I'm using comsol semiconductor module to model the charge transport in a chunk of semiconductor I vary the thickness of the semiconductor and solve the model as a parametric sweep. When I solve the model with two different initial solutions ( one is the zero solution or the intrinsic conditions, and second is the solution for the model for the previous thickness) model output is different. I have tried reducing the 'Relative tolerance' in the stationary solver (my study is stationary), but still the same. What other parameters in the solver I should check out?
Unsteady problems essentially require proper initial condition. If the problem is steady state one, any initial condition will evolve to same solution [This is essentially a case in the question ]. With the variation in thickness of the semiconductor, domain becomes different for, say, t1 and t2 there can be variations in the data. In my opinion solver parameters should not have huge impact if they have already reasonably acceptable values. Run as two separate problems (may be with same or different initial conditions); check if you extrapolate low grid size data to higher grid size. You may compare your data with the analytical or experimental data, if it is available.
It is the best and the worst feature of COMSOL: it can calculate everything and present nice charts and graphs.
The case you have encountered is very common for comsol users, as comsol allows you to start calculations at pretty complex systems and results do converge.
My suggestion will be to decompose your model and simplify task for comsol - this often helps to get stable solutions. Then you may add complexity r extra parameters, to gradually improve your model.
We are also COMSOL users (semiconductor module in the stationary solver) and find a similar problem many times. But I would say that this is a general drawback of numerical solvers: all searching-solution algorithms are very dependent on the initial conditions. To such a point that to spent time providing a good set of initial conditions, as close as possible to the final solution, is a task as important as defining the problem itself, and actually saves much time to the search of the final result.
Coming back to your problem, our protocol and my recommendation is to use as initial condition the solution of previous but more simple cases. Starting form a case where solution is easy and unquestionable, you should manually approach your current problem by gradually varying your parameters using each case the previous solution as initial condition. We have achieved a satisfactory solution of many cases that way.
For other finite element packages, I usually test my model with a known solution then gradually I turn in into my desired problem and geometry. Initial conditions are imperative too.
- Badges
- Science topic