Recommended Guideline for Creating a Ternary MTP by Combining Binary MTP

[Lim changmin]

Hello everyone,

I am seeking guidance on the recommended workflow for developing a Moment Tensor Potential (MTP) for the amorphous Ge-Se-Te ternary system.

A key challenge is the lack of existing DFT or experimental data for the amorphous ternary phase, which makes direct training and validation difficult. To address this, I have formulated a bottom-up strategy which involves creating and validating the constituent binary potentials first:

1. Develop and validate a Ge-Te MTP against published research.

2. Develop and validate a Ge-Se MTP against published research.

3. Develop and validate an Se-Te MTP.

4. Finally, combine these three validated binary MTPs to describe the full ternary system.

My underlying hypothesis is that if a combined potential accurately reproduces the structural properties (e.g., RDF, CN, angular distribution) of the constituent binary systems, it will also provide a reliable description of the ternary system. I have already successfully developed the Ge-Te and Ge-Se potentials, and their results show excellent agreement with previous studies.

With the recent release of QuantumATK X-2025.06, I noticed several improvements to MTP training, including a new 'Load mtp file' block and the ability to save the fit in an MTPParameters object.

This leads to my main questions:

1. Is my proposed methodology—developing and then combining three separate binary MTPs—a scientifically valid and recommended approach for creating a potential for a ternary system like this?

2. Does QuantumATK 2025.06 provide a specific tool or a recommended workflow (perhaps using the new features mentioned above) to combine these separately trained MTPs into a single, functional potential for the Ge-Se-Te system?

Thank you in advance for your support and insights.

[Anders Blom]

Quick check, do you really need Ge-Se-Te or Ge-Sn-Te (commonly, GST)? If it's the latter, we already have such a potential provided in the package.

Otherwise, you are in for a ride, but a potentially very rewarding one!

Your methodology is sort of correct, but you don't fit separately for 1-3, you combine those steps into one. That is, you include a bunch of basic crystal structures for all combinations in one training set (which you can test against the crystal experimental data). Then, instead of your step 4, you use active learning on a set of random alloys (and/or amorphous structures) of Ge-Se-Te to refine the potential for the full system. Ideally, for different stoichiometries! This is similar to the workflow we used for TiSi2 in the tutorial below, just a bit more complex for your case since you have 3 elements.

https://docs.quantumatk.com/tutorials/mtp-training-c-am-TiSi/mtp-training-c-am-TiSi.html