LAMMPS pair_style mace uses neighbor/ghost cutoff larger than r_max

[Pourasad]

Hi.

When using pair_style mace in LAMMPS with pre-trained foundation models, the neighbor/ghost cutoff requested by the pair style is much larger than the model r_max, leading to very large neighbor counts and poor performance.

Details

Model
• MACE-matpes-r2scan-omat-ft
• Exported with create_lammps_model.py
• LAMMPS reports: The r_max is: 6.

Observed in LAMMPS:

• master list distance cutoff = 13
• ghost atom cutoff = 13
• Ave neighs/atom ≈ 480

This occurs for:
• N = 250 (L ≈ 16.8 Å)
• N = 1024 (L ≈ 26.8 Å, i.e. L > 2 × r_max)

Increasing system size does not reduce the effective cutoff.

Is the large ghost/neighbor cutoff (~13 Å) intentional in the MACE LAMMPS interface?
Is there a way to align the LAMMPS neighbor cutoff with r_max?

Thanks for your attention in advance.