Predictive GFM 2025

[allaffa]

Main changes:

1. Consistency check across all large scale datasets for the format of the torch.geometric.data objects

              data_object = Data(
                   dataset_name="dataset_name",
                   natoms=natoms,
                   pos=pos,
                   cell=None,  # even if not needed, cell needs to be defined because ADIOS requires consistency across datasets
                   pbc=None,  # even if not needed, pbc needs to be defined because ADIOS requires consistency across datasets
                   edge_shifts=None,  # even if not needed, edge_shift needs to be defined because ADIOS requires consistency across datasets
                   atomic_numbers=atomic_numbers,  # Reshaping atomic_numbers to Nx1 tensor
                   chemical_composition=chemical_composition,
                   smiles_string=smiles_string,
                   x=x,
                   energy=energy,
                   energy_per_atom=energy_per_atom,
                   force=forces,
               )

2. apply graphgps_transform to compute structural and positional Laplacian encodings
3. Allow for parsed input argument choice to set compute_grad_energy. Default values is False
4. default value of energy_per_atom is set to False, because we do not need to normalize for machine learning force fields
5. Added chemical composition as one-dimensional vector with 118 entries. each entry counts the number of atoms for that chemical species in the atomistic structures

[allaffa]

@RylieWeaver @ArCho48 @zachfox
I added smiles_string as an attribute to each Data object. This attribute is set to None for inorganic compounds, for which the SMILES representation does not make sense, and also for those organic molecules for which xyz2mol struggles reconstructing the nature of the chemical bonds between atoms. This would require running quantum mechanical calculations, which obviously is insane to think about in this context.

[allaffa]

@pzhanggit
This PR corresponds to the branch where we will try to perform our imbalanced, multi-source work.
Please take a look at the structure of the Data objects.

[allaffa]

@zachfox
Whenever we will move ahead with the conditional DM, please take a look at the Data structures of this PR

[pzhanggit]

@pzhanggit This PR corresponds to the branch where we will try to perform our imbalanced, multi-source work. Please take a look at the structure of the Data objects.

Thank you, Max @allaffa . Introducing a dataset_name ID looks good for our multi-source work. Let me know when you complete the datasets generation, and I'll start the multi-source model training.

About the changes in hydragnn/utils/descriptors_and_embeddings/smiles_utils.py, I suggest we move them to a standalone file.

[pzhanggit]

See my replies to Max's comment

[allaffa]

@pzhanggit This PR corresponds to the branch where we will try to perform our imbalanced, multi-source work. Please take a look at the structure of the Data objects.

Thank you, Max @allaffa . Introducing a dataset_name ID looks good for our multi-source work. Let me know when you complete the datasets generation, and I'll start the multi-source model training.

About the changes in hydragnn/utils/descriptors_and_embeddings/smiles_utils.py, I suggest we move them to a standalone file.

@pzhanggit
Thanks, I moved the functionalities in a separate xyz2mol.py script