Can we create MTP with dopant atom included?

[Lim changmin]

Hello, I have some questions regarding MTP training with dopant atom attached.

We are planning to put carbon dopant in GeTe amorphous structure with different doping density.

However, since there are no force field related to C, Ge, Te, we decided to create MTP.

Here is the question, to create MTP, as far as I know, making initial training set is essential.

But since there are no initial training set regarding to different carbon doping density in GeTe, I wonder how should we proceed with the first step of creating MTP.

Is there any tips for it?

[Anders Blom]

Step 1 would be to make an MTP for GeTe. This should be relatively straightforward following the standard crystalline training protocol.
Step 2 is to use active learning to extend this MTP to a-GeTe
In step 3 you can then add C to either or both several GeTe crystal supercells and/or selected snapshots from the a-GeTe MD trajectory, and retrain on these. If you need to consider different kinds of C positions (interstitial, substitutional etc) make sure to cover all possibilities.
Finally, I would do a step 4 where you repeat the active learning for a-GeTe with C included.

Now, I am aware that the tutorials for MTP on the website are outdated. We are working on new ones which should be published very soon.

[pshinyeong]

Hello, I am working on the same thing and I have some questions regarding MTP with dopant atom

1. In step 4, you mentioned repeating active learning for a-GeTe with carbon included. Can I skip creating a base MTP with carbon and go directly to active learning after step 3?

2. Before Creating C-doped a-GeTe in step 3, I should use the MTP trained from Step 2 to create amorphous GeTe using Melt Quench method for more accurate a-GeTe?

3. Is it possible to use the defect training protocol with carbon as an interstitial or substitutional dopant in step 1 to skip step 3 altogether?



Thank you in advance for your guidance!

[Anders Blom]

Didn't notice your follow-up question, sorry.
1. I guess you mean after step 2. Yes, that is a possibility for sure, as long as you start the C atom in many different positions.
2. Yes, and this is true whether you skip 3 or not.
3. Probably, as long as you do step 4. The point of 1+2 is that it's a simpler system, and I prefer to build something complex on something simpler.

[Lim changmin]

Thank you for your reply.

Again, I have some other questions related to previous ones.

1. I want to train crystal GeTe with ratio 1:3 for Ge and Te each. However, from materials project, there was no Ge1Te3 crystal structures. So can I just use GeTe instead for MTP training to get Ge1Te3 MTP?

2. From your answer about putting C atom into GeTe, can I just use amorphous C doped GeTe with packmol as initial training set and skip step 1, 2(only getting GeTe MTP) ? I am little bit confused that why should I create only GeTe MTP first before training atom-doped GeTe.

Thank you

[Lim changmin]

Oh, and other question regarding to previous ones, if I were to investigate amorphous structure only, then can I skip step 1(crystal mtp) and go straight to step 2(amorphous mtp)?

[Anders Blom]

Not really, because you need a base MTP to run the active learning algorithm.

[Lim changmin]

Dear Anders Blom,

Thank you for your kind response.

I have successfully created the GeTe MTP and am now working on incorporating carbon atoms into the GeTe structure using the "defect training set" or "defect formation set".

However, as far as I understand, both methods seem to allow the introduction of only a single interstitial atom. I may be mistaken, but I wanted to check if there is a way to modify this feature to accommodate multiple interstitial atoms.

My goal is to introduce at least three carbon atoms into the GeTe structure. Would it be possible to modify the "Defect" feature to accommodate multiple interstitial atoms? Alternatively, would you advise modifying the atomic coordinates directly in the script?

I would greatly appreciate any guidance you could provide on this matter.

Thank you for your time and assistance.

Best regards,
Changmin

[AsifShah]

Dear Lim changmin,

If I am not mistaken,  I guess yes you can introduce many C atoms into your crystal structures. For instance, if your supercell has 20 atoms, you can do the following by creating new configurations as:
1. Replace some Ge with C in supercell
2. Replace some Te with C in supercell
3. Replace some Ge and Te with C in supercell

Depending on your accuracy then you can use all these configurations and do some active learning at high temperature with initial coefficients from your already generated MTP. For further accuracy you can then also do same on amorphous structures.

[Lim changmin]

Thank you for your response.

I am currently working on training an MTP exclusively for amorphous structures, as there is no corresponding crystalline phase.

I have some questions regarding Step 1 of the standard crystalline training protocol. For training amorphous structures, active learning requires both MTP training and a training set generation function. In this process, MTP training is typically initialized using the random displacement function applied to a crystalline structure.

My question is: Can the crystal random displacement function be used for amorphous structures as well? Or is it strictly applicable to crystalline phases?

I am asking because, in the tutorial (https://docs.quantumatk.com/tutorials/mtp_hfo2/mtp_hfo2.html), the workflow in Step 1 explicitly refers to the "bulk phase." I am a bit confused by the term "crystal" in "crystal training random displacement function" and whether it applies to amorphous systems.

I would appreciate any clarification on this.