|
Dear @gabor1, I have two different datasets (dataset 1 and dataset 2) that include the same metal (M) in different electronic states, which we can call M₁ and M₂. I was wondering if it is possible to develop a single potential from a combined data pool that includes both datasets by defining custom atoms M₁ and M₂. In addition, how can I define the reference-energy entry for both atoms? I would also like to perform fine-tuning. Is this possible? Thanks. Paula Abufager |
Replies: 12 comments 39 replies
|
If you always want to keep M1 and M2 separate, you will never want to change one oxidation state into the other, you can pretend that M1 and M2 are different atoms (just use a fake atom type). Alternatively, you can just call them both the same atom, and let MACE ingest all the data - it will need to infer the electronic structure from the local geometry. for many cases that works well! |
|
One way you can do that cleanly also is using the new embedding functionality of mace. Just create a categorical embedding with 0 being M1 and 1 being M1. You add to your input the following Then add oxydations_states to the atoms.arrays. You will need to modify the AtomicData though to correctly extract it. See the test here: mace/tests/test_embedding_train.py Line 13 in 0139da1 AtomicData here: https://github.com/ACEsuit/mace/blob/0139da1b864b29054f07db6627887a847a42050e/mace/data/atomic_data.py Plus if you |
|
Is there a more complete example for arbitrary per-atomic features? Is modifying |
|
bring the discussion from #1218 here @bernstei @rreocreux |
|
It seems like the embedding functionality is the best, and it's already almost there, just needs a modification in the data extractor ? |
|
yeah someone needs to take the pieces that exist and just put them together into an example that works. Several people could do it (@RokasEl @hatemhelal etc.) I'll ask around who has the capacity. |
|
I've started working on a general implementation. The bits that read the xyz and store it in |
|
Dear @gabor1, I would like to know if there is any chance of getting an example on how to deal with embedded atoms. Thank you very much in advance. Paula |
|
@bernstei is putting together a PR and an example, see below. |
|
@ilyes319 can you confirm the format of |
|
Looks from the code like it's a dict of dicts, so I'll update the docs. |
|
It is the same format as here https://github.com/ACEsuit/mace-foundations/blob/main/mace_omol/mace-omol.sh |
|
So, it turns out that doing this well for arbitrary features is actually quite tricky, basically because there are several different places where you need to know what atoms.info/atoms.arrays entries you need in the Actually, this is already broken, I think. You can see that I'm trying to figure out what can be done, but there are so many places that at least a piece of the |
|
The code changes are remarkably compact so far. |
|
The continuous per graph works? That should work because of thats elect temperature. |
|
Yes, it does. Just not per-atom. Could be that I'm packing the information into the |
|
Yeah - both the categorical and continuous are being packed into |
|
Looks like |
|
Please see #1222. Seems to work in my initial test https://github.com/ACEsuit/mace/blob/arb_per_atom_descriptors/tests/test_embedding_train.py#L213 |
|
In principle you should be able to, using the test I added as the framework for an example. I was hoping to have some feedback from @ilyes319 before recommending that. |
|
Hey Paula, yes please do give it a try, update to the latest dev branch. |
|
I am not sure how to include the embedding_spec block in the input file for training. I am not using config.yaml. Please find below my input file (after running, it does not recognize the embedding part). In the trainG1.xyz and testG1.xyz files, I have included total_spin = X (with X being the corresponding value) on the line where REF_energy is defined for each configuration. Thank you very much. Best. Paula. mace_run_train |
|
my recollection is that it will attempt to interpret the string argument as a python (source file) dict, using something like the python |
|
you have an example here: https://github.com/ACEsuit/mace-foundations/blob/main/mace_omol/mace-omol.sh |
|
Dear @ilyes319 In total_spin_key (embedding part), I replaced spin (total_spin) with total_spin (spin), but it did not work. Any help would be welcome. Thanks. Paula |
|
I'm confused - are you trying to have a new embedding that is per-atom, or per config? Because this whole discussion is about per-atom, but your arguments above seem to be setting up a per-config (graph) rather than per-atom embedding. |
|
Give that the only place that |
Since the spin is localized on the single metal atom present in each configuration, I have set the spin to be per-configuration (graph) as in the example https://github.com/ACEsuit/mace/blob/arb_per_atom_descriptors/tests/test_embedding_train.py#L213 . In any case, I can also define the spin per atom — no problem. I can try this approach as well, although I'm not entirely sure about the format of the input file. |
|
I see. That sounds like it should work. To do per-atom, you need to set the key of the top level dict passed as |
|
Obviously (I hope), you'll also need to add a column for each atom containing the |
|
Dear @bernstei , I would like to know if there is any known incompatibility between PyTorch 2.4.0 and the development version of MACE. Thank you very much, Best. Paula 2025-11-10 13:52:04.787 INFO: Training complete |
|
Loading pytorch-gpu/py3/2.8.0
|
|
As far as I can tell, this happens because there are some parameters of type EmptyParam in the model and that causes some issues when doing a deep copy of it. I have no idea if that's expected or not. |
|
So just a bit of context, Paula and I are trying to train a model with custom atoms on Jean-Zay (hence Rémi's comment). Even though the model is learning, the final model cannot be written. Have you experienced that @bernstei @ilyes319 ? I know some of you are also using Jean-Zay. |
|
What version of pytorch are you using? |
|
PyTorch 2.8.0, but I also tried PyTorch 2.4.0 (same problem) |
If you always want to keep M1 and M2 separate, you will never want to change one oxidation state into the other, you can pretend that M1 and M2 are different atoms (just use a fake atom type).
Alternatively, you can just call them both the same atom, and let MACE ingest all the data - it will need to infer the electronic structure from the local geometry. for many cases that works well!