Importance of lattice vectors for arbitrary molecule inputs

[davidcjuergens]

Hello,

How important are the lattice vectors in the .extxyz file inputs to make a dataset for SevenNet? If I don't want the lattice parameters to have an effect, can I simply set them to the identity matrix? Thanks

David

[YutackPark]

Set atoms.pbc = False. You can define structure that doesn't have lattice vectors or pbc. Remove DFT stress value (if exist). SevenNet will do the right thing.

…

----------------------------------------------------------------- *Yutack Park* Ph.D. Candidate Materials Data and Informatics Lab. Department of Materials Science and Engineering Seoul National University *E-mail:* ***@***.*** ***@***.***> *Homepage:* http://mtcg.snu.ac.kr ----------------------------------------------------------------- 2025년 3월 7일 (금) 오전 5:20, David Juergens ***@***.***>님이 작성:

Hello, How important are the lattice vectors in the .extxyz file inputs to make a dataset for SevenNet? If I don't want the lattice parameters to have an effect, can I simply set them to the identity matrix? Thanks David — Reply to this email directly, view it on GitHub <#178>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2RQY26AQ5HHIVTN5J7W5ZD2TCUYZAVCNFSM6AAAAABYPTDSTGVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZYGA2DQMZSGY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[davidcjuergens]

Where in the code can I set this parameter? Thanks

David

[davidcjuergens]

Update: My current solution is to adjust the _read_ase_readable method in SevenNetGraphDataset as follows:

@staticmethod
    def _read_ase_readable(
        filename: str,
        cutoff: float,
        num_cores: int = 1,
        tag: str = '',
        transfer_info: bool = True,
        allow_unlabeled: bool = False,
        **ase_kwargs,
    ) -> list[AtomGraphData]:
        
        pbc = ase_kwargs.pop('pbc', False) # ase_reader cannot take pbc kwarg, so pop
        atoms_list = dataload.ase_reader(filename, **ase_kwargs)
        for atoms in atoms_list:
            atoms.pbc = pbc # set pbc attr for atoms here

I am happy to submit a pull request if you think this would be a nice addition

[YutackPark]

I'm glad to hear you have solved an issue by yourself and I do want to patch the source. But I can't fully understand the problem you originally had. What I mentioned first was about preprocessing the .extxyz file before it passed to SevenNet. But I think your solution is overriding pbc attribute of atoms object inside the SevenNet.

Could you tell me why this was needed? If you want to use molecule data that doesn't have stress label, SevenNet will automatically ignore them so you don't have to bothered with PBC or lattice vectors.

[davidcjuergens]

Thanks for following up. My case is the following:

I just want to train SevenNet on the QM9 dataset (the individual molecules, without symmetry), so I made a .extxyz file from Qm9. I then fed this .extxyz file to SevenNetGraphDataset init function. i.e.,

xyz_files = ["/path/to/my/qm9/qm9.extxyz"]

dataset = SevenNetGraphDataset(
        cutoff=5.0,
        files=xyz_files,
        allow_unlabeled=True,
        process_num_cores=1
    )

In my .extxyz file, I set the lattice vectors to the identity matrix in an attempt to have a null effect from them. I noticed though, that after dataset construction, a single Qm9 example had ~2e+4 edges (just looking at dataset[0]). So clearly something was wrong and the molecule's atoms were presumably being duplicated in some kind of crystalline symmetry.

I then saw your advice to set atoms.pbc to false, and didn't know where else to set it False besides within the _read_ase_readable method as above. With the altered method, I then change my initialization of SevenNetGraphDataset to

dataset = SevenNetGraphDataset(
        cutoff=5.0,
        files=xyz_files,
        allow_unlabeled=True,
        process_num_cores=1,
        pbc=False,
    )

The graph # of edges now looks more reasonable (about 20 edges in dataset[0]).

Does this help explain?
Best,
David