facebookresearch
Adaptation of my research to UMA?
What would you like to report?
In my research, i was using the eq2_153M_ec4_allmd.pt model to predict the adsorption energy of some small molecules on a slab of Ni(111). My idea was to calculate it using the equiformer model and the fine tune the model for my specific models, reducing the MAE. However, as i understand, this model is not working anymore of Fairchem_v2. Can i do the same thing with UMA? And if so, i should just use the "uma-sm" checkpoint?
I asked the same question on the huggingface forum, but i can't see it anymore, so i thought maybe it'd be more appropriate to ask this in here.
Yup sorry about that, we closed the forum so we can consolidate everything here.
Correct, the older models are no longer supported in v2. Fortunately uma-sm is a plug in replacement here. See the example below:
from ase.build import fcc100, add_adsorbate, molecule
from ase.optimize import LBFGS
from fairchem.core import pretrained_mlip, FAIRChemCalculator
predictor = pretrained_mlip.get_predict_unit("uma-sm", device="cuda")
calc = FAIRChemCalculator(predictor, task_name="oc20")
# Set up your system as an ASE atoms object
slab = fcc100("Cu", (3, 3, 3), vacuum=8, periodic=True)
adsorbate = molecule("CO")
add_adsorbate(slab, adsorbate, 2.0, "bridge")
slab.calc = calc
# Set up LBFGS dynamics object
opt = LBFGS(slab)
opt.run(0.05, 100)
The one difference here though is the energy returned from this is a total energy, so if you want to compute adsorption energy you should use the same code and run a bare slab relaxation as well. Hope this helps!
@mshuaibii Ah, i see. So, i don't need to calculate the energy of the isolated adsorbate as well?
Regarding my original question, i have just two more questions:
- In the fairchem_v1, the energy was already the adsorption energy, within a reference reaction, which was:
x CO + (x + y/2 - z) H2 + (z-x) H2O + w/2 N2 + * -> CxHyOzNw*
As i understand, in fairchem_v2, using UMA and the task_name = oc20, i don't have to consider the reference reaction anymore. I just need to do E(Adsorbate + Slab) - E(Slab). Is that correct?
- In my original research, i was fine tuning my model, using the eq2_153M_ec4_allmd.pt checkpoint to generate a config.yml. And with that config.yml file, i used the 'main.py' function to run the finetuning on the database i created (which i splited in train.db, val.db and test.db). In fairchem_v2, is it the same procedure, but now using uma_sm.pt? Is the procedure of finetuning a model to a certain database different now?
Thank you!
Sorry you still definitely need to include some energy corresponding to the adsorbate reference. If you want to be consistent with previous models and the OC20 literature you can get those directly from Table 5 in the OC20 paper:
So it would be E_ML(adsorbate+slab) - E_ML(Slab) - gas_ref(Adsorbate). For example, E(H2O) = (2*-3.477 -7.204) = 14.158eV
@mshuaibii Ok, so, i am running the calculation with
from ase.build import fcc100, add_adsorbate, molecule from ase.optimize import LBFGS from ase.io import read, write from fairchem.core import pretrained_mlip, FAIRChemCalculator
predictor = pretrained_mlip.get_predict_unit("uma-sm", device="cpu") calc = FAIRChemCalculator(predictor, task_name="oc20")
slab = read('pt-c2h6.xyz')
slab.calc = calc
opt = LBFGS(slab) opt.run(0.05, 100)
write('pt-c2h6-free-2.json', slab, format='json')
And then i ran the same calculation, but only with the bare slab. I get these two values, and, for the adsorbate, i use the values you mentioned. So, for C2H6, i'd use (2*(-7.282) + 6*(-3.477)), right?
Would that be consistent with the UMA model?
@mshuaibii ok, thank you. I think the my Eads result still needs improvement, but i'll try estimating the energy with AdsorbML, maybe. I saw that there is now a "uma-s-1" model. Is it a improvement on "uma-sm"?
uma-s-1 is the same as uma-sm, we just renamed to introduce versioning. Can you share an example script of your use case with the results + what you expect? How big is the error you are seeing? Happy to spot check to make sure nothing is missing or off.
