Hi,
I try to observe the heating of a system using an electric pulse, i.e. Joule heating (--> Laplace equation and heat equation). The conductivity of the material of interest depends on the local electric field and local temperature, so the system is highly interdependent and therefore I am using the fully–coupled solving approach.
The meshing itself is not great 0.02 min. quality, however, due to high aspect ratios (a cylinder of 5nm thickness vs. 200nm diameter) hard to improve. I split the thickness manually into 8 layers, which is what I need for the physics I like to observe (the material itself is highly inhomogeneous with respect to conductivity so the local electric field depends on the number of layers I actually simulate).
In the beginning the solving works quite well, but at some time more and more NLfails occur. I already adjusted the absolute tolerance to a fairly low level (sometimes 1e7, 1e-8) and use a rel. tolerance of 1e-3 and tolerance factor of 1e-4 which to my understanding is similar to using 1e-7 since I am not using the segregated solver. I found that hint in the forum and it should help to shift the problem from the time-stepping to the solving and in fact I do get significantly less errors.
I also change the solver from BDF to Generalised alpha due to some sharp gradients that can and will occur. I also manually adjusted the time-stepping to put more emphasis on the Non–linear solver but also that does not really help.
What I find quite puzzling is that the error seems to be quite small (see attachment), however the number of NLfailuers increases continuously showing that there is a problem.
It would be great if someone has a suggestion of how to tackle the problem or at least knows why there are so many NLfailures.
Many thanks!
Toby
I try to observe the heating of a system using an electric pulse, i.e. Joule heating (--> Laplace equation and heat equation). The conductivity of the material of interest depends on the local electric field and local temperature, so the system is highly interdependent and therefore I am using the fully–coupled solving approach.
The meshing itself is not great 0.02 min. quality, however, due to high aspect ratios (a cylinder of 5nm thickness vs. 200nm diameter) hard to improve. I split the thickness manually into 8 layers, which is what I need for the physics I like to observe (the material itself is highly inhomogeneous with respect to conductivity so the local electric field depends on the number of layers I actually simulate).
In the beginning the solving works quite well, but at some time more and more NLfails occur. I already adjusted the absolute tolerance to a fairly low level (sometimes 1e7, 1e-8) and use a rel. tolerance of 1e-3 and tolerance factor of 1e-4 which to my understanding is similar to using 1e-7 since I am not using the segregated solver. I found that hint in the forum and it should help to shift the problem from the time-stepping to the solving and in fact I do get significantly less errors.
I also change the solver from BDF to Generalised alpha due to some sharp gradients that can and will occur. I also manually adjusted the time-stepping to put more emphasis on the Non–linear solver but also that does not really help.
What I find quite puzzling is that the error seems to be quite small (see attachment), however the number of NLfailuers increases continuously showing that there is a problem.
It would be great if someone has a suggestion of how to tackle the problem or at least knows why there are so many NLfailures.
Many thanks!
Toby