No forces are read, but training is running

[wilkek]

I got this output for my simulation, and the property forces is not being read. What does it then actually train on, and is there an easy fix?

2025-09-30 17:44:01.395 INFO: Since ASE version 3.23.0b1, using energy_key 'energy' is no longer safe when communicating between MACE and ASE. We recommend using a different key, rewriting energies to 'REF_energy'. You need to use --energy_key='REF_energy', to tell the key name chosen.
2025-09-30 17:44:01.406 INFO: Since ASE version 3.23.0b1, using forces_key 'forces' is no longer safe when communicating between MACE and ASE. We recommend using a different key, rewriting energies to 'REF_forces'. You need to use --forces_key='REF_forces', to tell the key name chosen.
2025-09-30 17:44:01.406 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.338 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.
2025-09-30 17:44:29.339 WARNING: Failed to extract forces: The property "forces" is not available.

[pdobbelaere]

This feels like a MACE problem, which might not be a Psiflow thing. Can you share the MACE configuration used for training and a few of your training structures? Also, are you using the Psiflow container for model training or your own MACE installation? If possible, share your Python environment.
I imagine it will be something silly like ASE hiding the atomic forces in a SinglePointCalculator and MACE not knowing where to check for them.