Thank you @mshuaibii ! Here is the script of the Pt + adsorbate system:
from ase.build import fcc100, add_adsorbate, molecule
from ase.optimize import LBFGS
from ase.constraints import FixAtoms
from ase.io import read, write
from fairchem.core import pretrained_mlip, FAIRChemCalculator
predictor = pretrained_mlip.get_predict_unit("uma-sm", device="cpu")
calc = FAIRChemCalculator(predictor, task_name="oc20")
slab = read('pt-c2h6.xyz')
slab.calc = calc
constraint = FixAtoms(indices=range(18))
slab.set_constraint(constraint)
opt = LBFGS(slab)
opt.run(0.05, 100)
write('pt-c2h6-2lay.json', slab, format='json')
And here is the bare slab calculation
from ase.build import fcc100, add_adsorbate, molecule
from ase.constraints import FixAtoms
from ase.optimize import LBFGS
from ase.io import read, write
from fairchem.core import pretrained_mlip, FAIRChemCalculator
predictor = pretrained_mlip.get_predict_unit("uma-sm", device="cpu")
calc = FAIRChemCalculator(predictor, task_name="oc20")
slab = read('4ml-9ptcluster.xyz')
slab.calc = calc
constraint = FixAtoms(indices=range(9))
slab.set_constraint(constraint)
opt = LBFGS(slab)
opt.run(0.05, 100)
write('pt-9pt-2lay.json', slab, format='json')
The energy of the first one is -227.174212 eV and the bare cluster is -187.158260 eV. With the value of E(C2H6) = -35.4260 eV of the isolated adsorbate, the adsorption energy is of about -4 eV, but the experimental Eads is -0.2819 ev. Am i missing something on the python inputs? BFGS optimizer gives me the same result. Thank you again for helping me.
Here are the xyz files: pt-c2h6.xyz
44
Lattice="8.394416784003516 0.0 0.0 4.197208392001758 7.269778184901512 0.0 0.0 0.0 47.23325771917803" Properties=species:S:1:pos:R:3:tags:I:1 pbc="T T T"
Pt 0.00000000 0.00000000 20.00000000 4
Pt 2.77185858 0.00000000 20.00000000 4
Pt 5.54371716 0.00000000 20.00000000 4
Pt 1.38592929 2.40049995 20.00000000 4
Pt 4.15778787 2.40049995 20.00000000 4
Pt 6.92964646 2.40049995 20.00000000 4
Pt 2.77185858 4.80099990 20.00000000 4
Pt 5.54371716 4.80099990 20.00000000 4
Pt 8.31557575 4.80099990 20.00000000 4
Pt 1.38592929 0.80016665 22.26321306 3
Pt 4.15778787 0.80016665 22.26321306 3
Pt 6.92964646 0.80016665 22.26321306 3
Pt 2.77185858 3.20066660 22.26321306 3
Pt 5.54371716 3.20066660 22.26321306 3
Pt 8.31557575 3.20066660 22.26321306 3
Pt 4.15778787 5.60116655 22.26321306 3
Pt 6.92964646 5.60116655 22.26321306 3
Pt 9.70150504 5.60116655 22.26321306 3
Pt -0.00000000 1.60033330 24.52642611 2
Pt 2.77185858 1.60033330 24.52642611 2
Pt 5.54371716 1.60033330 24.52642611 2
Pt 1.38592929 4.00083325 24.52642611 2
Pt 4.15778787 4.00083325 24.52642611 2
Pt 6.92964646 4.00083325 24.52642611 2
Pt 2.77185858 6.40133319 24.52642611 2
Pt 5.54371716 6.40133319 24.52642611 2
Pt 8.31557575 6.40133319 24.52642611 2
Pt 0.00000000 0.00000000 26.78963917 1
Pt 2.77185858 0.00000000 26.78963917 1
Pt 5.54371716 0.00000000 26.78963917 1
Pt 1.38592929 2.40049995 26.78963917 1
Pt 4.15778787 2.40049995 26.78963917 1
Pt 6.92964646 2.40049995 26.78963917 1
Pt 2.77185858 4.80099990 26.78963917 1
Pt 5.54371716 4.80099990 26.78963917 1
Pt 8.31557575 4.80099990 26.78963917 1
H 6.74122348 4.12098752 30.40968098 0
H 6.96865908 3.17148090 28.92735405 0
H 6.15545920 4.74331029 28.85021347 0
H 4.44489025 2.88637523 28.71238163 0
H 4.21066085 3.78975603 30.21710220 0
C 4.92324334 3.14528633 29.68489244 0
C 6.26123108 3.83287284 29.46315005 0
H 5.02667482 2.20430380 30.24188399 0
and 4ml-9ptcluster.xyz
36
Lattice="8.394417164447733 0.0 0.0 4.197208582223866 7.2697785143758695 0.0 0.0 0.0 47.23325985984131" Properties=species:S:1:pos:R:3:tags:I:1:forces:R:3 energy=-0.5338616371154785 pbc="T T T"
Pt 0.00000000 0.00000000 20.00000000 4 -0.31235087 -0.16667318 -0.23576574
Pt 2.77185858 0.00000000 20.00000000 4 -0.04428514 -0.27879435 -0.42013121
Pt 5.54371716 0.00000000 20.00000000 4 0.17418820 -0.36001483 -0.35575607
Pt 1.38592929 2.40049995 20.00000000 4 -0.27495486 0.11659463 -0.43486473
Pt 4.15778787 2.40049995 20.00000000 4 -0.00221496 0.00802770 -0.62820607
