unusually high simulation time
unusually high simulation time
Hi,
I am getting unusually high simulation time when I simulate solidification simulation with high angle grain boundaries. Simulation is still running for 4 days and not even 50% is completed. However when I did the same simulation with low angle grain boundaries, the simulation finished within 11 hrs. Is this expected? I am not getting any errors. but I cannot understand why its taking a longer time for high angle grain boundary.
BR
Chamara
I am getting unusually high simulation time when I simulate solidification simulation with high angle grain boundaries. Simulation is still running for 4 days and not even 50% is completed. However when I did the same simulation with low angle grain boundaries, the simulation finished within 11 hrs. Is this expected? I am not getting any errors. but I cannot understand why its taking a longer time for high angle grain boundary.
BR
Chamara
Re: unusually high simulation time
Hi Chamara,
Such behaviour could come either from a strongly reduced phase-field time stepping or from numerical instabilities. The first step should be to check out where the additional calculation time is going. Please check the .TabP output to see that in comparison between the two simulations. If the slowing down is due to a reduced time-stepping, you should see a strongly increased time consumption of the "PF time" and "List time" column of .TabP. The time-stepping itself can be checked using the .TabT output.
Numerical instabilities, on the other hand, would probably lead to a strongly increased "TQ time".
If the job has been just running badly (IO-Problems, etc.), only the "Wallclock" time would be up. Instead, for a "perfectly" running job, the Wallclock time should be close to the sum of all the following columns.
Bernd
Such behaviour could come either from a strongly reduced phase-field time stepping or from numerical instabilities. The first step should be to check out where the additional calculation time is going. Please check the .TabP output to see that in comparison between the two simulations. If the slowing down is due to a reduced time-stepping, you should see a strongly increased time consumption of the "PF time" and "List time" column of .TabP. The time-stepping itself can be checked using the .TabT output.
Numerical instabilities, on the other hand, would probably lead to a strongly increased "TQ time".
If the job has been just running badly (IO-Problems, etc.), only the "Wallclock" time would be up. Instead, for a "perfectly" running job, the Wallclock time should be close to the sum of all the following columns.
Bernd
Re: unusually high simulation time
Hi Bernd,
Thanks a lot.
I just checked .TabT and .TabP files between two simulations. In HAGB simulation during the time where the grains meet up the time step is in E-08 range (mostly) where as for the LAGB it varies between E-07 and E-08.
I see very strong increase in TQ and PF times in HAGB simulation once the grains meet up.
In LAGB - TQ time is between 640 and 780.
In HAGB - TQ time is between 13000 to 330000
I am not getting any errors.
Any recommendation?
BR
Chamara
Thanks a lot.
I just checked .TabT and .TabP files between two simulations. In HAGB simulation during the time where the grains meet up the time step is in E-08 range (mostly) where as for the LAGB it varies between E-07 and E-08.
I see very strong increase in TQ and PF times in HAGB simulation once the grains meet up.
In LAGB - TQ time is between 640 and 780.
In HAGB - TQ time is between 13000 to 330000
I am not getting any errors.
Any recommendation?
BR
Chamara
Re: unusually high simulation time
Hi Chamara,
This sounds like if you have fluctuations of the phase-field parameter which create new interface cells/fragments, which always require new thermodynamic data for the 0/1 or other 0/x and 1/x phase interactions. You can check in the .TabTQ (giving the computation time for TQ by phase interaction) output whether other phases x are involved in the fluctuations.
Probably, the fluctuations are triggered in the triple junctions between the melt and two or more fcc grains. Depending on the misorientation model of the 1/1-interaction, the triple point contributions to the 0/1-interaction can be the source of the problems. Especially, the interface energy - if there is a large difference to that of the 0/1-interface - can cause strong contributions.
Phenomenologically, you can influence the fluctuation behaviour e.g. by modifying the interface mobility (you can easily reduce the 0/1-interface mobility by 2-3 orders of magnitude at low temperatures where only small rests of liquid remain) or using stabilisation. The hysteresis factor to the phase minimum can also help, e.g. using
#Phase minimum?
1.0E-6 200
instead of
1.0E-4
Bernd
This sounds like if you have fluctuations of the phase-field parameter which create new interface cells/fragments, which always require new thermodynamic data for the 0/1 or other 0/x and 1/x phase interactions. You can check in the .TabTQ (giving the computation time for TQ by phase interaction) output whether other phases x are involved in the fluctuations.
Probably, the fluctuations are triggered in the triple junctions between the melt and two or more fcc grains. Depending on the misorientation model of the 1/1-interaction, the triple point contributions to the 0/1-interaction can be the source of the problems. Especially, the interface energy - if there is a large difference to that of the 0/1-interface - can cause strong contributions.
