How to use a checkpoint after an epoch to run simulation

[dominicvarghese]

Hi everyone,

I was using the multihead replay fine-tuning with the MACE-MATPES-0-R2SCAN model on my VASP DFT dataset.

I gave the max number of epochs to be 30. From the current results,

2025-10-30 08:03:06.023 INFO: Epoch 10: head: pt_head, loss=0.00020118, RMSE_E_per_atom=    2.15 meV, RMSE_F=   12.57 meV / A, RMSE_stress=    0.39 meV / A^3
2025-10-30 08:03:08.770 INFO: Epoch 10: head: Default, loss=0.00006202, RMSE_E_per_atom=    2.36 meV, RMSE_F=    3.84 meV / A, RMSE_stress=    0.63 meV / A^3
2025-10-30 09:25:13.460 INFO: Epoch 11: head: pt_head, loss=0.00020515, RMSE_E_per_atom=    2.19 meV, RMSE_F=   11.39 meV / A, RMSE_stress=    0.39 meV / A^3
2025-10-30 09:25:16.254 INFO: Epoch 11: head: Default, loss=0.00006059, RMSE_E_per_atom=    2.36 meV, RMSE_F=    3.78 meV / A, RMSE_stress=    0.63 meV / A^3
2025-10-30 10:47:38.783 INFO: Epoch 12: head: pt_head, loss=0.00020811, RMSE_E_per_atom=    2.21 meV, RMSE_F=   10.53 meV / A, RMSE_stress=    0.39 meV / A^3
2025-10-30 10:47:41.572 INFO: Epoch 12: head: Default, loss=0.00005915, RMSE_E_per_atom=    2.35 meV, RMSE_F=    3.73 meV / A, RMSE_stress=    0.63 meV / A^3

I have the default head having the error which is ready for deployment. Is it possible to package the current .pt file to be converted to a .model file and be used in ASE as a calculator?

Thanks
Dominic

[dominicvarghese]

I was able to get the model following the discussion ACEsuit/mace#564.

I was having another doubt with replay fine-tuning. My dataset was obtained by using PBEsol pseudopotentials in VASP without spin polarization. Is it okay to then use the pretraining replay data which was R2SCAN in this case to do the fine-tuning?
Does the mismatch in these cause an error with the model learning the real behavior?

[ilyes319]

Sorry for the long delay, I watch the discussion of the mace repo more than here. It is fine to use another level of theory for the finetuning set, just be careful to change your E0s.