Add HPHT_CH4_H2O free energy profile of proton hopping benchmark#436
Add HPHT_CH4_H2O free energy profile of proton hopping benchmark#436ThomasThevenet wants to merge 1 commit intoddmms:mainfrom
Conversation
ElliottKasoar
added
the
new benchmark
There was a problem hiding this comment.
Hi @ThomasThevenet, thanks for this, and apologies for the delay in reviewing. It's looking great so far!
If you can please share the data needed to run this, or remind me where you have if you already have, I can test it more thoroughly, but I've left some initial comments/thoughts/questions.
Please also try to add the documentation when you can too.
Let us know if anything is unclear!
| write_dir = OUT_PATH / model_name | ||
| write_dir.mkdir(parents=True, exist_ok=True) | ||
|
|
||
| calc = model.get_calculator() |
There was a problem hiding this comment.
| calc = model.get_calculator() | |
| calc = model.get_calculator(precision="low") |
You may need to rebase for this to work, but we've added a change where the default precision is "high" (usually float64) since we generally want this for static tests, but previously if you didn't specify a precision explicitly, it would default to the dtypes we set in models.yml, which tended to be equivalent to "low", and I think make sense since we're doing MD here
| level_of_theory: PBE + D3 | ||
|
|
||
| DF_MAE: | ||
| good: 0 | ||
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of reaction free energies" | ||
| level_of_theory: PBE + D3 | ||
|
|
||
| DF#_MAE: | ||
| good: 0 | ||
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of free energy barriers" | ||
| level_of_theory: PBE + D3 |
There was a problem hiding this comment.
| level_of_theory: PBE + D3 | |
| DF_MAE: | |
| good: 0 | |
| bad: 50 | |
| unit: kJ/mol | |
| tooltip: "Mean absolute error of reaction free energies" | |
| level_of_theory: PBE + D3 | |
| DF#_MAE: | |
| good: 0 | |
| bad: 50 | |
| unit: kJ/mol | |
| tooltip: "Mean absolute error of free energy barriers" | |
| level_of_theory: PBE + D3 | |
| level_of_theory: PBE+D3 | |
| DF_MAE: | |
| good: 0 | |
| bad: 50 | |
| unit: kJ/mol | |
| tooltip: "Mean absolute error of reaction free energies" | |
| level_of_theory: PBE+D3 | |
| DF#_MAE: | |
| good: 0 | |
| bad: 50 | |
| unit: kJ/mol | |
| tooltip: "Mean absolute error of free energy barriers" | |
| level_of_theory: PBE+D3 |
These need to be without spaces to be matched up correctly
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of free energy profiles" | ||
| level_of_theory: PBE + D3 | ||
|
|
||
| DF_MAE: | ||
| good: 0 | ||
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of reaction free energies" | ||
| level_of_theory: PBE + D3 | ||
|
|
||
| DF#_MAE: | ||
| good: 0 | ||
| bad: 50 |
There was a problem hiding this comment.
Genuine question: How much have you considered these thresholds, and are you happy with the idea behind what they represent?
If you're happy with them, great, but if not, we're very happy to discuss their intent further!
| FEP_MAE: | ||
| good: 0 | ||
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of free energy profiles" | ||
| level_of_theory: PBE + D3 | ||
|
|
||
| DF_MAE: | ||
| good: 0 | ||
| bad: 50 | ||
| unit: kJ/mol | ||
| tooltip: "Mean absolute error of reaction free energies" | ||
| level_of_theory: PBE + D3 | ||
|
|
||
| DF#_MAE: |
There was a problem hiding this comment.
I would consider making these slightly more 'human readable'. E.g. you don't need underscores (if it doesn't work, especially with a #, you may need to wrap it in ".
I'd potentially even writing out "Free Energy Profile MAE" etc. since otherwise it's very unlikely people will have any dense of what they are without the tooltips (which of course are there to give more detail, but I think for example DF and DF# are quite confusing on their own).
| import pytest | ||
| from ase import units | ||
| from ase.io import write, read | ||
| #from ase.io.extxyz import read |
There was a problem hiding this comment.
| #from ase.io.extxyz import read |
|
|
||
| starting_frames_dir= ( | ||
| download_s3_data( | ||
| key="inputs/molecular_reactions/HPHT_CH4_H2O/HPHT_CH4_H2O.zip", |
There was a problem hiding this comment.
Apologies if I missed this, but have you shared this with us? If not, could you so that we can upload it?
| opt = LBFGS(atoms, logfile= write_dir / f"{structure_name}.opt") | ||
| opt.run(fmax=0.2) | ||
| #VELOCITIES INITIALISATION | ||
| rng = np.random.default_rng(seed=13) | ||
| MaxwellBoltzmannDistribution(atoms, temperature_K=3000, rng=rng) | ||
| Stationary(atoms) | ||
| ZeroRotation(atoms) | ||
|
|
||
| #OUTPUT SETUP | ||
|
|
||
|
|
||
| traj_path = write_dir / f"{structure_name}.extxyz" | ||
| log_path = write_dir / f"{structure_name}.log" | ||
|
|
||
| def write_frame(): | ||
| write(traj_path, atoms, format="extxyz", append=True) | ||
|
|
||
| #MOLECULAR DYNAMICS | ||
| dyn = NoseHooverChainNVT( | ||
| atoms, | ||
| timestep=0.5 * units.fs, | ||
| temperature_K=3000, | ||
| tdamp=100 * units.fs, | ||
| tchain=3, | ||
| tloop=2, | ||
| logfile=str(log_path), | ||
| ) |
There was a problem hiding this comment.
It's not a requirement, but can I suggest looking at swapping this out for janus_core's molecular dynamics (see examples here: https://stfc.github.io/janus-core/tutorials/python/md.html)?
It will essentially do all of this in a single command (geometry optimisation, setting the temperature distribution, removing rotations, logging).
If you're interested and run into any problems with this, let me know!
|
|
||
| MODELS = get_model_names(current_models) | ||
| BENCHMARK_NAME = "HPHT_CH4_H2O" | ||
| DOCS_URL = "https://ddmms.github.io/ml-peg/user_guide/benchmarks/molecular.html#proton" |
There was a problem hiding this comment.
The last part proton will need to match the title you create in the molecular.rst file documenting this test (lower case, with spaces replaced by -, so you may need to change this.
2 participants
Pre-review checklist for PR author
PR author must check the checkboxes below when creating the PR.
Summary
"HPHT_CH4_H2O" is a new benchmark of the "Molecular Reactions" category which tests the free energy profile of proton hopping reaction in CH4/H2O liquid from molecular dynamics simulations at 3000 K and 22-69 GPa. The associated analysis performs a molecular recognition algorithm to collect reaction coordinate values and compute the free energy profile associated to the reaction : H3O+ + CH4 = H2O + CH5+. Reference profiles were computed at the DFT (GGA) level (PBE+D3).
Currently three metrics are available :
The application displays parity plot for MAE of both free energy of reaction and barrier. Interaction of the user with parity plots dynamically displays the predicted and reference free energy profiles of the selected structure.
Linked issue
None.
Progress
to do:
Prepare reference data
Upload reference data to S3 bucket
Write documentation
Talk about the full pipeline test
Testing
Analysis and application components have been tested locally from precomputed molecular dynamics trajectories. Full MD-based pipeline is not manageable with local ressources and has not been tested.
New decorators/callbacks
No new decorators were introduced. A new Dash callback leveraging clickData was implemented to enable interactive visualization of free energy profiles. Selecting a point in the parity plot dynamically displays the corresponding structure's profiles for all MLIP models.