LAMMPS: neighbor/ghost cutoff larger than r_max in pair_style mace

[sungs368]

Hi,

Thank you for developing MACE — it has been extremely useful for my work.

I have a question about the LAMMPS interface of MACE, specifically the neighbor/ghost cutoff requested by pair_style mace (in my case pair_style mace/kk). When using a pre-trained foundation model, the neighbor/ghost cutoff used by LAMMPS appears to be significantly larger than the model’s r_max. This leads to very large neighbor counts and, as a result, poor performance (especially on GPU).

My setup / observation:

Model:

• Foundation model: mace-mpa0-medium (LAMMPS exported model)
• Model r_max reported by LAMMPS: 6 Å

LAMMPS run:

• Using KOKKOS/GPU with pair_style mace/kk (perpetual neighbor list)
• GPU environment:
  • NVIDIA RTX 4070 Ti
  • CUDA 12.2
• LAMMPS prints (Neighbor list info):
  • master list distance cutoff = 14
  • ghost atom cutoff = 14
  • (additional context: binsize = 14, etc.)

I attached a screenshot of my LAMMPS log output showing these values.

My questions:

1. Is it expected/intentional that the LAMMPS neighbor/ghost cutoff becomes ~14 Å even when the model r_max is 6 Å?
2. Is there any recommended way to align/reduce the effective LAMMPS neighbor/ghost cutoff closer to r_max (e.g., settings, compilation options, or a different pair_style option)?
3. If this is a known limitation of the current LAMMPS interface, is there any planned fix or best-practice workaround for improving performance?

For reference, I saw similar reports/discussions here (but I wanted to provide my own numbers/log output as well):

• Issue LAMMPS pair_style mace uses neighbor/ghost cutoff larger than r_max #1315
• Discussion Too slow speed for lammps gpu #562

Thanks again for your time and for making MACE available.

[ilyes319]

Hey, the mace lammps pairstyle is deprecated, is this is also a problem of the MLIAP lammps interface?

[sungs368]

Thank you very much for your reply — I really appreciate it.

At the moment, I am using a GPU-enabled LAMMPS build, and I have not yet tried building or using the ML-IAP LAMMPS interface.

I was wondering if there is any known way to address or mitigate the cutoff behavior in the current setup, or if there is no practical solution for this issue aside from switching to ML-IAP.

Thanks again for your time and help.

[gabor1]

not really with the old pairstyle - this is a major reason for switching to MLIAP. But you might try the Symmetrix package, which works for the older MACE models.

[sungs368]

Thank you very much for your reply once again.

I have rebuilt and am currently using the ML-IAP interface, which resolved the cut-off issue. However, the performance has not improved, so I will also try using the Symmetrix package as you suggested.

I truly appreciate you taking the time to respond despite your busy schedule.

Best regards

[ilyes319]

The performance should be much better because of cuequivariance. Did you make sure to install the cueq kernels and compile the model on the GPU?

…

________________________________ From: sungs368 ***@***.***> Sent: Wednesday, February 11, 2026 6:50:54 AM To: ACEsuit/mace ***@***.***> Cc: Ilyes Batatia ***@***.***>; Comment ***@***.***> Subject: Re: [ACEsuit/mace] LAMMPS: neighbor/ghost cutoff larger than r_max in pair_style mace (Discussion #1358) Thank you very much for your reply once again. I have rebuilt and am currently using the ML-IAP interface, which resolved the cut-off issue. However, the performance has not improved, so I will also try using the Symmetrix package as you suggested. I truly appreciate you taking the time to respond despite your busy schedule. Best regards — Reply to this email directly, view it on GitHub<#1358 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ALTF4VZHFMTU46JOPVLBP7D4LLGM5AVCNFSM6AAAAACT3HLXWWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNZWGQ4TMOA>. You are receiving this because you commented.Message ID: ***@***.***>