Error in MoS2 forcefield implementation in QuantumATK

[AsifShah]

Dear Admin,

The actual MoS2 forcefield given in the original paper "https://www.frontiersin.org/journals/nanotechnology/articles/10.3389/fnano.2022.1034795/full#supplementary-material" is quite different from what is implemented in QuantumATK ReaxFF_CHOSMoNiAuTi_2022. Upon inspecting the implementation in QuantumATK against with original paper it, I found multiple lines have errors and one line is completely missing.

When using this forcefield it gives so many distortions in MoS2. Kindly, check into this or please let me know if I am missing anything.'

regards
Asif

[Anders Blom]

Would you mind sending me the specific differences by private message here on the Forum? We'll be happy to take a look at it.

[Anders Blom]

Ok, I can confirm there is a difference, but to me it looks like it's only one line difference, viz. one of the torsions is left out (we have 47, the article has 48). Perhaps the article was updated after it was published. But you said multiple lines, I don't see that, so please share your differences.

Note that you can easily use the ReaxFF version from the article in QuantumATK, by downloading and saving it as a text file, and then use the ReaxFFPotential class (https://docs.quantumatk.com/manual/Types/ReaxFFPotential/ReaxFFPotential.html) to define the potential set. Then you can also directly compare if the article version gives stable structures.

It would also be helpful for us if you can share some scripts that show that the QuantumATK version give these distortions, and then we can verify if the article version works better.

[AsifShah]

Dear Anders Blom,

Sure. Please find the attached pics highlighting errors. The yellow lines indicate places where mismatch is found between Reax parameters from article (left side) and QATK implementation (Right side). Also, torsion has one missing line and Angles also misplaced or missing line.
I am also attaching a original reactive forcefield (with parameters from article) file but unfortunately it gives some runtime error. Please help fix it if possible.

PS: This is a quick manual check I did. There might be other errors as well.

Thanks
regards
Asif Altaf Shah

[Anders Blom]

Some of these are comparing the wrong lines with each other, there are two different lines in the file starting with "4 4 7", and the other one matches. And the "4 4 4 4" line is correct, just has truncated decimal zeroes in one case. I can't even see a line with "1 3 1 2" in the current version of QuantumATK, so it seems this was partly fixed after your release (I checked in W-2024.09), but I do agree on the 48th torsion still missing even in W.

However, now things get interesting. To use the original file in QuantumATK, just save it as a text file and use file="reaxff.txt" instead. If you do this, however, we discover that the file is not correct. The missing torsion angle "0 4 9 0" cannot be used because there is no particle 9. Maybe it should be 8, but trying to "fix" a ReaxFF potential like that is most likely not going to work.

So, we now have 3 potentials, all slightly different:
1) The one shipped with whichever version of QuantumATK you use
2) The original from the paper
3) The one shipped with QuantumATK W-2024.09

You say that 1 gives weird structures, but we also cannot use 2 (someone should maybe notify the authors...). So hopefully 3 is actually the correct one, I have attached it to this post and shown how to use it in the Python script.

Finally, I ran your script, and yes it does bend the structure, but you have an oxygen molecule there, so I am not too surprised. I am also somewhat skeptical to the quality of the potential as it makes very long Mo-S bonds, 2.65 Å compared to 2.42 Å in databases for MoS2. Maybe this is an effect of the fact that the potential they publish is invalid, and in reality they used a different one for the paper (which they don't want to share...?), but clearly the forcefield as published does not appear to work well for MoS2, and the paper also never checks this (they have a few other bond comparisons with DFT, but not Mo-S). I would also ask the authors why they state that they build on the "force field based on ReaxFF Mo/S/O/H (Islam et al., 2014)" when the paper by Islam has nothing to with Mo/S...

[AsifShah]

Dear Anders Blom,

Thanks a lot for your response.
Yes, the Mo-S bonds are overestimated. Also, though there is O2 molecule but the distortion is high than what is obtained from DFT.

I tested the script u shared on Gold and the results do not make sense. The Au atom does not bond at S vacancy which is clearly in contrast with DFT. (See attached pic)

I think you are right the authors perhaps do not want to share the real forcefield. This one looks useless.

[Anders Blom]

MoS2 2D material would be a nice candidate to fit an MTP for, and then one can add additional atoms as needed. Not trivial, but probably a lot better than this potential, yes. The article makes references to a few other existing ReaxFFs for MoS2 (and, notably, claims to improve on them: Ostadhossein et al., 2017; Hong et al., 2017; Chen et al., 2020) - perhaps you can test those, provided they supply the parameters.