Formation energy trends for pure system and system with vacancy from canonical MC simulations

[pandeydhanshree]

          Hello

I am trying to perform canonical MC calculations with A,B and vacancy (Va) as the components of the alloy at a certain temperature. So, my training DFT data contains structures like AB, BVa, AVa, and ABVa. Now I wish to see the formation energy trends at different temperatures for, let say, a: Pure AB system ("comp": [0.49, 0.51],) b: Low vacancy concentration ("comp": [0.49, 0.49, 0.02],) c: High vacancy concentration ("comp": [0.45454545, 0.45454545, 0.09090910],) The results I obtain (attached below) are not consistent with what we obtain from 0K cluster expansion/DFT energies (structures with vacancy have higher formation energy). Can you please have a look at the plots and let me know what is the issue. Thank you and best regards Dhanshree Casm-MC-results.pdf

Originally posted by @pandeydhanshree in https://github.com/prisms-center/CASMcode/issues/338#issuecomment-1911736125

[xivh]

Do you expect the vacancies to raise the formation energy or to lower the formation energy? Does the formation energy of the first step of your heating run match the DFT?

Also, can you upload your prim.json?

[pandeydhanshree]

"Do you expect the vacancies to raise the formation energy or to lower the formation energy? " I would except the formation energy for vacancies to be higher (less negative) than that of without vacancy. I hope that make sense, right? "Does the formation energy of the first step of your heating run match the DFT?" No, the formation energy does not match.

The prim.json is "{ "basis" : [ { "coordinate" : [ 0.00000, 0.00000, 0.00000], "occupants" : [ "W", "Ta", "Va"]}

], "coordinate_mode" : "Cartesian", "lattice_vectors" : [ [ -1.758031, 1.758031, 1.758031 ], [ 1.758031, -1.758031, 1.758031], [ 1.758031, 1.758031, -1.758031] ], "title" : "WTaVac" } "

[xivh]

I just created a project with your prim, and I only see two composition axes. Which axis are you using? Can you do casm composition -d to get the axes, and then query site_frac or comp_n and see if it matches what you expected?

A pure AB system is not necessarily "comp": [0.49, 0.51].

[pandeydhanshree]

Below is the output I am getting when casm composition -d is executed Possible composition axes:

   KEY     ORIGIN          a          b     GENERAL FORMULA
   ---        ---        ---        ---     ---
     0         Va         Ta          W     W(b)Ta(a)Va(1-a-b)
     1         Ta         Va          W     W(b)Ta(1-a-b)Va(a)
     2          W         Va         Ta     W(1-a-b)Ta(b)Va(a)

Currently selected composition axes: 0

I did the query site_frac/comp_n and I get what I expect. For example, some of the initial lines for "ccasm query -k site_frac" is Print:

• site_frac

name selected site_frac(W) site_frac(Ta) site_frac(Va)

SCEL1_1_1_1_0_0_0/0           1                             1.00000000  0.00000000  0.00000000
SCEL1_1_1_1_0_0_0/1           1                             0.00000000  1.00000000  0.00000000
SCEL9_1_9_1_0_0_0/0           1                             0.88888889  0.11111111  0.00000000
SCEL9_9_1_1_0_6_3/0           1                             0.55555556  0.33333333  0.11111111
SCEL9_9_1_1_0_8_4/0           1                             0.22222222  0.66666667  0.11111111
SCEL10_1_10_1_0_0_0/0         1                             0.80000000  0.20000000  0.00000000
SCEL10_2_5_1_0_0_1/0          1                             0.20000000  0.80000000  0.00000000
SCEL10_2_5_1_0_0_1/1          1                             0.30000000  0.60000000  0.10000000
SCEL10_2_5_1_0_0_1/2          1                             0.50000000  0.40000000  0.10000000

Regarding "A pure AB system is not necessarily "comp": [0.49, 0.51].", what do you suggest then for pure AB system? Thanks and regards Dhanshree

[xivh]

Okay, if you are using composition axis 0, then I think the composition axis is not the problem.

If I understand your plots correctly, on the top plot you have the zero-vacancy structures near the convex hull at formation energy -100 meV/atom, and the low-vacancy and high-vacancy structures in the upper group at formation energy +300meV/atom? Then in the lower plot from monte carlo, the formation energies of the zero-vacancy structures look close to the DFT energies at around -100 meV/atom, but the formation energies of the low-vacancy and high-vacancy structures are much lower i.e. breaking the predicted convex hull?

