Inconsistent force predictions between `tensorpot` and `pyace` backends

[Alexandrina-Chen]

Dear python-ace developers,

I trained an ACE potential (output_potential.yaml) using the tensorpot backend and wanted to use the resulting potential with the ASE pyace calculator.

When plotting F_pred vs F_ref on my test dataset, I noticed that the resulting force MAE and RMSE differ significantly from those reported in log.txt during training.

To ensure this was not an issue in my own Python code, I performed a no-fit run using:

pacemaker input.yaml --no-fit --potential output_potential.yaml

The input.yaml file is identical to the one used during training, except:

• The backend is set to pyace instead of tensorpot
• The filenames are set to fitting_data_info.pckl.gzip and test_data_info.pckl.gzip (both generated during training)

This run gave me the same forces as my own ASE calculator, confirming consistency between them. However, a subset of forces predicted by pyace differs noticeably from those predicted by tensorpot.

I attached a minimal dataset for reproducing this issue:

• output_potential.yaml – the trained potential
• difference_data.pckl.gzip – ASE Atoms, reference energies, forces, and positions
• difference_data_pred_tensorpot.pckl.gzip – predictions by tensorpot
• difference_data_pred_pyace.pckl.gzip – predictions by pyace
• input.yaml – configuration file used
• Two scatter plots of F_tensorpot vs F_pyace

I am using:

• python version: 3.10.13
• pacemaker/pyace version: 0.2.7+172.gd1c213a
• ace_evaluator version: 2023.11.25

The file:
example.tar.gz

Please let me know if additional information is needed to reproduce this issue.

---

Questions:

1. Is this difference between tensorpot and pyace expected, or does it indicate an issue?
2. Would it be better to train the potential directly with backend=pyace to avoid such inconsistencies?
3. If I run MD simulations with LAMMPS, is there any known issue regarding different force predictions?

Thank you very much for your help!

Best,
Sijia