Pt 6.92964646 2.40049995 20.00000000 4 0.23018552 -0.08997133 -0.55403936
Pt 2.77185858 4.80099990 20.00000000 4 -0.22530563 0.33798230 -0.34853876
Pt 5.54371716 4.80099990 20.00000000 4 0.03447652 0.24354135 -0.52992260
Pt 8.31557575 4.80099990 20.00000000 4 0.26745248 0.15026639 -0.46366435
Pt 1.38592929 0.80016665 22.26321306 3 -0.33770394 -0.26172283 0.28692761
Pt 4.15778787 0.80016665 22.26321306 3 -0.02599625 -0.40194958 0.34261608
Pt 6.92964646 0.80016665 22.26321306 3 0.34787211 -0.48480666 0.29408494
Pt 2.77185858 3.20066660 22.26321306 3 -0.29877108 0.11824526 0.20432311
Pt 5.54371716 3.20066660 22.26321306 3 0.01767213 -0.01681863 0.25907558
Pt 8.31557575 3.20066660 22.26321306 3 0.40290275 -0.13287738 0.21112162
Pt 4.15778787 5.60116655 22.26321306 3 -0.25182065 0.48609477 0.05739736
Pt 6.92964646 5.60116655 22.26321306 3 0.04976374 0.37698632 0.10724723
Pt 9.70150504 5.60116655 22.26321306 3 0.42119065 0.25606629 0.06632011
Pt -0.00000000 1.60033330 24.52642611 2 -0.43969834 -0.19299160 -0.23514140
Pt 2.77185858 1.60033330 24.52642611 2 -0.05670149 -0.30899432 -0.27734968
Pt 5.54371716 1.60033330 24.52642611 2 0.24180123 -0.42035222 -0.22310247
Pt 1.38592929 4.00083325 24.52642611 2 -0.39924374 0.14384851 -0.17450312
Pt 4.15778787 4.00083325 24.52642611 2 -0.01883423 0.03337597 -0.21485627
Pt 6.92964646 4.00083325 24.52642611 2 0.29727450 -0.10432465 -0.16424841
Pt 2.77185858 6.40133319 24.52642611 2 -0.29030764 0.46974319 -0.16933322
Pt 5.54371716 6.40133319 24.52642611 2 0.07420910 0.38521412 -0.20935112
Pt 8.31557575 6.40133319 24.52642611 2 0.38585782 0.24765463 -0.15989113
Pt 0.00000000 0.00000000 26.78963917 1 -0.30347216 -0.14542174 0.34515211
Pt 2.77185858 0.00000000 26.78963917 1 -0.07683324 -0.23726018 0.40833247
Pt 5.54371716 0.00000000 26.78963917 1 0.18301935 -0.33111823 0.23034151
Pt 1.38592929 2.40049995 26.78963917 1 -0.22559471 0.06862979 0.54017049
Pt 4.15778787 2.40049995 26.78963917 1 0.00387780 -0.02380829 0.61236179
Pt 6.92964646 2.40049995 26.78963917 1 0.27744064 -0.13308601 0.42513713
Pt 2.77185858 4.80099990 26.78963917 1 -0.17265809 0.30443525 0.48137310
Pt 5.54371716 4.80099990 26.78963917 1 0.04166083 0.22511321 0.55236053
Pt 8.31557575 4.80099990 26.78963917 1 0.31119263 0.11680170 0.36831415
Ah you're comparing to experiments, that's always one layer more complex.
For starters, the adsorption energy is a global minima problem so the way you have it setup here is only evaluating 1 configuration which may not actually correspond to the global minima. You can try AdsorbML - https://github.com/facebookresearch/fairchem/blob/main/src/fairchem/applications/AdsorbML/tutorials/adsorbml_walkthrough.ipynb. This pipeline evaluates the adsorption energy by enumerating many configurations to find the minima. We generally do N=100 placements.
@brunosamp4 The experimental energy you mention is a small negative number. My guess is you mean the adsorption energy of C2H6 relative to gas phase C2H6 (and the small negative number means there's a weak attraction).
The -4 eV you're seeing is very large, and much more negative than the experimental one. @mshuaibii is 100% correct that you should find the best adsorption site on the surface, but it can only make the number lower by finding a better site, which will make your problem worse.
I think the issue you have is that you need to consider the gas phase reference that Muhammed mentioned, which predicts the energy relative to a linear combination of CO, H2O, and H2, not C2H6(g). So, you need to subtract the reaction energy of 2CO + 5H2 → C2H6 + 2H2O, which happens to be -3.59 eV or so based on experimental thermochemistry). That gets you quite close to the experimental value!
Please follow this tutorial to make sure you understand the correction and can arrive at the -3.59 eV yourself (and verify that I didn't make any mistakes!) https://fair-chem.github.io/catalysts/examples_tutorials/OCP-introduction.html
@zulissimeta Oh i see, i didn't realize that we should use the reference reaction here as well. Thank you very much! Now i'm getting good results.