Questions about training data preparation

[turbosonics]

Hi all,

I have some questions about training data preparation. These questions may sounds silly, but I couldn't find anything useful regarding these problems from NequIP papers I read so far...

I wish to perform NequIP training for condensed phase simple cubic crystal which involves diffusion of a light element. I hope NequIP captures diffusivity, density, and modulus correctly.

So, I performed some VASP DFTs that can be categorized as four groups as following:

1. DFT-1: 30k frames of 3000K AIMD
2. DFT-2: 38k frames of 300K AIMD
3. DFT-3: 40 geometry optimizations (=40 frames) for volume-energy relation (eqn of state).
4. DFT-4: 9 frames of NEB for cation-vacancy diffusion.

1. Under/over-representation.
   I combined the result DFT data of all four groups to a single file to perform NequIP training. DFT is VASP but I converted them to extxyz. In this case, data amount from DFT-1 and DFT-2 greatly outnumbers compared to those of DFT-3 and DFT-4. It is ~68k frames vs 49 frames.
   Maybe my understanding towards NequIP is not enough yet so I may be wrong, but what I'm worrying is that the trend of data from DFT-3 and DFT-4 is overwhelmed by DFT-1 and DFT-2 because of data amount difference (number of frames).
   Is this worry a valid one? Or is this something I don't need to worry about, and NequIP will process DFTs regardless of DFT type (AIMD, NEB, Geo Opt) and frame number?
   Or, if my worry is valid, is there any hyperparameters that I can tweak to resolve this problem?

2. NPT AIMD for NequIP training reference data?
   This question is related to question 1. If ~40 frames of equation of state (=40 geo opts) is not enough, then would it be better to run ~40ps NPT AIMD with huge pressure to get volume-energy relation for ~20k number of frames? I hope my parameters to be prepared for expansion/compression event for better mechanical property description. My loss functions set up is energy 1 (PerAtomMSELoss), stress 1, and force 1.

3. Input data format preference (if any).
   Instead of using OUTCAR file directly, I convert them to extxyz using ase. I think extxyz is much easier to process data, instead of playing with OUTCAR file. But is this something recommended for NequIP? Or would reading VASP raw data file (=OUTCAR file) be more preferred by NequIP for any unknown reasons?