0.5.5

What's Changed

• Update NaCl examples to use orb-v3 and streamline dependencies by @timduignan in #106
• Fix conservative model loading for finetuning by @vsimkus in #112
• correct_read_me_docs: docs: Update README.md by @evansdoe in #100
• Add r_max as argument with default 6A. by @ameya98 in #115
• Fix aggregate nodes when last batch in a graph has no edges/nodes by @vsimkus in #119
• Omol models by @benrhodes26 in #120

New Contributors

• @evansdoe made their first contribution in #100
• @ameya98 made their first contribution in #115

Full Changelog: 0.5.4...v0.5.5

0.5.4

What's Changed

• Fix cuML OOM issue by @vsimkus in #93
• update publishing mechanism to not use poetry by @DeNeutoy in #94

Full Changelog: 0.5.3...0.5.4

0.5.3

What's Changed

• don't squash singleton dims in calc by @DeNeutoy in #89
• force fp64 in brute force edge construction by @DeNeutoy in #91

Full Changelog: v0.5.2...0.5.3

0.5.2

What's Changed

• add readme example for confidences by @DeNeutoy in #84
• fix missing model url by @DeNeutoy in #85
• Docker image by @benrhodes26 in #86

Full Changelog: v0.5.1...v0.5.2

0.5.1

What's Changed

• Update readme with paper links by @benrhodes26 in #77
• Pin dm-tree==0.1.8 due to MacOS compilation issues by @vsimkus in #78
• Featurization fixes and a speed benchmarking script by @benrhodes26 in #81

Full Changelog: v0.5.0...v0.5.1

0.5.0

April 2025: We have released the Orb-v3 set of potentials. These models improve substantially over Orb-v2, in particular:

Model compilation using PyTorch 2.6.0+, enabling faster inference while maintaining support for dynamic graph sizes
Wider architecture (1024 vs 512) with fewer layers (5 vs 15) compared to v2, resulting in 2-3x faster performance with similar parameter count
Two variants available: direct models and conservative models (forces/stress computed via backpropagation)
Trained on the larger, more diverse OMat24 dataset
Improved edge embeddings using Bessel-Spherical Harmonic outer products (8 Bessel bases, Lmax=3)
Enhanced stability through Huber loss and a ZBL pair repulsion term added to forces
Models available with both unlimited neighbors and 20-neighbor maximum configurations
New confidence head providing intrinsic uncertainty estimates for predictions

v0.4.2

What's Changed

• added jupyter notebook and examples for sagemaker by @Arthurhussey in #45
• update notebook by @Arthurhussey in #46
• Replace torch.cuda.amp.autocast -> torch.autocast by @BenedictIrwin in #43
• added eol announcement by @Arthurhussey in #47
• fixed marketplace product id by @Arthurhussey in #48
• fix: mark the return type as GraphRegressor instead of torch.nn.Module by @caic99 in #50
• Feature/md tutorial by @timduignan in #49
• Colab for MD by @timduignan in #51
• correct ordering of properties by @DeNeutoy in #55
• added azure model card and examples by @Arthurhussey in #56

New Contributors

• @BenedictIrwin made their first contribution in #43
• @caic99 made their first contribution in #50
• @timduignan made their first contribution in #49

Full Changelog: v0.4.1...v0.4.2

v0.4.1

What's Changed

• apache-2 license by @DeNeutoy in #31
• moved weights to s3 by @Arthurhussey in #38
• Fix ase_atoms_to_atom_graphs not respecting given flag for graph computation and add device parameter to graph computation functions. by @nimashoghi in #35
• Wrap by default by @benrhodes26 in #41
• Pass device in calculator by @benrhodes26 in #42
• Relax numpy and torch constraints by @benrhodes26 in #40

New Contributors

• @nimashoghi made their first contribution in #35

Full Changelog: v0.4.0...v0.4.1

v0.4.0 - Orb V2

What's Changed

Oct 2024: We have released a new version of the models, orb-v2. This version has 2 major changes:

• v2 models use a smoothed cosine distance cutoff for the attention mechanism. This is a more physically motivated cutoff that is better suited for MPNNs.
• The force predictions now have net zero forces, meaning they are much more stable for MD simulations.
• The models are generally more accurate (Increase in 2-3% on the matbench discovery dataset).

Relevant PRs

• Orb v2 models by @DeNeutoy in #27
• Finetuning improvements by @benrhodes26 in #29
• Add net torque removal by @DeNeutoy in #30

Full Changelog: v0.3.2...v0.4.0

V0.3.2

What's Changed

• Fix device issue by @benrhodes26 in #17
• Finetune script by @zhiyil1230 in #13

Full Changelog: v0.3.1...v0.3.2