TorchX Potentials

[karolina2]

Hi,

What are the differences between: TorchX_MACE_MP_0_L0_2023, TorchX_MACE_MP_0_L1_2023 and TorchX_MACE_MP_0_L2_2024? According to the QuantumATK website they are all based on one paper by Gábor Csányi's group and co-workers: 'A foundation model for atomistic materials chemistry', where only MACE-MP-0 name is used. Could you provide some details (training sets, levels of theory - whether it was PBE, PBE-D3 or revPBE-D3) about TorchX potentials implemented in ATK?

Kind regards,
Karolina

[Anders Blom]

The differences between L=0,1,2 are described in the reference paper https://arxiv.org/abs/2401.00096 (page 14) and are related to the size of the model, differing by the maximal message equivariance.

A quick flip to page 15 further shows that "The DFT calculations use the PBE exchange-correlation functional with Hubbard U terms applied to some transition metal oxide systems, but no additional dispersion correction".

[karolina2]

Thank you!