2
$\begingroup$

I'm looking to numerically compute the solution to

$$ k(x,u) \partial_t u - \Delta u = f \quad\quad\text{ in } \Omega \times [0,T]$$

where $k$ is a continuous but nonlinear (in $u$) real-valued function and $\Omega \subseteq \mathbb{R}^2$. I want to use the Finite Element Method, but I cannot seem to manage the nonlinear $k$. I wrote the code myself and it already works for $k = const$, i.e. in the "standard case".

My problem is, that I cannot get rid of the $k(u)$ factor and therefore don't reach a formula like $$ M \dot{u} - A u = F $$ like in page 80 here (mass matrix $M$, stiffness matrix $A$). I also can't seem to reach any formulation which would allow me to apply the Newton method or something similar. Is it even possible to use a FEM approach in this case?

Any help or reference is appreciated.

$\endgroup$
  • 3
    $\begingroup$ Hi mstrkft and welcome to scicomp! In general, if your PDE is non-linear, then you should expect your system of equations will also be non-linear and will need to consider a non-linear solver such as newton method. $\endgroup$ – Paul Dec 16 '14 at 20:14
  • $\begingroup$ Hi Paul, thank you for your answer. I already managed to derive a implicit Finite Volume solver based on the Newton method. However, I want to use a FEM approach, but I can't seem to bring this into a form, where I can use the Newton method. $\endgroup$ – mstrkrft Dec 17 '14 at 8:17
2
$\begingroup$

As a first shot, I suggest you use Rothe's method (time discretization first), rather than the method of lines (space discretization first).

If you employ implicit Euler for that, at every time step $l$, you will have to solve the system

$$ -\tau \Delta u^l+k(u^l)u^l = k(u^{l-1})u^{l-1} + \tau [ \Delta u^{l-1} + f^l], $$

which is a semilinear elliptic problem. Then you can apply, e.g., Newton or fixed point iteration.

$\endgroup$
2
$\begingroup$

You will have to iterate out the problem. After discretization, you get a problem of the form $$ M(U^n) U^n + \Delta t \; A U^n = F^n(U^{n-1}) $$ where the mass matrix depends on the solution $U^n$ of the n-th time step. This nonlinear system of equations has to be solved by an iteration, e.g., the Newton method.

An example for a different question, but using the same idea, is given here: https://www.dealii.org/developer/doxygen/deal.II/step_15.html

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.