This page was generated from unit-2.9-fourthorder/fourthorder.ipynb.

2.9 Fourth order equations

We consider the Kirchhoff plate equation: Find \(w \in H^2\), such that

\[\int \nabla^2 w : \nabla^2 v = \int f v\]

A conforming method requires \(C^1\)-continuous finite elements, which are complicated to construct. We consider the following two alternatives, which get along with simpler finite element spaces:

Hybridized \(C^0\)-continuous interior penalty method:

A simple way out is to use continuous elements, and treat the missing \(C^1\)-continuity by a Discontinuous Galerkin method. A DG formulation is

\[\sum_T \nabla^2 w : \nabla^2 v - \int_{E} \{\nabla^2 w\}_{nn} \, [\partial_n v] - \int_{E} \{\nabla^2 v\}_{nn} \, [\partial_n w] + \alpha \int_E [\partial_n w] [\partial_n v]\]

[Baker 77, Brenner Gudi Sung, 2010]

We consider its hybrid DG version, where the normal derivative is a new, facet-based variable:

\[\sum_T \nabla^2 w : \nabla^2 v - \int_{\partial T} (\nabla^2 w)_{nn} \, (\partial_n v - \widehat{v_n}) - \int_{\partial T} (\nabla^2 v)_{nn} \, (\partial_n w - \widehat{w_n}) + \alpha \int_E (\partial_n v - \widehat{v_n}) (\partial_n w - \widehat{w_n})\]

The facet variable is the normal derivative \(n_E \cdot \nabla w\), which is oriented along the arbitrarily chosen edge normal vector. We cannot use the FacetFESpace since it does not have the orientation. We use the NormalFacetFESpace, which provides vector-valued functions in normal direction on the facets, like the normal traces of an \(H(div)\) space.

[1]:
from ngsolve import *
from ngsolve.webgui import Draw
mesh = Mesh (unit_square.GenerateMesh(maxh=0.1))
[2]:
order = 3

V1 = H1(mesh, order=order)
V2 = NormalFacetFESpace(mesh, order=order-1)
V = V1*V2

(w,what), (v,vhat) = V.TnT()

Some proxy functions and GridFunctions provide additional differential operators. We can get them via the Operator function. w.Operator("hesse") provides the Hessian, a matrix-valued function. Note that we can use InnerProduct(.,.) for \(\nabla^2 w : \nabla^2 v\):

[3]:
n = specialcf.normal(mesh.dim)
h = specialcf.mesh_size

def jumpdn(v,vhat):
    return n*(grad(v)-vhat)
def hesse(v):
    return v.Operator("hesse")
def hessenn(v):
    return InnerProduct(n, hesse(v)*n)

dS = dx(element_boundary=True)
lhs = InnerProduct (hesse(w), hesse(v)) * dx \
     - hessenn(w) * jumpdn(v,vhat) * dS \
     - hessenn(v) * jumpdn(w,what) * dS \
     + 3*order*order/h * jumpdn(w,what) * jumpdn(v,vhat) * dS

We solve the clamped plate. The boundary conditions \(w = 0\) and \(\partial_n w = 0\) are essential boundary conditions for the two components:

[4]:
gf = Solve(lhs == 1*v*dx, w[BND(".*")] == 0, what[BND(".*")] == 0)
gfw = gf[w]

Draw (gfw, mesh, "disp_DG")
Draw (grad (gfw), mesh, "grad")
Draw (hesse (gfw), mesh, "hesse");
Warning in Webgui.Draw: function has more than 3 components, drawing Norm(cf) instead
Warning in Webgui.Draw: function has more than 3 components, drawing Norm(cf) instead

The Hellan-Herrmann-Johnson Mixed Method:

[Hellan 67, Herrmann 67, Johnson 73, Arnold+Brezzi 85, Comodi 89]

We introduce the bending moments as a new, matrix-valued variable \(\sigma = \nabla^2 w\). Then we can write the plate equation as a saddle point problem:

\[\begin{split}\DeclareMathOperator{\Div}{div} \begin{array}{ccccll} \int \sigma : \tau & - & \int \tau : \nabla^2 w & = & 0 & \forall \, \tau \\ -\int \sigma : \nabla^2 v & & & = & -\int f v & \forall \, v \end{array}\end{split}\]

Integration by parts in the upper-right and lower-left terms:

\[\begin{split}\DeclareMathOperator{\Div}{div} \begin{array}{ccccll} \int \sigma : \tau & + & < \Div \tau , \nabla w > & = & 0 & \forall \, \tau \\ < \Div \sigma , \nabla v > & & & = & -\int f v & \forall \, v \end{array}\end{split}\]

The Hellan-Herrmann-Johnson method uses finite elements \(\sigma_h\) which have a continuous normal-normal component. Then, the term \(<\Div \sigma, \nabla v>\) must be understood in distributional form:

\[\sum_T \int_T \Div \sigma \, \nabla v - \sum_E \int_E [\sigma_{nt}] (\nabla v)_t\]

Since \(\sigma_{nn}\) is continuous, the jump occurs only in the tangential component of \(\sigma_n\). Luckily, it hits the tangential component of \(\nabla v\), which is single-valued for \(H^1\) finite elements. Thus, we can rewrite the term element by element:

\[\sum_T \int_T \Div \sigma \, \nabla v - \int_{\partial T} \sigma_{nt}\nabla_t v\]

The Hellan-Herrmann-Johnson method does not require access to the neighbour element as in DG methods. In [Pechstein-Schöberl 11] the function space \(H(\Div \Div)\) was introduced. The HHJ finite elements are (nearly) conforming for this space. A basis was also given in [PS11].

For the clamped plate, \(w = 0\) is an essential boundary condition for the displacement. The condition \(\partial_n w = 0\) is a natural boundary condition now.

[5]:
order = 3

V = HDivDiv(mesh, order=order-1)
Q = H1(mesh, order=order)
X = V*Q

print ("ndof-V:", V.ndof, ", ndof-Q:", Q.ndof)

(sigma, w), (tau, v) = X.TnT()

n = specialcf.normal(mesh.dim)

def tang(u): return u-(u*n)*n

lhs = (InnerProduct (sigma, tau) + div(sigma)*grad(v) + div(tau)*grad(w))*dx \
      + (-(sigma*n) * tang(grad(v)) - (tau*n)*tang(grad(w)))*dx(element_boundary=True)

gfu = Solve(lhs == -1*v*dx, w[BND(".*")] == 0)

Draw (gfu[sigma], mesh, name="sigma")
Draw (gfu[w], mesh, name="disp");
ndof-V: 3165 , ndof-Q: 1096
Warning in Webgui.Draw: function has more than 3 components, drawing Norm(cf) instead
Warning in Webgui.Draw: function has more than 3 components, drawing Norm(cf) instead

Both methods give nearly the same deflection in the center of the plate:

[6]:
mip = mesh(0.5, 0.5)
print ("w, C0-IP:", gfw(mip))
print ("w, HHJ:  ", gfu[w](mip))
w, C0-IP: 0.0012647281576770196
w, HHJ:   0.0012653014281189066