Skip to content

Releases: mdi-group/mace-field

General learning of electric response in inorganic materials

03 Dec 14:56
0886c7f

Choose a tag to compare

State of the MACE-Field package used for preprint: https://arxiv.org/abs/2508.17870.

Here we provide the MACE-Field models trained for this paper:

  • MACE-field-mp-0b3-medium-mh.model mh finetuned on: MPtraj replay set (10000) of energies, forces and stresses; 2500 MP-Ferroelectric polarisations; and ~6000 MP-Dielectric BECs and polarisabilities.
  • MACE-field-dielectric.model trained on ~6000 energies, BECs and polarisabilities from the MP-Dielectric dataset.
  • MACE-field-ferroelectrics.model trained on energies, forces and polarisations from the MP-Ferroelectric dataset.
  • MACE-field-BaTiO3.model trained on energies, forces, stresses, polarisations, BECs and polarisabilities from AIMD trajectories for BaTiO3 (Allegro-pol dataset).
  • MACE-field-SiO2.model trained on energies, forces, stresses, polarisations, BECs and polarisabilities from AIMD trajectories for alpha quartz (SiO2) (Allegro-pol dataset).

NOTE: These models are compatible with this release version using the ScaleShiftFieldMACE model. These no longer work with the latest MACE-Field code. So please use the older package code tagged here for these models.

What's Changed

New Contributors

Full Changelog: https://github.com/mdi-group/mace-field/commits/v1.0