Semi-discrete form for diffusion

6.2. Semi-discrete form for diffusion#

In this section we go back to the Poisson equation one last time. We have seen this PDE before in 1D and 2D, but only for steady-state (time-independent) problems. Here we make one more extension by allowing time-dependent behavior to arise. We will see one common way to treat time in FEM problems, by arriving at a semi-discrete system of equations we can solve numerically.

This page also allows you to switch between tensor notation and index notation. Try it out and see how different parts of the formulation become easier to write in one or the other notation.

Strong form equation#

Tensor notation

As always, we start the formulation with the PDE in its strong form:

(6.4)#

- \nabla \cdot q - ρ c \dot{u} + f = 0, with q = - κ \cdot \nabla u

where $u$ is our solution field of interest: in heat conduction problems it would represent temperature; in mass diffusion it would be a concentration; in Darcy flow it would be the hydraulic head. Furthermore, $q$ is a gradient-driven flux vector that depends on a diffusivity matrix $κ$ . The new term here is $ρ c \dot{u}$ , with $ρ c$ being a capacity parameter and $\dot{u}$ the first time derivative of our solution, i.e. a velocity field.

The problem is completed by Dirichlet, Neumann and initial boundary conditions:

(6.5)#

\begin{array}{r} \begin{aligned} u = g & on Γ_{D} \\ - q \cdot n = h & on Γ_{N} \\ u = u_{0} & at t = 0 \end{aligned} \end{array}

where $g$ is a prescribed value for the solution field and $n$ is the normal to the surface where Neumann boundary conditions are applied. Since we are solving a time-dependent problem, the third condition fixes an initial field $u_{0}$ from which we will predict into the future.

Index notation

As always, we start the formulation with the PDE in its strong form:

(6.6)#

- q_{i, i} - ρ c \dot{u} + f = 0, with q_{i} = - κ_{i j} u_{, j}

where $u$ is our solution field of interest: in heat conduction problems it would represent temperature; in mass diffusion it would be a concentration; in Darcy flow it would be the hydraulic head. Furthermore, $q_{i}$ is a gradient-driven flux vector that depends on a diffusivity matrix $κ_{i j}$ . The new term here is $ρ c \dot{u}$ , with $ρ c$ being a capacity parameter and $\dot{u}$ the first time derivative of our solution, i.e. a velocity field.

The problem is completed by Dirichlet, Neumann and initial boundary conditions:

(6.7)#

\begin{array}{r} \begin{aligned} u = g & on Γ_{D} \\ - q_{i} n_{i} = h & on Γ_{N} \\ u = u_{0} & at t = 0 \end{aligned} \end{array}

where $g$ is a prescribed value for the solution field and $n_{i}$ is the normal to the surface where Neumann boundary conditions are applied. Since we are solving a time-dependent problem, the third condition fixes an initial field $u_{0}$ from which we will predict into the future.

Weak form#

Tensor notation

To get to the weak form we first multiply the PDE by a weight function $w$ and integrate over the domain $Ω$ :

(6.8)#

- \int_{Ω} w (\nabla \cdot q) d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω = 0

where $w$ for this problem is always a scalar field, irrespective of the dimensionality of $Ω$ . The next step is to eliminate the derivative $\nabla \cdot q$ using the divergence theorem and integration by parts. Here we can directly use the result we already derived in Eq. (1.32):

(6.9)#

\int_{Ω} \nabla w \cdot q d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω - \int_{Γ_{N}} w q \cdot n d Γ = 0

and we see that the derivative has now been transferred to $w$ . This also causes an integral over $Γ$ to appear. From the requirement that $w = 0$ at $Γ_{D}$ , only the integral over $Γ_{N}$ remains. Finally, note that we did not touch the time derivative $\dot{u}$ , which means we are effectively treating $u$ and $\dot{u}$ as two independent fields.

As a final step, we substitute the Neumann boundary condition of Eq. (6.5) and the constitutive relation $q = - κ \cdot \nabla u$ to get to the final weak form:

(6.10)#

- \int_{Ω} \nabla w \cdot κ \nabla u d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω + \int_{Γ_{h}} w h d Γ = 0

which is valid for any $w \in V$ , an infinite-dimensional function space. Weak solutions to the PDE are equally valid and exact as solutions to the strong form. The only limitation is that they only obey the original PDE in a distribution sense. This is a sensible tradeoff, in exchange we get a whole new set of solutions that do not need to be twice differentiable.

