Papers
Topics
Authors
Recent
Search
2000 character limit reached

ORB-v3: Scalable Atomic Simulation

Updated 4 July 2026
  • The paper introduces ORB-v3, a universal ML interatomic potential that achieves near–state-of-the-art accuracy while reducing latency and memory use by over 10× and 8×, respectively.
  • ORB-v3 employs a graph-based architecture with enhanced radial and angular embeddings, using widened MLPs in a GNS-style message-passing framework to optimize speed and efficiency.
  • Benchmarking on high-temperature MOFs and large-scale simulations shows ORB-v3’s effective trade-off between accuracy and scalability, despite challenges in bond-breaking regimes.

ORB-v3 is a family of universal machine-learning interatomic potentials designed to combine near–state-of-the-art accuracy with low latency, low memory usage, and scalability to very large atomistic systems (Rhodes et al., 8 Apr 2025). It extends the Orb-v2 line while systematically varying conservatism, graph sparsity, and training data regime, and it explicitly argues that non-equivariant, non-conservative architectures can still model energies, forces, stresses, and even higher-derivative-dependent properties with strong accuracy (Rhodes et al., 8 Apr 2025). Independent benchmarking on high-temperature metal-organic frameworks places ORB-v3 among the strongest off-the-shelf universal MLIPs near equilibrium, while also showing substantial degradation in bond-breaking and decomposition regimes, especially when the model is used generatively in long molecular-dynamics trajectories (Edwards et al., 28 Apr 2026).

1. Conceptual scope and scientific role

ORB-v3 is framed as a universal interatomic potential, meaning a single pre-trained model family intended to operate across broad chemistry and structure space rather than being fit to one material or one reaction class (Rhodes et al., 8 Apr 2025). In this setting, the model takes atomic positions and species as input and predicts a scalar potential energy,

E=E({Ri,Zi}),E = E(\{\mathbf{R}_i, Z_i\}),

together with derivatives such as forces,

Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,

for conservative variants, and also stresses or virials (Rhodes et al., 8 Apr 2025).

The stated motivation is the gap between the cost of ab initio simulation and the scale of problems of interest. Density-functional workflows are described as too expensive for nanosecond trajectories and for systems in the 10410^4–10510^5 atom regime, whereas many materials and chemistry workflows require exactly those scales together with reliable derivative-based observables such as phonons, elastic constants, and thermal conductivities (Rhodes et al., 8 Apr 2025). ORB-v3 therefore targets a combined objective: accuracy, latency, and system-size scalability, rather than optimizing only one axis.

Within that agenda, ORB-v3 is also a direct intervention in an architectural debate. The model family is presented as evidence against the claim that strict roto-equivariance and strict conservatism are always prerequisites for accurate universal MLIPs. The paper reports that non-equivariant, non-conservative variants can remain accurate even on tasks that depend on higher derivatives of the potential-energy surface, provided that the architecture, data, and regularization are chosen appropriately (Rhodes et al., 8 Apr 2025).

2. Architecture and model family

ORB-v3 retains the basic architecture and diffusion pretraining scheme of Orb-v2, but revises the representation and systematically enumerates model variants (Rhodes et al., 8 Apr 2025). The core representation is graph-based: atoms are nodes, edges connect neighboring atoms within a cutoff of about $10$ Å, and periodic systems are handled through dynamic supercell tiling so that neighbor lists remain correct under periodic boundary conditions (Rhodes et al., 8 Apr 2025).

The edge representation changes materially relative to Orb-v2. Orb-v2 used unit vectors and $20$ Gaussian radial basis functions, whereas ORB-v3 uses the outer product of radial Bessel basis functions and spherical-harmonic angular embeddings, specifically $8$ Bessel bases and spherical harmonics with Lmax⁡=3L_{\max}=3, together with a smooth envelope cutoff (Rhodes et al., 8 Apr 2025). The message-passing backbone follows a GNS-style architecture; its multilayer perceptrons are widened from $512$ to $1024$ while the depth is reduced to Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,0 layers, a design choice explicitly described as favoring width over depth for speed (Rhodes et al., 8 Apr 2025).