Have you tried looking at those structures or seeing what their correlations are?

[pandeydhanshree]

Yes, exactly. You got it correctly. Yes, I have looked at the structures obtained from MC simulations. Overall, I see the clustering of vacancies. But I have not compared the structures (same Ta/W concentration) from DFT and MC simulations. This is exactly what you are suggesting right? But, irrespective of the above point, can one trust the trend obtained from MC calculations?

[xivh]

Yes, if there are structures predicted below the convex hull, then you should calculate those and refit the cluster expansion. If you have structures that far below the convex hull, then you can't trust the monte carlo results. What you want to see is that at 0K you get your DFT convex hull, and as the temperature goes up, the free energy decreases. The formation energy should not decrease because it means there is a more stable potential ground state there.

[pandeydhanshree]

You mean I should take the final structure obtain from MC (few thousand atoms) and do the DFT calculations and refit the cluster expansion?

[xivh]

Yes, you can try reducing the size of the Monte Carlo (but this may change your results) or importing the structure into CASM in order to make it primitive before DFT.

Another option is to enumerate a large set of test configurations (different volumes or orderings than your existing data), predict them with your cluster expansion, and see if any break the convex hull. Then you can select a few of these to calculate with DFT. If you choose to do this method, I recommend copying your CASM project and enumerating in the new project. Then you can import the enumerated structures back into the original project. It will keep the original project clean, and it will also be faster because having a lot of configurations in the database slows things down.

[pandeydhanshree]

Thank you so much for your suggestion. I will try these possibilities. Best regards Dhanshree

[pandeydhanshree]

Hello Regarding the point "Another option is to enumerate a large set of test configurations (different volumes or orderings than your existing data), predict them with your cluster expansion, and see if any break the convex hull. ", how should I do it. I mean after enumerating different configurations (as new project), what steps do I follow? Do I perform the step "casm select --set 'is_calculated' -o train" with my DFT calculated structures? In short, I am bit confused with what steps to follow when you say " predict them with your cluster expansion" Can you please help me with that. Thank you in advance and regards.

[xivh]

If you have the same bspecs.json, eci.json, etc., then you can do casm query -k clex to get the formation energy predictions per unit cell and plot them against your existing convex hull.

If you copied the old project, then you can try casm query -k clex_hull_dist(CALCULATED) to get the hull distance to your fitting dataset. Check casm query --help properties, I don't remember if it is per atom or per unit cell.

[pandeydhanshree]

Okay. Thanks a lot !!

[darjaved]

Can you explain a bit how to copy the original project and import the new structures to improve the CE fit. I am doing my DFT calculation outside CASM project. with VASP ( not with CASM interface). @xivh @pandeydhanshree

[pandeydhanshree]

Hello @darjaved You can import your VASP structure files using the casm import command. For example, execute the command: casm import -s settings.json -b subdirectories.txt You can see the details for settings.json using "casm import". The subdirectories.txt file contains the paths to your VASP structures that you wish to import. I hope it was helpful. With best regards Dhanshree

[darjaved]

I already have a casm project and now from the MC sumulations I found i need to include more structures, so i will do the DFT calculations of those new structures. Since i have already done the fitting part if i import more structures it will change my whole eci’s. I have already done the weight optimization. How to import new structures so to make sure I don’t have to repeat whole weight optimisation again. Like what does copying old projects mean?

On Tue, 6 Aug 2024 at 2:28 PM, pandeydhanshree @.***> wrote:

Hello @darjaved https://github.com/darjaved You can import your VASP structure files using the casm import command. For example, execute the command: casm import -s settings.json -b subdirectories.txt You can see the details for settings.json using "casm import". The subdirectories.txt file contains the paths to your VASP structures that you wish to import. I hope it was helpful. With best regards Dhanshree

— Reply to this email directly, view it on GitHub https://github.com/prisms-center/CASMcode/issues/351#issuecomment-2270751725, or unsubscribe https://github.com/notifications/unsubscribe-auth/A6AKEKWXHBK3Y5OKI4DJIIDZQCF4PAVCNFSM6AAAAABMBSCIC6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDENZQG42TCNZSGU . You are receiving this because you were mentioned.Message ID: @.***>