René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"[Bug]: TorchSim backend returns NaN on some structures","articleBody":"### Contact Details\n\ngael.huynh@uclouvain.be\n\n### Bug Description\n\nSome structures seems to return NaN when using the TorchSim backend while using `Potential.predict_properties` works perfectly fine. Note that this error only occurs for some structures (see example below where only 1 structure presents this issue).\n\n### MatterSim Version\n\n1.2.5\n\n### Python Version\n\n3.12.12\n\n### Reproduction Steps\n\n1. Install dependencies `pip install mattersim==1.2.5`\n2. Run the code below\n3. The second structure should return NaN values when using the TorchSim backend\n\n```py\nimport numpy as np\nimport requests\nimport torch\nfrom mattersim.forcefield.potential import Potential\nfrom pymatgen.core import Structure\n\n\ndef fetch_alexandria_entry(id_string):\n    url = f\"https://alexandria.icams.rub.de/pbe/v1/structures/{id_string}\"\n\n    response = requests.get(url)\n    response.raise_for_status()\n    data = response.json()[\"data\"]\n    attrs = data[\"attributes\"]\n\n    return Structure(\n        lattice=attrs[\"lattice_vectors\"],\n        species=attrs[\"species_at_sites\"],\n        coords=attrs[\"cartesian_site_positions\"],\n        coords_are_cartesian=True,\n    )\n\n\ndef direct_inference(model, structures):\n    from mattersim.datasets.utils.build import build_dataloader\n    from pymatgen.io.ase import AseAtomsAdaptor\n\n    adapter = AseAtomsAdaptor()\n    atoms = [adapter.get_atoms(struct) for struct in structures]\n\n    dataloader = build_dataloader(atoms, only_inference=True)\n    predictions = model.predict_properties(\n        dataloader,\n        include_forces=True,\n        include_stresses=True,\n    )\n\n    return predictions\n\n\ndef torchsim_inference(model, structures):\n    import torch_sim as ts\n    from mattersim.torchsim import get_torchsim_wrapper\n\n    device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n    wrapper = get_torchsim_wrapper(model, device=device)\n\n    state = ts.static(structures, model=wrapper)\n    return state\n\n\ndef main():\n    alexandria_ids = [\"agm003221802\", \"agm006228157\"]\n    structures = [fetch_alexandria_entry(id) for id in alexandria_ids]\n\n    model = Potential.from_checkpoint(\n        load_path=\"mattersim-v1.0.0-1M\", load_training_state=False\n    )\n\n    results = torchsim_inference(model, structures)\n    for i in range(len(structures)):\n        for key, prop in results[i].items():\n            if torch.isnan(prop).any():\n                print(f\"NaN value found in {key} for structure {i} (TorchSim)\")\n\n    results = direct_inference(model, structures)\n    for i in range(len(structures)):\n        for key, prop in zip([\"energy\", \"forces\", \"stress\"], results[i]):\n            if np.isnan(prop).any():\n                print(f\"NaN value found in {key} for structure {i} (ASE)\")\n\n\nif __name__ == \"__main__\":\n    main()\n```\n\n### Expected Behavior\n\nI would expect the TorchSim backend to return the same values as `Potential.predict_properties` and to not return NaN values on valid structures.\n\n### Actual Behavior\n\nSome structures will return NaN properties when using TorchSim.\n\n### Error Logs\n\n```python\n\n```\n\n### Code of Conduct\n\n- [x] I agree to follow the project's Code of Conduct","author":{"url":"https://github.com/akwarii","@type":"Person","name":"akwarii"},"datePublished":"2026-06-09T09:47:36.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":0},"url":"https://github.com/162/mattersim/issues/162"}