questions about units convert

[yuchanpei]

Dear developers,
I am trying to train a nequip potential with data in extxyz format. I use the following code to construct the dataset in jupyter notebook：
E = float(energy_dict[pdbid])
sdfmol = ase.Atoms(symbols_str,positions=xyz_tuple_list)
calculator = SinglePointCalculator(sdfmol, energy=E)
sdfmol.calc = calculator
io.write("test_ligand_pocket_with_spcalc.xyz",sdfmol, format='extxyz', append=True
The unit of energy I use is kcal/mol. Should I convert the unit from kcal/mol to eV when I construct my own dataset?
When I use kcal/mol as an energy unit to train a nequip, should I use energy_units_to_eV in NequIPCalculator.from_deployed_model() to convert the energy unit since I use the ASE dataset loader to load my data?

Thank you in advance for any help you might be able to provide.

Best regards,
Cristina

[Linux-cpp-lisp]

Hi Cristina,

Thanks for your interest in our code!

In general nequip is agnostic to units; the trained model will predict in whatever units the data is in. (Of course, the data must be self-constant and all in the same units.)

A very important caveat to this is dataset: ase, which inherits the unit conversion behavior of ase.io.read. This can be a bit subtle: if your data is in kcal/mol, for example, but in an extxyz file, ASE can't know that and will assume it is eV (ASE units) and pass it along to nequip, yielding a nequip model with kcal/mol units. However for some formats (mostly those coming out of DFT codes), ASE "knows" the units of the data and will convert them to eV and Angstrom in ase.io.read. Thus you will get a nequip model in eV and Angstrom. You can easily check what ASE is doing with your files by playing with ase.io.read in a Python REPL.

For running simulations in LAMMPS, it is important to use the correct LAMMPS units to match those predicted by the network (which are those of the data)--- quoting the tutorial:

CAUTION: the reference data here are in kcal/mol for the energies and kcal/mol/A for the forces. The NequIP model will therefore also be predicting outputs in these units. We are therefore using units real in LAMMPS (see docs). If your reference data are in other units, you should using the corresponding units command in LAMMPS (e.g. if you use eV, A then units metal would be appropriate, which would then also change time units from fs to ps).

For the ASE calculator, NequIPCalculator, we offer the energy_units_to_eV in case you trained a model that is not in eV, but that you want to use to run MD in ASE, which requires it to be in eV. NequIPCalculator is not relevant for dataset: ase or training, just inference.