Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.

Load vector is always zero vector in linear problems? Hmm...

Please login with a confirmed email address before reporting spam

Attached is a model with one 1D element and 2 Dirichlet-type BCs on both ends of it implemented similarly to Chien Liu's blog post here: www.comsol.com/blogs/implementing-the-weak-form-in-comsol-multiphysics/

Using Linear discretisation, stationary solver, Laplace equation in the domain, this is about the simplest FEM model possible, so it should be very easy to interpret.

As described in Chien Liu's blog post that follows the one linked above in that series
www.comsol.com/blogs/implementing-the-weak-form-with-a-comsol-app
we can add an Assemble node under the solver node, and view the System Matrix under Derived Values.

In our case we have 4 DOFs: two corresponding to the dependent variable's value at each end of the element, and two for the lambdas used to implement the Dirichlet BCs. The stiffness matrix looks about right:
0 1 0 0
1 1 0 -1
0 0 0 1
0 -1 1 1
But the load vector is all zeros:
0
0
0
0
(Sometimes when testing with some trivial changes, the middle entries show up as small values around the machine epsilon ~1e-16.)

The BCs are 2.0 on the left end and 7.0 on the right, and the solution shows up correctly in the plot.

The eliminated stiffness matrix and load vector are the same (up to a scaling constant), because we are using weak constraints (nothing is eliminated, and the Lagrange multipliers are explicitly solved for). If we implement the Dirichlet BCs the usual way instead, and check the "Use weak constraints" checkbox, then exactly the same thing happens as above. Also, we can see under Equation View, that we have the same equations as when we use the "Weak Contribution" method to implement the Dirichlet BCs.

If we use LiveLink and get the Kc, Lc matrices, then, of course, the solution Kc\Lc is just the trivial all-zeros solution (as Lc is a zero vector, and so is L), which is incorrect.

There is a comment under Chien Liu's post (second link above) from Jing Wang July 6, 2015, reporting the same issue, which has been referred to Comsol support, but with no public follow-up. My guess is that, since the blog post refers to Comsol 5.0, that it used to work as expected under 5.0, but has changed since then.

The interesting thing is that I found a way to get the expected load vector, using this strange procedure: in "Stationary Solver 1" under "Linearity" choose "Linear perturbation", with "Values of linearization point" "Prescribed by" "Solution", "Solution": "Solution 1 (sol1)". If we solve and assemble now, we get a flat zero solution, but the load vector is now evaluated to the expected values.

However, under no combination of these Linearity settings could I get both the expected solution in the plots and the expected load vector from System Matrix evaluation.

My interpretation is that either this is a bug, or Comsol since 5.1 has been using another layer of linearisation point picking, that we can't access. Either way, we can't export the matrices into Matlab and solve the system there with higher precision using the procedure dictated by common sense.


3 Replies Last Post Jun 22, 2016, 8:21 AM EDT
Chien Liu COMSOL Employee

Please login with a confirmed email address before reporting spam

Posted: 3 years ago Jun 20, 2016, 8:51 AM EDT
Dear Evgeni,

Thank you for the comment. Did you perform this step as described in the text of the blog post: “Make sure to drag the Assemble node up to position it before the Stationary Solver node.” ?

If this step is not done and the Assemble node is left after (below) the Stationary Solver node, then the content of the load vector may be altered by the solver as you have observed.

Hope this helps.

Sincerely,

Chien
Dear Evgeni, Thank you for the comment. Did you perform this step as described in the text of the blog post: “Make sure to drag the Assemble node up to position it before the Stationary Solver node.” ? If this step is not done and the Assemble node is left after (below) the Stationary Solver node, then the content of the load vector may be altered by the solver as you have observed. Hope this helps. Sincerely, Chien

Lars Gregersen COMSOL Employee

Please login with a confirmed email address before reporting spam

Posted: 3 years ago Jun 20, 2016, 10:39 AM EDT
Dear Evgeni,

LiveLink for Matlab has been updated for version 5.2a such that you can specify which solver nodes should be evaluated. The default behaviour to solve (i.e. prepare the system matrices) up until the step before the (stationary) solver.

In short: mphmatrix works for this scenario in version 5.2a.


--
Lars Gregersen
Comsol Denmark
Dear Evgeni, LiveLink for Matlab has been updated for version 5.2a such that you can specify which solver nodes should be evaluated. The default behaviour to solve (i.e. prepare the system matrices) up until the step before the (stationary) solver. In short: mphmatrix works for this scenario in version 5.2a. -- Lars Gregersen Comsol Denmark

Please login with a confirmed email address before reporting spam

Posted: 3 years ago Jun 22, 2016, 8:21 AM EDT
Dear Chien and Lars, thank you for the prompt help. Indeed I have missed the step concerning the order of the Assemble and the Solver nodes. Fixing that gives me the expected load vector when I evaluate it inside Comsol GUI. Also, good to know that I must be the last person to run into this confusion with LiveLink, given the release of 5.2a about five days ago :) The best outcome is that there isn't a terrible gap in my understanding, but rather it were the mysterious internal workings of Comsol, who didn't give up the load vector without a fight.

Dear Chien and Lars, thank you for the prompt help. Indeed I have missed the step concerning the order of the Assemble and the Solver nodes. Fixing that gives me the expected load vector when I evaluate it inside Comsol GUI. Also, good to know that I must be the last person to run into this confusion with LiveLink, given the release of 5.2a about five days ago :) The best outcome is that there isn't a terrible gap in my understanding, but rather it were the mysterious internal workings of Comsol, who didn't give up the load vector without a fight.

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.