7
$\begingroup$

I am working on a project that requires generating a large dataset of Co3O4 and Fe3O4 structures under various temperatures and pressures using LAMMPS. The structures will be used for training machine learning models. However, I have not been able to find reliable force fields for these two transition metal oxides in the literature or online resources.

Therefore, I would like to consult the community:

Are there any proven force fields available for modeling Co3O4 and Fe3O4 in LAMMPS? The force field needs to yield reasonably accurate structures, but some minor discrepancies in lattice parameters or bond lengths are acceptable.

My expertise is mainly in running LAMMPS simulations, but I have limited experience in developing force fields for new materials. I would greatly appreciate any comments, suggestions, or references to relevant works. Thank you very much!

$\endgroup$

1 Answer 1

5
$\begingroup$

I am afraid your project to develop a new force field will not work from LAMMPS. The main reason is that, LAMMPS is, in simplest terms, not a first-principles simulation engine, but a molecular dynamics engine. In other words, if there is no force field defined for a material, LAMMPS cannot compute forces and energies from quantum mechanics and approximations, because it just does not have that capacity. LAMMPS's main job is to take existing force fields and propagate dynamics in time according to equations of motion.

You need either a quantum chemical code (e.g. ORCA), or a density functional theory code (e.g. VASP, Quantum Espresso, CP2K, or FHI-aims). The examples are not exhaustive, there are others for each.

Co3O4 and Fe3O4 will be particularly challenging for machine learning since they will have multiple sets of forces and energies for a given set of atomic positions based on the spin configurations. To my knowledge, there does not yet exist a well-tested machine-learning framework that can handle that.

$\endgroup$
3
  • $\begingroup$ Thank you for the helpful suggestions! To clarify, my goal is to use machine learning to map between structural features and spectroscopic properties. The plan is to generate a large and diverse dataset of structures using LAMMPS, and then calculate the spectra from these structures using first-principles methods. $\endgroup$
    – li jiwei
    Commented Apr 27 at 12:35
  • $\begingroup$ Given the challenge in finding accurate force fields, I'm wondering if it's critical to have highly precise structures for my machine learning project that aims to predict spectra from structures. Would it be valid to use approximate force fields to efficiently generate a large quantity of training data, and then let the machine learning model learn to compensate for the inaccuracies?Thank you again for the insightful advice! $\endgroup$
    – li jiwei
    Commented Apr 27 at 12:36
  • $\begingroup$ Hi @lijiwei, I think I can say a few things about these follow-up questions. But this StackExchange has a rule for one question per post. Please ask a follow-up question in a new post. A suggested framing of such questions for you can be something like "I'm trying to do [your goal]. I plan to generate a diversity of structures using [method]. What are benefits / drawbacks / alternatives?" If you do ask these questions, I would also advise you to clarify what you mean by "let the machine learning model learn to compensate for the inaccuracies" as I suspect something may be fuzzy there. $\endgroup$ Commented Apr 27 at 12:43

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .