variational problem
u = ufl.TrialFunction(V)
v = ufl.TestFunction(V)
f = ufl.div(ufl.grad(ufl.div(ufl.grad(u_ex))))
g_N = ufl.dot(ufl.grad(u_ex), n)

a = (
    ufl.inner(ufl.div(ufl.grad(u)), ufl.div(ufl.grad(v))) * ufl.dx
    - ufl.inner(ufl.avg(ufl.div(ufl.grad(u))), ufl.jump(ufl.grad(v), n)) * ufl.dS
    - ufl.inner(ufl.jump(ufl.grad(u), n), ufl.avg(ufl.div(ufl.grad(v)))) * ufl.dS
    + beta / h_avg * ufl.inner(ufl.jump(ufl.grad(u), n), ufl.jump(ufl.grad(v), n)) * ufl.dS
    - ufl.inner(ufl.div(ufl.grad(u)), ufl.dot(ufl.grad(v), n)) * ufl.ds
    - ufl.inner(ufl.dot(ufl.grad(u), n), ufl.div(ufl.grad(v))) * ufl.ds
    + beta_boundary / h * ufl.inner(ufl.dot(ufl.grad(u), n), ufl.dot(ufl.grad(v), n)) * ufl.ds
)
L = (
    ufl.inner(f, v) * ufl.dx
    - ufl.inner(g_N, ufl.div(ufl.grad(v))) * ufl.ds
    + beta_boundary / h * ufl.inner(g_N, ufl.dot(ufl.grad(v), n)) * ufl.ds
)
```

We create a {py:class}`LinearProblem <dolfinx.fem.petsc.LinearProblem>`
object that brings together the variational problem, the Dirichlet
boundary condition, and which specifies the linear solver. In this
case we use a direct (LU) solver. The {py:func}`solve
<dolfinx.fem.petsc.LinearProblem.solve>` will compute a solution.

```python
problem = LinearProblem(
    a,
    L,
    bcs=[bc],
    petsc_options_prefix="demo_biharmonic_",
    petsc_options={
        "ksp_type": "preonly",
        "pc_type": "lu",
    },
)
uh = problem.solve()

assert isinstance(uh, fem.Function)
assert problem.solver.getConvergedReason() > 0
```

We compute the relative $L^2$-error between the computed
and exact solution:

```python
error = fem.form(ufl.inner(uh - g_D, uh - g_D) * ufl.dx)
local_error = fem.assemble_scalar(error)
glob_error = error.mesh.comm.allreduce(local_error, op=MPI.SUM)
ref_scale = fem.form(ufl.inner(u_ex, u_ex) * ufl.dx)
local_scale = fem.assemble_scalar(ref_scale)
scale = ref_scale.mesh.comm.allreduce(local_scale, op=MPI.SUM)
print("||u_h-u_ex||_{L^2}^2/||u_ex||_L^2^2: ", f"{glob_error / scale:.5e}")
assert glob_error.real / scale.real < 1e-6
```

The solution can be written to a VTX-file using {py:class}`VTXWriter
<dolfinx.io.VTXWriter>` which can be opened with ParaView

```python
if has_adios2:
    out_folder = Path("out_biharmonic")
    out_folder.mkdir(parents=True, exist_ok=True)
    with io.VTXWriter(msh.comm, out_folder / "biharmonic.bp", [uh]) as file:
        file.write(0.0)
```

and displayed using [pyvista](https://docs.pyvista.org/).

```python
try:
    import pyvista

    cells, types, x = plot.vtk_mesh(V)
    grid = pyvista.UnstructuredGrid(cells, types, x)
    grid.point_data["u"] = uh.x.array.real
    grid.set_active_scalars("u")
    plotter = pyvista.Plotter()
    plotter.add_mesh(grid, show_edges=True)
    warped = grid.warp_by_scalar()
    plotter.add_mesh(warped)
    if pyvista.OFF_SCREEN:
        plotter.screenshot(out_folder / "uh_biharmonic.png")
    else:
        plotter.show()
except ModuleNotFoundError:
    print("'pyvista' is required to visualise the solution")
    print("Install 'pyvista' with pip: 'python3 -m pip install pyvista'")
```

## References
```{bibliography}
```
