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.
Problem with Joules Heating module
Posted Dec 16, 2011, 11:17 p.m. EST 1 Reply
Please login with a confirmed email address before reporting spam
Hi,
I am trying to simulate the Micro Heater model for some application, When simulate the model using Joule's Heating module with stationary solver i m getting the result but the same model when i am trying with Time dependent solver i am getting the following error message, can you kindly help me in solving this problem.
The error message is :
Feature: Time dependent solver (sol 1/t1)
Error : Failed to find consistant initial values
Regards,
Kathir
I am trying to simulate the Micro Heater model for some application, When simulate the model using Joule's Heating module with stationary solver i m getting the result but the same model when i am trying with Time dependent solver i am getting the following error message, can you kindly help me in solving this problem.
The error message is :
Feature: Time dependent solver (sol 1/t1)
Error : Failed to find consistant initial values
Regards,
Kathir
1 Reply Last Post Dec 17, 2011, 3:49 a.m. EST
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hi
that can come from many things, I suspect it is related to a) the mesh density, and b) what it states: your initial conditions are too different from the expected solution, the solver does not find it's track from the starting point to THE (uniquely accepted) solution.
My best guess to solve it is: a) define some initial conditions (typically a T gradient through your system) that is closer to the solution FOR t=0, or mid way between t=0 and t=your_first_time_step_value. b) be sure you do not have any discontinuity in the BC definitions (sudden Dirac type turn on of values such as T(t=0) = 0, T(t>0) = 300[K]
and c) check the diffusivity of your material and the relation mesh size to time stepper duration (but the true time step COMSOL solver will use, not the only the one you defined).
Diffusion equations, such as HT, but also for chemistry and species diffusion, have solutions with very steep gradients, often around t=0 these MUST be resolved by the mesh otherwise you are not getting correct results, and the solver fails. It's just as important than for a Nyquist criteria in discrete time sampling (everyone knows and respect that no ? ;) for FEM its the same so alpha[m^2/s]=k/rho/Cp is linking the time steps Dt and the mesh size DZ and one MUST respect DZ<sqrt(alpha*Dt), if possile by a factor 2 or 4 or even more (check Nagi's responses onthe FORUM, he know more than me on this subject).
This is the main reason for negative temperatures (cooling) during the first steps, that people interprete wrongly or as "physics" or the saturating heat flux observed from a constant temperature boundary, or the sqrt(-1) in species negative concetrations. Unfortunately COMSOL does NOT propose a simple plot of mesh density versus time step, alpha which is similar in many ways to a Reynolds cell polt, or a Nusselt, Prandtl ... plot in CFD. Unfortunately, there has not been enough publications around this so many people (of the younger generation) forget this and take this numerical error for a physical property !
How many of you out there do systematically validate and verify all dependent variables for each of your models ? (I must admit I do not do it systematically, but I do spend a few days each time I enter a new physics field, or re-enter it after some months doing other physics simulations to be sure I have my warning flags turned on and the correct one).
Example: try out a simple time dependent HT model with about 1m average length geometrical scale and 100-300K temperature step on one boundary (something we experience each day cooking or in the fab), and an hour time range. Then take same model but a MEMSscale device 5 um large, 100 um thick with 3 layers metal < 1 um, oxyde SiO2 ~ 2um and a thick bulk Si, and set a 100-300[K] heating on the metal layer, here reaction times are in the usec domain, but the mesh ! needs to be in the nm ! because of the scaling laws of alpha and its related variables k/rho/Cp and because of the volume to surface ration, that changes drasticalyy between these to model dimensions, and our perception, jumping from one model to the other is fully fooled. Be aware !
By the way try and cmpare the results with the free BDF time stepper and witha strict or intermediate for a reasonnable defined range (I often use log or power typ scales such as
dt = 2^{range(-16,1,16)} or dt = 10^{range(-3,0.5,3}
By the way, this applies to all diffusion problems, chemistry, Heat ... for historical reasons we treat them differently, but behind the hood they are the same, this is some of the nice things with COMSOL we are reuniting the different physics in a much more logical way: the basic math rules, and no longer humain history and his desire to have his name on a formula, even if it was "just" a variant from an other existing one ...
--
Good luck
Ivar
that can come from many things, I suspect it is related to a) the mesh density, and b) what it states: your initial conditions are too different from the expected solution, the solver does not find it's track from the starting point to THE (uniquely accepted) solution.
My best guess to solve it is: a) define some initial conditions (typically a T gradient through your system) that is closer to the solution FOR t=0, or mid way between t=0 and t=your_first_time_step_value. b) be sure you do not have any discontinuity in the BC definitions (sudden Dirac type turn on of values such as T(t=0) = 0, T(t>0) = 300[K]
and c) check the diffusivity of your material and the relation mesh size to time stepper duration (but the true time step COMSOL solver will use, not the only the one you defined).
Diffusion equations, such as HT, but also for chemistry and species diffusion, have solutions with very steep gradients, often around t=0 these MUST be resolved by the mesh otherwise you are not getting correct results, and the solver fails. It's just as important than for a Nyquist criteria in discrete time sampling (everyone knows and respect that no ? ;) for FEM its the same so alpha[m^2/s]=k/rho/Cp is linking the time steps Dt and the mesh size DZ and one MUST respect DZ<sqrt(alpha*Dt), if possile by a factor 2 or 4 or even more (check Nagi's responses onthe FORUM, he know more than me on this subject).
This is the main reason for negative temperatures (cooling) during the first steps, that people interprete wrongly or as "physics" or the saturating heat flux observed from a constant temperature boundary, or the sqrt(-1) in species negative concetrations. Unfortunately COMSOL does NOT propose a simple plot of mesh density versus time step, alpha which is similar in many ways to a Reynolds cell polt, or a Nusselt, Prandtl ... plot in CFD. Unfortunately, there has not been enough publications around this so many people (of the younger generation) forget this and take this numerical error for a physical property !
How many of you out there do systematically validate and verify all dependent variables for each of your models ? (I must admit I do not do it systematically, but I do spend a few days each time I enter a new physics field, or re-enter it after some months doing other physics simulations to be sure I have my warning flags turned on and the correct one).
Example: try out a simple time dependent HT model with about 1m average length geometrical scale and 100-300K temperature step on one boundary (something we experience each day cooking or in the fab), and an hour time range. Then take same model but a MEMSscale device 5 um large, 100 um thick with 3 layers metal < 1 um, oxyde SiO2 ~ 2um and a thick bulk Si, and set a 100-300[K] heating on the metal layer, here reaction times are in the usec domain, but the mesh ! needs to be in the nm ! because of the scaling laws of alpha and its related variables k/rho/Cp and because of the volume to surface ration, that changes drasticalyy between these to model dimensions, and our perception, jumping from one model to the other is fully fooled. Be aware !
By the way try and cmpare the results with the free BDF time stepper and witha strict or intermediate for a reasonnable defined range (I often use log or power typ scales such as
dt = 2^{range(-16,1,16)} or dt = 10^{range(-3,0.5,3}
By the way, this applies to all diffusion problems, chemistry, Heat ... for historical reasons we treat them differently, but behind the hood they are the same, this is some of the nice things with COMSOL we are reuniting the different physics in a much more logical way: the basic math rules, and no longer humain history and his desire to have his name on a formula, even if it was "just" a variant from an other existing one ...
--
Good luck
Ivar