Phenomenologically, you can influence the fluctuation behaviour e.g. by modifying the interface mobility (you can easily reduce the 0/1-interface mobility by 2-3 orders of magnitude at low temperatures where only small rests of liquid remain) or using stabilisation. The hysteresis factor to the phase minimum can also help, e.g. using
#Phase minimum?
1.0E-6 200
instead of
1.0E-4
Bernd
Re: unusually high simulation time
Hi Bernd
Thanks again.
When I check the .TabTQ, I can see 0/1 interaction time increases significantly compare to 0/2 and 1/2.
I used misorientation model with constant factor (0.1) for low angle GB.
My 0/1 interface energy is 1.2E-05 and 1/1 interface energy is 1.2E-04. I did not use any stabilization value for 0/1 interface energy.
I use mob_corr function 0/1 interaction.
Thanks again.
When I check the .TabTQ, I can see 0/1 interaction time increases significantly compare to 0/2 and 1/2.
I used misorientation model with constant factor (0.1) for low angle GB.
My 0/1 interface energy is 1.2E-05 and 1/1 interface energy is 1.2E-04. I did not use any stabilization value for 0/1 interface energy.
I use mob_corr function 0/1 interaction.
Re: unusually high simulation time
Hi Chamara,
A factor of 10 in interface energies inside a triple junction is already a lot - are you sure that the interface energy 1/1 is such high? Reducing it would probably solve the problem...
Otherwise you should try the stabilizing measures I mentioned before.
Bernd
A factor of 10 in interface energies inside a triple junction is already a lot - are you sure that the interface energy 1/1 is such high? Reducing it would probably solve the problem...
Otherwise you should try the stabilizing measures I mentioned before.
Bernd
Re: unusually high simulation time
I took the GB energy data form literature for Ni. There the highest value is around 1.2E-04 J/cm^2.
if it is round 8 times, is it still high?
if it is round 8 times, is it still high?
Re: unusually high simulation time
I cannot say exactly, this would be rather a question for Janin. But consider it in terms of triple point angles: A factor of two is already sufficient for complete wetting of the grain boundaries!
Re: unusually high simulation time
admin wrote: ↑Tue May 28, 2019 3:09 pmHi Chamara,
This sounds like if you have fluctuations of the phase-field parameter which create new interface cells/fragments, which always require new thermodynamic data for the 0/1 or other 0/x and 1/x phase interactions. You can check in the .TabTQ (giving the computation time for TQ by phase interaction) output whether other phases x are involved in the fluctuations.
Probably, the fluctuations are triggered in the triple junctions between the melt and two or more fcc grains. Depending on the misorientation model of the 1/1-interaction, the triple point contributions to the 0/1-interaction can be the source of the problems. Especially, the interface energy - if there is a large difference to that of the 0/1-interface - can cause strong contributions.
Phenomenologically, you can influence the fluctuation behaviour e.g. by modifying the interface mobility (you can easily reduce the 0/1-interface mobility by 2-3 orders of magnitude at low temperatures where only small rests of liquid remain) or using stabilisation. The hysteresis factor to the phase minimum can also help, e.g. using
#Phase minimum?
1.0E-6 200
instead of
1.0E-4
Bernd
Hi Bernd,
If you are using mob_corr option how can you define a low mobility value for 0/1 interface at lower temperatures?
Side question: does the mob_corr works if the phase is defined as an faceted face?
Re: unusually high simulation time
Hi Chamara,
If you use mob_corr you specify a physical mobility value, and MICRESS automatically calculates the corresponding numerical one. If you want to have a transition from diffusion limited growth at high temperatures to mobility control at low temperatures, you can use a temperature dependent definition of the physical mobility value from file. At high temperatures where you want to have diffusion control, you put a high value like 1.0, at lower temperatures you specify a low physical value which is below the numerical value for diffusion control, e.g.
# temperature/K phys. mobility/cm2s-1
2000 1.0
1500 1.0
1480 1.E-4
1400 2.E-5
1300 1.E-5
1000 1.E-5
mob_corr as well works for faceted growth.
Bernd
If you use mob_corr you specify a physical mobility value, and MICRESS automatically calculates the corresponding numerical one. If you want to have a transition from diffusion limited growth at high temperatures to mobility control at low temperatures, you can use a temperature dependent definition of the physical mobility value from file. At high temperatures where you want to have diffusion control, you put a high value like 1.0, at lower temperatures you specify a low physical value which is below the numerical value for diffusion control, e.g.
# temperature/K phys. mobility/cm2s-1
2000 1.0
1500 1.0
1480 1.E-4
1400 2.E-5
1300 1.E-5
1000 1.E-5
mob_corr as well works for faceted growth.
Bernd