High computation time and other questions

solid-solid phase transformations, influence of stresses and strains
Post Reply
pmecozzi
Posts: 10
Joined: Mon Apr 15, 2013 3:20 pm
anti_bot: 333

High computation time and other questions

Post by pmecozzi » Mon Oct 26, 2020 2:07 pm

Dear Bernd,

I have performed simulations of austenite formation in a Fe-0.67 wt% C -1.15 wt% Mn steel during heating at 1 K/s. The starting microstructure consists of pearlite and 9% of pre-eutectoid ferrite. I used a low grid size (5 nm) to resolve the fine cementite lamella.
The simulation was run with Thermo-calc coupling: the GES file was derived using Thermo-calc version 2020a (TCFE9 and MOBFE4).
The initial concentration was set as equilibrium 2. The value of ferrite composition at 973 K was derived by Thermo-calc (0.012 wt% C, 0.76 wt % Mn) (see the driving file PearliteDiss_TQpear_1Cs-1_Portion1_red1_in).
My first simulation run without problems but it was very slow (see the log file PearliteDiss_TQpear_1Cs-1_Portion1_red1_log.txt).
• Could you please suggest any change in the driving file to decrease the time step?
Besides, when I saw the microstructure after 0.49 s (see the file PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.49s. png) I found that only one of the two austenite nuclei, which MICRESS set in the microstructure, grew (see the file PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.03s. png).
• Are there any reasons for the different behaviour of the two austenite grains?

In the second simulation, I increased the initial temperature from 1003 K to 1015 K to better fit some experimental data (see the driving file PearliteDiss_TQpear_1Cs-1_Portion1_red2_in.txt).
In this case, MICRESS set only one nucleus (instead of two in the previous simulation) and this nucleus did not grow, despite the high undercooling (3084.6 K!!!). I suppose that this was a consequence of the change in the linearisation parameters of the phases BCC/FCC by increasing the T0 of 12 K but I was not expecting such a big change(see the log filePearliteDiss_TQpear_1Cs-1_Portion1_red2_log.txt).

Then I run the third simulation by changing the setting of the initial concentration from “equilibrium 2” to “from file”.
The files for the concentration of Mn and C were derived by the initial concentration maps of the previous simulation; all the other settings have remained unchanged.
In this case, the relative fraction of the ferrite and cementite phase changes after the first step, before the austenite nucleation (see PearliteDiss_TQpear_1Cs-1_Portion1_red3_tabf.txt).
• What is the reason for the variation of the relative fraction of ferrite and cementite?
Also after the nucleus was set, the following error message appeared
trying hard phases 2 1 level: 4 zp= 115196 error= 3….
……………..
……………..
(see the PearliteDiss_TQpear_1Cs-1_Portion1_red3_log.txt).

Finally, I run the fourth simulation by changing the temperature at which the initial equilibrium is calculated (see the driving file PearliteDiss_TQpear_1Cs-1_Portion1_red4_in); the error message disappeared but no nuclei were set!

Please let me know which other input/output files you will need for helping me in defining the correct driving file for simulating the pearlite to austenite formation.

Best Regards

Pina
Attachments
PearliteDiss_TQpear_1Cs-1_Portion1_red4_log.txt
(27.23 KiB) Downloaded 42 times
PearliteDiss_TQpear_1Cs-1_Portion1_red4_in.txt
(35.4 KiB) Downloaded 40 times
PearliteDiss_TQpear_1Cs-1_Portion1_red3_log.txt
(30.57 KiB) Downloaded 55 times
PearliteDiss_TQpear_1Cs-1_Portion1_red3_in.txt
(35.4 KiB) Downloaded 39 times
PearliteDiss_TQpear_1Cs-1_Portion1_red2_log.txt
(30.74 KiB) Downloaded 48 times
PearliteDiss_TQpear_1Cs-1_Portion1_red2_in.txt
(34.68 KiB) Downloaded 45 times
PearliteDiss_TQpear_1Cs-1_Portion1_red1_log.txt
(55.6 KiB) Downloaded 47 times
PearliteDiss_TQpear_1Cs-1_Portion1_red1_in.txt
(34.68 KiB) Downloaded 41 times
PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.49s.png
PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.49s.png (18.66 KiB) Viewed 608 times
PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.03s.png
PearliteDiss_TQpear_1Cs-1_Portion1_red1_phas_0.03s.png (18.12 KiB) Viewed 608 times

Bernd
Posts: 1222
Joined: Mon Jun 23, 2008 9:29 pm

Re: High computation time and other questions

Post by Bernd » Tue Oct 27, 2020 10:51 am

Hi Pina,

Thanks for your detailed description of your simulations.

The resolution of your simulation is quite high (Δx=5 nm). This may be the simple reason that your simulations are so slow, given the fact that carbon is diffusing quite fast. So my guess is that most time is lost in calculating diffusion of carbon in bcc. You should check the .TabP output to verify that most time is spent in the diffusion solver, and .TabD to see that C in bcc has the highest diffusivity.

