Today, we release a new suite of Orb foundation models for simulation, designed to reach unprecedented scale and accuracy. In this third release, we introduce Orb in a number of flavors – each of which strikes a different balance between accuracy and computational efficiency. While our most accurate model achieves state-of-the-art performance on many physical property prediction tasks, our fastest model is an order of magnitude more computationally efficient but remarkably retains comparable accuracy. Our fastest models make simulating quantum properties at the mesoscale a possibility for the first time.
Orb-v3 comprises a diverse set of scalable, universal all-atom foundation models strategically positioned along the performance-speed-memory Pareto frontier. Access to models across this frontier is important as it allows users to choose the right trade off for their use case.
This dramatically improves its generalization performance across the periodic table.
A simple and differentiable regularization strategy which strengthens the symmetry-induced inductive biases without sacrificing performance.
These improve precision towards fine-grained computational workflows; notable examples include optional conservatism (i.e. computation of the interatomic forces via gradients of the energy) and smoother radial embeddings
Our sparse direct model, orb-v3-direct-20, is 10x faster and 5x more memory efficient than MACE-MPA-0, and 40x faster and 40x more memory efficient than Sevennet. We believe this scalability is essential to perform accurate atomic simulations at scale.
Orb-v3 breaks through a long-standing barrier in simulation: it brings quantum-level accuracy to the mesoscale, where the collective quantum behavior of tens or even hundreds of thousands of atoms gives rise to entirely new phenomena. This makes it possible — for the first time — to simulate complex processes like crystal nucleation and growth, the self-assembly of nanostructures such as metal-organic frameworks, and the functional dynamics of biomolecules like enzymes. These are the scales where many real-world breakthroughs in advanced materials and biochemistry will happen — and Orb-v3 makes them accessible to computer design.
Model weights and code are available under an Apache 2.0 License on Github at https://github.com/orbital-materials/orb-models.
Models are also available to use interactively via our launch partner, Rowan.
