Orb: A Fast, Scalable Neural Network Potential

TL;DR

Orb presents a scalable, universal interatomic potential framework that leverages diffusion pretraining and a non-equivariant Graph Network Simulator to achieve state-of-the-art accuracy and substantial speedups on the Matbench Discovery benchmark. It demonstrates robust generalization across molecules and crystalline materials, including MD stability in MD17 and MOF-5, and reasonable adsorption energetics in Mg-MOF-74 when combined with dispersion corrections. The work argues that relaxing strict symmetry constraints, coupled with diffusion-based pretraining and torque-removal adjustments, can rival conventional conservative, equivariant models while enabling large-scale, high-fidelity simulations. By releasing open-source code and providing extensive benchmarks, Orb aims to accelerate materials discovery and scalable atomistic modeling in both research and industry contexts.

Abstract

We introduce Orb, a family of universal interatomic potentials for atomistic modelling of materials. Orb models are 3-6 times faster than existing universal potentials, stable under simulation for a range of out of distribution materials and, upon release, represented a 31% reduction in error over other methods on the Matbench Discovery benchmark. We explore several aspects of foundation model development for materials, with a focus on diffusion pretraining. We evaluate Orb as a model for geometry optimization, Monte Carlo and molecular dynamics simulations.

Figures (6)

• Figure 1: top: Element distribution of the full diffusion pretraining dataset for Orb. bottom: Distribution of the dataset with respect to system size. Due to the efficiency of Orb's architecture, we can train on systems with up to 5000 atoms, two orders of magnitude larger than MatterGen zeni2024mattergengenerativemodelinorganic, a similar generative model of materials.
• Figure 2: Effect of pretraining and model size on Alexandria and MPtraj Energy MAE, Forces MAE and Forces within a threshold of 0.03. Diffusion pretraining helps universally with improvements between 17% and 70%, even on Alexandria, which is an order of magnitude larger MPtraj.
• Figure 3: left: Model forward pass speed (excluding featurization) compared to MACE on a single NVIDIA A100 GPU. At large system sizes, Orb is between 3 to 6 times faster than MACE. right: End to end model inference speed for a 100 atom system on a single NVIDIA A100 when implemented as a Calculator object in the Atomic Simulation Environment Python library. The D3 dispersion correction adds a substantial cost which is amortized by Orb models, as the corrections are incorporated into training datasets. All measurements reported as the median of 50 runs.
• Figure 4: (a) Distribution of dihedral angles across the MD trajectory for MOF-5, shown at six time intervals (0–500, 500–1000, 1000–1500, 1500–2000, 2000–2500, and 2500–3000 ps) corresponding to increasing temperatures from 300 K to 800 K; (b) top: snapshots of MOF-5 at 2,500 ps (stable framework) and 3,500 ps (decomposed structure), middle: root mean squared deviation (RMSD) from the initial structure, remaining constant until around 3,000 ps, after which significant structural degradation occurs, bottom: simulation temperature over time, showing incremental heating; (c) schematic of the four carbon atoms ($C1$–$C4$) in the MOF-5 and linker structure used to calculate the dihedral angle.
• Figure 5: (a) Free energy surface of MACE + D3 (left) and Orb-D3 (right) obtained from Widom insertion in Mg-MOF-74. Lowest free energies are represented by the blue region near the open-metal centers, indicating favorable adsorption sites for $\text{CO}_2$. (b) $\text{CO}_2$ adsorption locations in Mg-MOF-74 showing the two sites with the most favorable adsorption energies, obtained with Widom insertion, with values of -54.5 kJ/mol and -54.4 kJ/mol for the respective sites. While both Orb and MACE predict similar locations of energy minima, the free energy minima of ORB are closer in magnitude to the experimental heat of adsorption (-44 kJ/mol).