Index notation

To get to the weak form we first multiply the PDE by a weight function $w$ and integrate over the domain $Ω$ :

(6.11)#

- \int_{Ω} w q_{i, i} d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω = 0

where $w$ for this problem is always a scalar field, irrespective of the dimensionality of $Ω$ . The next step is to eliminate the derivative $q_{i, i}$ using the divergence theorem and integration by parts. Here we can directly use the result we already derived in Eq. (1.32):

(6.12)#

\int_{Ω} w_{, i} q_{i} d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω - \int_{Γ_{h}} w q_{i} n_{i} d Γ = 0

and we see that the derivative has now been transferred to $w$ . This also causes an integral over $Γ$ to appear. From the requirement that $w = 0$ at $Γ_{D}$ , only the integral over $Γ_{N}$ remains. Finally, note that we did not touch the time derivative $\dot{u}$ , which means we are effectively treating $u$ and $\dot{u}$ as two independent fields.

As a final step, we substitute the Neumann boundary condition of Eq. (6.7) and the constitutive relation $q_{i} = - κ_{i j} u_{, j}$ to get to the final weak form:

(6.13)#

- \int_{Ω} w_{, i} κ_{i j} u_{, j} d Ω - \int_{Ω} w ρ c \dot{u} d Ω + \int_{Ω} w f d Ω + \int_{Γ_{h}} w h d Γ = 0

which is valid for any $w \in V$ , an infinite-dimensional function space. Weak solutions to the PDE are equally valid and exact as solutions to the strong form. The only limitation is that they only obey the original PDE in a distribution sense. This is a sensible tradeoff, in exchange we get a whole new set of solutions that do not need to be twice differentiable.

Semi-discrete form#

Tensor notation

To get to a tractable FEM problem we need to restrict the space $V$ . We do this through the Galerkin method, by defining shape function-based approximations to all of our fields:

(6.14)#

u^{h} = N a, {\dot{u}}^{h} = N \dot{a}, w^{h} = N c

where the shape function matrices and DOF vectors take the shape:

(6.15)#

\begin{array}{r} N = [\begin{array}{c} N_{1} & N_{2} & \dots & N_{n n} \end{array}] a = {\begin{array}{c} a_{1} \\ a_{2} \\ ⋮ \\ a_{n n} \end{array}} \dot{a} = {\begin{array}{c} {\dot{a}}_{1} \\ {\dot{a}}_{2} \\ ⋮ \\ {\dot{a}}_{n n} \end{array}} c = {\begin{array}{c} c_{1} \\ c_{2} \\ ⋮ \\ c_{n n} \end{array}} \end{array}

with $n n$ being the number of nodes in our finite element mesh. Now the problem has two DOFs per node, namely a solution value $a_{n}$ and a velocity ${\dot{a}}_{n}$ . We will also need gradients of $u^{h}$ and $w^{h}$ :

(6.16)#

\nabla u^{h} = B a, \nabla w^{h} = B c

with:

(6.17)#

\begin{array}{r} B = [\begin{array}{c} N_{1, x} & N_{2, x} & \dots & N_{n n, x} \\ N_{1, y} & N_{2, y} & \dots & N_{n n, y} \end{array}] \end{array}

We now substitute the approximations above into the weak form. The goal is to reach a so-called semi-discrete form, indicating that only spatial derivatives are discretized while time derivatives remain untouched and are treated as additional DOFs.

Substitution yields:

(6.18)#

- \int_{Ω} {(B c)}^{T} κ B a d Ω - \int_{Ω} {(N c)}^{T} ρ c N \dot{a} d Ω + \int_{Ω} {(N c)}^{T} f d Ω + \int_{Γ_{h}} {(N c)}^{T} h d Γ = 0

Making use of Eq. (1.17) we can remove $c$ , $a$ and $\dot{a}$ from the integrals:

(6.19)#

- c^{T} (\int_{Ω} B^{T} κ B d Ω) a - c^{T} (\int_{Ω} N^{T} ρ c N d Ω) \dot{a} + c^{T} (\int_{Ω} N^{T} f d Ω) + c^{T} (\int_{Γ_{h}} N^{T} h d Γ) = 0

and cancel out $c$ from all terms. We then gather the expressions into the more compact form:

(6.20)#

K a + M \dot{a} = f

where the time derivative makes the new term $M \dot{a}$ appear. We have then reached the final discretized form and can write down the terms one more time for completeness:

