Comparison of MACE in ASE and LAMMPS

[sandy1593]

Hello everyone, Good day

I’ve been running MD simulations on MOF-74 using MACE potentials to benchmark performance. Specifically, I compared simulations in ASE and LAMMPS using the same medium and small mace-mp-0 models. In ASE, the structure remains stable even at 600 K, which aligns with experimental results suggesting MOF-74 should be stable up to that temperature. However, in LAMMPS, the structure becomes unstable, even at 300 K.

For LAMMPS, I used create-lammps-model.py to convert the MACE model into a LAMMPS-compatible pair potential. I'm starting to suspect there may be an issue introduced during this conversion process or in the integration of MACE with LAMMPS itself.
If anyone has insights on properly integrating MACE potentials with LAMMPS, or has encountered similar stability issues, I’d appreciate your input.

Thanks in advance for your help!

[gabor1]

perhaps you could check the energy and forces for a given configuration? that removes some of the differences between the two md runs.

[sandy1593]

Thanks Prof. for your quick response. Energy is different after minimization for ASE and LAMMPS calculations.
Also, I tried to perform calculation of CO2 adsorption energy for Mg-mof-74. I got different energy values for ASE and LAMMPS.
The result obtained using ASE are more believable.
I am still curious to know whether create-lammps-model.py script is working efficiently or not?

[wcwitt]

I am still curious to know whether create-lammps-model.py script is working efficiently or not?

We would be foolish to rule this out entirely, but in the overwhelming majority of cases the explanation is something else. Could be inconsistent timestep or units, differences in the thermostat or minimization algorithms, different initialization, etc.

If you find a single snapshot where ASE and LAMMPS give different energy or forces, that is much more concerning and cause to investigate - that was Gabor's suggestion.

[jfaraudo]

Is there any update on this? I am currently using MACE in ASE but I am planning to move to LAMMPS, so I am curious to know whether everything is self-consistent between MACE implementations in ASE and LAMMPS.

[wcwitt]

As far as we know there is no cause for concern. You will want the ML-IAP interface or symmetrix:

• https://mace-docs.readthedocs.io/en/latest/guide/lammps_mliap.html
• https://github.com/wcwitt/symmetrix