The family is organized by a naming convention,

Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,1

with Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,2, Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,3, and Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,4 (Rhodes et al., 8 Apr 2025). The three axes correspond to distinct modeling choices.

“Conservative” variants compute forces and virial stresses by analytic differentiation of the energy with respect to positions and a symmetric cell-strain tensor, so energy–force consistency is exact by construction (Rhodes et al., 8 Apr 2025). “Direct” variants instead use explicit force and stress heads, so the force field is not required to be the gradient of a scalar energy (Rhodes et al., 8 Apr 2025). The sparsity parameter distinguishes a hard cap of Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,5 neighbors per atom from using all neighbors within the cutoff. The Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,6-neighbor setting gives major computational savings, but the hard maximum-neighbor rule can introduce discontinuities when neighbors enter or leave the retained set (Rhodes et al., 8 Apr 2025).

These variants imply four principal operating regimes. Conservative-Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,7 is the most physically constrained and is the strongest basis for strict energy conservation and smooth higher derivatives. Conservative-Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,8 preserves force–energy consistency but inherits neighbor-set discontinuities. Direct-Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,9 is continuous but non-conservative. Direct-10410^40 is the most throughput-oriented configuration and the one most explicitly optimized for very large systems (Rhodes et al., 8 Apr 2025).

3. Training data, objectives, and regularization

ORB-v3 is trained in two major data regimes: OMat24 and MPA (Rhodes et al., 8 Apr 2025). For OMat24, the model uses only the approximately 10410^41 million AIMD-sampled structures and excludes the roughly 10410^42 of the dataset labeled as “rattled” structures (Rhodes et al., 8 Apr 2025). The exclusion is motivated by pathological behavior on heteronuclear diatomics, where rattled configurations induced large kinks in the potential-energy surface; the final filtering choice is reported to yield smooth, physically reasonable diatomic curves (Rhodes et al., 8 Apr 2025). The MPA regime corresponds to MPTraj plus Alexandria and is retained partly for compatibility with legacy evaluation pipelines such as Matbench-Discovery (Rhodes et al., 8 Apr 2025).

Supervised targets include energy, forces, virial stress for conservative models, and a confidence signal (Rhodes et al., 8 Apr 2025). The paper states that Huber loss with 10410^43 is used for energy, force, and stress targets, rather than pure MAE, and that a non-learnable Zeigler–Biersack–Littmark pair-repulsion term is added to enforce physically correct short-range repulsion (Rhodes et al., 8 Apr 2025).

A specific regularizer, “equigrad,” is introduced for conservative variants. It penalizes the derivative of the energy with respect to an infinitesimal rotation parameterization, so that exact rotational invariance would correspond to a zero penalty term (Rhodes et al., 8 Apr 2025). Because conservative models already backpropagate through the energy to obtain forces and stresses, this regularization is described as incurring no extra backward cost (Rhodes et al., 8 Apr 2025). The regularizer complements rotational data augmentation rather than replacing it.

The paper also reports a noteworthy training pathology and its remedy. Direct ORB-v3 models trained on MPA, even when initialized from OMat24 pretraining, tended to overfit forces and to perform poorly on second- and third-derivative-sensitive tasks. The stated fix is distillation from a conservative teacher, specifically orb-v3-conservative-inf-mpa, whose static predictions on the full MPA dataset are used as teacher labels for direct models (Rhodes et al., 8 Apr 2025). The authors identify the origin of this behavior as an open question and mention Hessian-based distillation as a possible future direction (Rhodes et al., 8 Apr 2025).

ORB-v3 also includes a confidence head trained to predict a binned intrinsic force error per atom, using predefined bins 10410^44 and a cross-entropy objective on detached node states (Rhodes et al., 8 Apr 2025). The resulting signal is described as analogous to AlphaFold’s pLDDT and is reported to correlate with per-atom force MAE even on out-of-distribution settings such as small molecules and zeolites (Rhodes et al., 8 Apr 2025).

