EricBoi/mlip-visualization-models

Models for MLIP-Visualization

Converted weights for MLIP-Visualization, which runs machine-learning interatomic potentials in the browser and draws every forward and backward pass. The app downloads these files directly; they are plain safetensors plus a JSON file of hyperparameters per model.

file	model	source	licence
pet-mad-xs	PET-MAD XS	lab-cosmo/upet, converted with convert_pet.py	BSD-3-Clause
mace-mp-0b3-medium	MACE-MP-0b3 medium	mace-foundations/mace-mp-0, converted with convert_mace.py	MIT
mace-mp-0b2-small	MACE-MP-0b2 small	mace-foundations/mace-mp-0, converted with convert_mace.py	MIT
lorem-demo	LOREM demo	metatrain experimental LOREM, random initialization from convert_lorem.py	MIT
ani-2x	ANI-2x	TorchANI models.ANI2x(), converted with convert_ani.py	MIT
physnet-acetone	PhysNet · acetone MP2	mmml physnetjax, trained with train_physnet_demo.py	MIT
pet-mols-s-v1.0	PET-MOLS S v1.0	lab-cosmo/upet, converted with convert_pet.py	BSD-3-Clause

index.json lists what the app offers.

Credits

• PET-MAD / PET-MOLS: lab-cosmo/upet (BSD-3-Clause), converted with pet-kokkos's convert_pet.py. Cite the PET and PET-MAD papers when you use them.
• MACE-MP-0: mace-foundations/mace-mp-0 (MIT), Batatia et al., A foundation model for atomistic materials chemistry (2023), converted with scripts/convert_mace.py.
• LOREM demo: a random initialization of metatrain's experimental LOREM (Bigi et al., arXiv:2507.19382), exported with scripts/convert_lorem.py. Not a trained potential.
• ANI-2x: TorchANI (MIT), Devereux et al., J. Chem. Theory Comput. 16, 4192 (2020), converted with scripts/convert_ani.py.
• PhysNet (acetone): a small invariant PhysNet (mmml physnetjax, MIT, max_degree = 0) trained for this app on mmml's acetone-dimer MP2 example data, with scripts/train_physnet_demo.py.