(6.21)#

K = \int_{Ω} B^{T} κ B d Ω

(6.22)#

M = \int_{Ω} N^{T} ρ c N d Ω

(6.23)#

f = \int_{Ω} N^{T} f d Ω + \int_{Γ_{h}} N^{T} h d Γ

Index notation

To get to a tractable FEM problem we need to restrict the space $V$ . We do this through the Galerkin method, by defining shape function-based approximations to all of our fields:

(6.24)#

u^{h} = N_{n} a_{n}, {\dot{u}}^{h} = N_{n} {\dot{a}}_{n}, w^{h} = N_{m} c_{m}

where the shape function matrices and DOF vectors take the shape:

(6.25)#

\begin{array}{r} N_{n} = [\begin{array}{c} N_{1} & N_{2} & \dots & N_{n n} \end{array}] a_{n} = {\begin{array}{c} a_{1} \\ a_{2} \\ ⋮ \\ a_{n n} \end{array}} {\dot{a}}_{n} = {\begin{array}{c} {\dot{a}}_{1} \\ {\dot{a}}_{2} \\ ⋮ \\ {\dot{a}}_{n n} \end{array}} c_{m} = {\begin{array}{c} c_{1} \\ c_{2} \\ ⋮ \\ c_{n n} \end{array}} \end{array}

with $n n$ being the number of nodes in our finite element mesh. Now the problem has two DOFs per node, namely a solution value $a_{n}$ and a velocity ${\dot{a}}_{n}$ . We will also need gradients of $u^{h}$ and $w^{h}$ :

(6.26)#

u_{, j}^{h} = N_{n, j} a_{n}, w_{, i}^{h} = N_{m, i} c_{m}

with:

(6.27)#

\begin{array}{r} N_{n, j} = [\begin{array}{c} N_{1, x} & N_{2, x} & \dots & N_{n n, x} \\ N_{1, y} & N_{2, y} & \dots & N_{n n, y} \end{array}] \end{array}

We now substitute the approximations above into the weak form. The goal is to reach a so-called semi-discrete form, indicating that only spatial derivatives are discretized while time derivatives remain untouched and are treated as additional DOFs.

Substitution yields:

(6.28)#

- \int_{Ω} N_{m, i} c_{m} κ_{i j} N_{n, j} a_{n} d Ω - \int_{Ω} N_{m} c_{m} ρ c N_{n} {\dot{a}}_{n} d Ω + \int_{Ω} N_{m} c_{m} f d Ω + \int_{Γ_{h}} N_{m} c_{m} h d Γ = 0

Making use of Eq. (1.17) we can remove $c_{m}$ , $a_{n}$ and ${\dot{a}}_{n}$ from the integrals:

(6.29)#

- c_{m} (\int_{Ω} N_{m, i} κ_{i j} N_{n, j} d Ω) a_{n} - c_{m} (\int_{Ω} N_{m} ρ c N_{n} d Ω) {\dot{a}}_{n} + c_{m} (\int_{Ω} N_{m} f d Ω) + c_{m} (\int_{Γ_{h}} N_{m} h d Γ) = 0

and cancel out $c_{m}$ from all terms. We then gather the expressions into the more compact form:

(6.30)#

K_{m n} a_{n} + M_{m n} {\dot{a}}_{n} = f_{m}

where the time derivative makes the new term $M_{m n} {\dot{a}}_{n}$ appear. We have then reached the final discretized form and can write down the terms one more time for completeness:

(6.31)#

K_{m n} = \int_{Ω} N_{m, i} κ_{i j} N_{n, j} d Ω

(6.32)#

M_{m n} = \int_{Ω} N_{m} ρ c N_{n} d Ω

(6.33)#

f_{m} = \int_{Ω} N_{m} f d Ω + \int_{Γ_{h}} N_{m} h d Γ

Solving the semi-discrete system#

We have reached a semi-discrete system with $n n$ equations, the number of nodes in the FE mesh. However, we now have twice as many DOFs: solution values $a$ and velocities $\dot{a}$ . The problem seems under-determined, but that is only because we still did not take into account that $a$ and $\dot{a}$ are of course coupled in time and for now the FEM discretization is only taking care of spatial derivatives. We will address this on the next page.

Semi-discrete form for diffusion

Contents

6.2. Semi-discrete form for diffusion#

Strong form equation#

Weak form#

Semi-discrete form#

Solving the semi-discrete system#