Functionality for outputting ACE "descriptor"

[braxtonowens]

Along with training interatomic potentials, "physics-inspired atomic descriptors can be used to rationalize the relationship between atom configurations and material properties" Physics-Inspired Structural Representations for Molecules and Materials

Ideally, like the Dscribe python library, where given an ASE ATOMS object containing positional and atom-type information, a per-atom ACE description an be generated.

[cortner]

@wcwitt I think you've worked with per-atom descriptors. I'm still on holiday and without working laptop. Would you be willing to help? Maybe we can get more input on the kind of interfaces we should provide...

[wcwitt]

This is possible, although we don't have a good tutorial yet. Do you want to use ASE atoms specifically, or would a pure Julia example be sufficient?

[wcwitt]

@braxtonowens, that last question was for you. Is the ASE atoms part important, or would you be happy providing an xyz file, for example?

[braxtonowens]

Oh sorry, xyz files are fine.

[braxtonowens]

@wcwitt Just checking if there have been any developments regarding the extraction of a per-site description. As mentioned here: https://github.com/ACEsuit/ACE1.jl/discussions/37; A temporary script to do this with ACE1.jl would suffice. Thanks.

[wcwitt]

Sorry - ACEfit is still not quite ready for use, but I’ll try to send you a temporary script tomorrow that does what you need

[wcwitt]

@braxtonowens, can you try this and let me know how it goes? It runs for me but I haven't tested it carefully - we may need to iterate a few times.

One-liner to set up the environment:

using Pkg; Pkg.activate("."); pkg"registry add https://github.com/JuliaRegistries/General"; pkg"registry add https://github.com/JuliaMolSim/MolSim.git"; pkg"add ACE1pack, ACE1, JuLIP, IPFitting, ASE"

Dataset: Si_tiny.zip

Script:

using ACE1pack

dataset = JuLIP.read_extxyz("Si_tiny.xyz");

basis = ace_basis(species = [:Si],
      N = 3,        # correlation order = body-order - 1
      maxdeg = 12,  # polynomial degree
      r0 = 2.3,     # estimate for NN distance
      rin = 0.1,
      rcut = 5.5,
      pin = 2)      # no need to change

descriptors = []
for atoms in dataset

    # compute average descriptor for each structure
    descriptor = zeros(length(basis))
    for i in 1:length(atoms)
        descriptor += site_energy(basis, atoms, i)
    end
    descriptor /= length(atoms)

    push!(descriptors, descriptor)
end

[braxtonowens]

Thanks for the script. I tested this on my machine and for a singular file it seems to work. I however attempted to run this on two separate remote ubuntu servers hosted at BYU and they both resulted in seg faults at the line dataset = JuLIP.read_extxyz("my_aluminum.xyz");

I will continue testing with my own machine, I just thought I would let you know. If you would like any more feedback let me know.

[wcwitt]

That's not good - does my example run on those machines, and if so would you be willing to send me your file that fails?

[braxtonowens]

Sorry, it just turned out that the segmentation fault was caused from an incorrect path being passed to JuLIP.read_extxyz()

I believe I got what I wanted. Thanks for the help! Let me know If you guys get any other algorithms finished soon like LODE or MACE.

[wcwitt]

Great! Please do let me know if you have any more difficulty, as I will eventually turn this into an example for the website