Question about nequip training and evaluation

[baham2]

Hello,

Thank you for developing and maintaining this package.

I have trained a Nequip model on my periodic system using the attached config.yaml file. My dataset includes trajectories from two different temperatures of the same system. After training, I achieved a validation MAE of 0.0004 eV/atom for energy and a force MAE of 0.02 eV/Å, along with smooth learning curves. Following the training, I used the same config file to evaluate the model on the training and validation datasets (nequip-evaluate --model train_dir/best_model.pth --dataset-config config.yaml --batch-size 1 --output ./output.xyz). However, I have two questions:

1. Although I have the pbc key in my dataset (shape: nframes*3, type: boolean), the pbc is printed as F (pbc="F F F") in the output file. I am wondering if this is a typo or if the pbc is in fact turned off during evaluation (or possibly during training as well)?
2. I have plotted the energy error distribution for each temperature (attached). It appears that the Nequip predictions have a positive offset for the lower temperature and a negative offset for the higher temperature. Is this behavior expected, or could something be going wrong?

Thank you for your help.
config.txt

[Declanwhx]

Hi, i too am getting my PBC for nequip-evaluate as "F F F". Does it really evaluate the MLFF without pbc or do we miss a relevant setting. My training input and my validation input includes pbc "T T T", and of course, we would expect significant difference between energies and forces if pbc are not included in the testing.

[Linux-cpp-lisp]

I'm closing this as outdated since it uses the old code, but please feel free to file again if you face unexpected PBC issues with the new code.

Regarding offsets, this depends a great deal on how you select your per-type offset energies (or "E0s") which is an important consideration for all MLFFs. A good starting point if you can get them are the isolated atom energies from your underlying reference DFT calculations.