Integration of OpenMM

[JMorado]

This PR builds upon previous features introduced in PR #55, implementing the following changes:

• Refactored the code to align its structure with that of md.py, aiming for uniformity, consistency, and improved readability
• Added the possibility to use either a Langevin (NVT) or a Verlet (NVE) integrator
• Implemented restart functionality
• Added functionality to run the OpenMM MD in a temporary directory
• Added option to select the OpenMM Platform
• Added tests

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 92.62295% with 18 lines in your changes missing coverage. Please review.

Project coverage is 68.26%. Comparing base (e809fe8) to head (fb54bd1).
Report is 10 commits behind head on main.

Files with missing lines	Patch %	Lines
mlptrain/sampling/md_openmm.py	91.48%	12 Missing ⚠️
mlptrain/training/active.py	0.00%	5 Missing ⚠️
mlptrain/sampling/md.py	66.66%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #87      +/-   ##
==========================================
+ Coverage   65.10%   68.26%   +3.16%     
==========================================
  Files          50       51       +1     
  Lines        4662     4847     +185     
==========================================
+ Hits         3035     3309     +274     
+ Misses       1627     1538      -89     

Flag	Coverage Δ
python-3.9	63.31% <19.67%> (-1.57%)	⬇️
python-3.9-mace	68.10% <90.98%> (+3.08%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[juraskov]

The code looks good, thanks! I have only a minor comment regarding triggering a NotImplemented error. If it is too much work to implement, we can merge the code without it.

[Hanwen1018]

Dear all, I just have one comment, for now OpenMM only support MACE, correct me if I am wrong. Do we need to ensure that we have MACE before running OpenMM dynamics?

[JMorado]

Hi @Hanwen1018,

Indeed, OpenMM-ML currently only supports MACE (and ANI, but that MLP is not part of mlp-train). MACE is installed when running ./install_mace.sh (as well as OpenMM and OpenMM-ML). If, for some unusual reason, the MACE package cannot be found, OpenMM-ML by default throws an ImportError:

https://github.com/openmm/openmm-ml/blob/main/openmmml/models/macepotential.py#L136-L143

Additionally, if you install the other environments (./install_gap.sh or ./install_ace.sh), you cannot run an OpenMM simulation:

https://github.com/JMorado/mlp-train/blob/openmm-features/mlptrain/sampling/md_openmm.py#L210-L220

Could you please clarify what you are suggesting?

[Hanwen1018]

Hi @Hanwen1018,

Indeed, OpenMM-ML currently only supports MACE (and ANI, but that MLP is not part of mlp-train). MACE is installed when running ./install_mace.sh (as well as OpenMM and OpenMM-ML). If, for some unusual reason, the MACE package cannot be found, OpenMM-ML by default throws an ImportError:

https://github.com/openmm/openmm-ml/blob/main/openmmml/models/macepotential.py#L136-L143

Additionally, if you install the other environments (./install_gap.sh or ./install_ace.sh), you cannot run an OpenMM simulation:

https://github.com/JMorado/mlp-train/blob/openmm-features/mlptrain/sampling/md_openmm.py#L210-L220

Could you please clarify what you are suggesting?

Thank you for clarification. I want to add a checking step before run OpenMM. It is not about whether MACE is installed. I want to check it is the MACE potential used in the _create_openmm_simulation function. As in _create_openmm_simulation, the argument is mlp (belong to mlt.potentials._base.MLPotential), which can also be ACE or GAP, and will make issue.

[JMorado]

I understand your point. In such a scenario, OpenMM-ML might throw an error that isn't very informative. I agree that an additional check would be beneficial. I'll implement that. Thanks!

[juraskov]

This is probably part of the same issue. For example, when I activate the mace environment and specify ACE in AL, it will still train the ACE and the code will fail only after the training is finished - it is a similar issue as with the plumed check when the error was triggered after the training. The environments are quite similar and on my system, the ACE training finished even though I activated a different environment (which is not great tbh). So the MLP type check should be able to resolve also this.