One approach to speed up the simulation would of course be to simulate with coarser resolution (e.g. going to Δx=10 nm would reduce the calculation time for diffusion by a factor of 16). If you want to keep the high resolution, another option would be to assume infinite diffusion of C in bcc. However, this is only viable if diffusion of C in bcc is not limiting the process, and results do not depend on this diffusivity. At least at the beginning of the transformation, this would require that there is a kinetic control either by the bcc-cementite or the bcc-fcc interfaces.

Another consequence of the high resolution are the high curvature undercooling values. This may be the reason that your fcc nuclei sometimes do not grow despite there is a driving force for phase transformation. You should not look at the high values of "overheating" which you see when nuclei are set: Local compositions can be extreme if you check for nucleation in the very outer parts of the interfaces (which you explicitly enable by setting the fraction range for nucleation between 0 and 1 - I would not do that!).

Another reason for not getting successful nucleation of fcc can be the limit to only 10 nuclei. If they have been lost by checking in extreme places, no further chances are left. I generally do not recommend to limit the number of nuclei.

Another point which may be worth to have a look at is the high number of updates of thermodynamic data due to "fast moving interfaces", e.g.:

Fast moving interface(s) relinearized FCC_A1/CEMENTITE (448)

When using global updating schemes for thermodynamic data, MICRESS does extra updates if the size of the interfaces change considerably. The number in parenthesis means the number of such updates since the last output time. The large number here could indicate fluctuations which could be due to a too high interface mobility. I know that you prefer not to assume diffusion limited kinetics but use tabulated mobility data. However, even if these values are physically correct, you still should use mobility correction ("mob_corr") to account for numerical effects of the finite interface thickness and to avoid too high values.

Bernd

pmecozzi
Posts: 10
Joined: Mon Apr 15, 2013 3:20 pm
anti_bot: 333

Re: High computation time and other questions

Post by pmecozzi » Tue Oct 27, 2020 4:37 pm

Dear Bernd,

Many thanks for answering all my questions.
At first I have to say that the carbon diffusivity in ferrite is indeed very high.
Simulation Temperature Diff Diff. C. Diff.
time[s] [K] P1 C1 P1 C2 P2 C1 P2 C2
0 1003 0.00E+00 0.00E+00 0.00E+00 0.00E+00
0.11376 1003.114 1.16E-08 5.34E-15 8.58E-07 7.29E-13
0.2418 1003.242 1.16E-08 5.36E-15 8.59E-07 7.33E-13
0.25 1003.25 1.16E-08 5.36E-15 8.59E-07 7.33E-13
0.387728 1003.388 1.16E-08 5.38E-15 8.61E-07 7.37E-13

However, I also checked the tabP file and I found that most of the time is spent in the phase-field solver and not in the diffusion solver.
Simulation CPU time Wallclock TQ time PF time Diff time
time [s] time [s] [s] [s] [s]
0 0 0 0 0 0ΔΔΙ
0.11376 80753.02 21600.07 274.83 14508.97 4153.449
0.2418 161920.1 43200.1 625.288 27833.14 8865.073
0.25 166910.6 44523.36 648.592 28628.45 9159.425
0.387728 248490 66123.4 1131.056 41247.75 14047.2


Nevertheless, I will reduce the resolution by doubling the grid size and halving the number of grid points in the x and z-directions; I have already done simulations with Δx=10 nm but I used the same number of grids points to simulate the transformation in a bigger area; of course, I did not save computation time.
However, since I am wondering that the reduction of the number of grids in the cementite lamella may affect the results, I will also simulate the transformation by maintaining the resolution low and setting the carbon diffusivity in ferrite infinite, as you have suggested.
Of course, I will also use your suggestion of not setting the range of nucleus position as [0 1] (hoping that I will reduce the overheating), increasing the number of nuclei in the nucleation input data and using the mob_corr option in the phase interaction input data.

Best Regards

Pina

Bernd
Posts: 1222
Joined: Mon Jun 23, 2008 9:29 pm

Re: High computation time and other questions

Post by Bernd » Wed Oct 28, 2020 2:10 pm

Hi Pina,

If it is really the "PF-time" and the "List time" in the .TabP which consumes most (it is difficult to see your pasted columns in your answer), then this means that most probably the phase-field time steps are very small. This can either be due to a too large interface mobility (which perhaps is cured automatically if you use "mob_corr"), or to some localized thermodynamic trouble. In many cases, using a "minimum time step" can optimize performance as explained here.

Bernd

pmecozzi
Posts: 10
Joined: Mon Apr 15, 2013 3:20 pm
anti_bot: 333

Re: High computation time and other questions

Post by pmecozzi » Thu Oct 29, 2020 10:41 am

Hi Bernd,

Many thanks for the further suggestions.
I also attach the tabP file since the numbers that I copied/pasted in the previous e-mail were indeed mixed up.

Best Regards

Pina
Attachments
PearliteDiss_TQpear_1Cs-1_Portion1_red1_TabP.txt
(1.53 KiB) Downloaded 38 times

Post Reply