4. Benchmarks, efficiency, and scaling behavior

The primary quantitative claim of ORB-v3 is that it expands the performance–speed–memory Pareto frontier, providing near–state-of-the-art accuracy with at least a 10410^45 reduction in latency and at least an 10410^46 reduction in memory relative to prior universal MLIPs (Rhodes et al., 8 Apr 2025). The benchmarks used to support that claim include Matbench-Discovery, an MDR phonon benchmark, elastic-modulus evaluation, and explicit large-system scaling studies (Rhodes et al., 8 Apr 2025).

On Matbench-Discovery with MPA training, orb-v3-conservative-inf-mpa achieves an 10410^47 score of 10410^48, a 10410^49 of 10510^50, and 10510^51 steps/s on a 10510^52-atom periodic system (Rhodes et al., 8 Apr 2025). Orb-v3-direct-inf-mpa records 10510^53, 10510^54, and 10510^55 steps/s, while orb-v3-direct-20-mpa records 10510^56, 10510^57, and 10510^58 steps/s (Rhodes et al., 8 Apr 2025). For comparison, the same table reports 10510^59, $10$0, and $10$1 steps/s for MACE-MPA-0; $10$2, $10$3, and $10$4 steps/s for SevenNet-MF-ompa; and $10$5 for Orb-v2 (Rhodes et al., 8 Apr 2025). These figures are used to argue that conservative and direct ORB-v3 variants occupy different favorable points on the same frontier rather than one regime strictly dominating the other.

The derivative-sensitive benchmarks are especially important because they bear on the paper’s central architectural claim. On the MDR phonon and mechanical-property evaluation using OMat24-AIMD training, orb-v3-conservative-inf-omat is reported as best across almost all property MAEs, with values of $10$6 K for $10$7, $10$8 J/mol·K for $10$9, $20$0 kJ/mol for $20$1, $20$2 J/mol·K for $20$3, and $20$4 GPa for $20$5 (Rhodes et al., 8 Apr 2025). Orb-v3-direct-inf-omat remains very close, with $20$6, $20$7, $20$8, $20$9, $8$0, and $8$1 on the same metrics (Rhodes et al., 8 Apr 2025). The paper interprets this as evidence that direct, non-conservative models can still be smooth enough for practical finite-difference phonon and elasticity workflows.

Scalability is treated as a first-class result rather than an implementation detail. ORB-v3 uses an adaptive graph-construction strategy: brute-force GPU distance computation via torch.cdist and topk for small systems, cuML nearest neighbors with algorithm="rbc" for intermediate sizes, and CPU KD-trees when GPU memory pressure dominates (Rhodes et al., 8 Apr 2025). In the explicit $8$2-atom benchmark, all baselines listed in the paper together with ORB-v3 conservative models run out of memory, whereas orb-v3-direct-20 completes in under $8$3 s on an NVIDIA H200 GPU with about $8$4 GB of GPU memory (Rhodes et al., 8 Apr 2025).

The paper also presents a large out-of-domain demonstration: a fully solvated Carbonic Anhydrase II system of about $8$5 atoms, simulated for $8$6 ps of Langevin dynamics at $8$7 K using orb-v3-direct-inf-omat (Rhodes et al., 8 Apr 2025). The reported outcome is that no unphysical behavior occurs and the structure remains close to the original PDB. This does not constitute a formal biochemical validation, but it functions as a scale and stability demonstration for mesoscale all-atom simulation.

5. External benchmarking in high-temperature MOF chemistry

An independent assessment of ORB-v3 appears in the high-temperature MOF benchmark of universal machine-learned interatomic potentials (Edwards et al., 28 Apr 2026). That study evaluates ORB-v3 without finetuning, through the orb-models package version $8$8, on $8$9 ps AIMD trajectories at Lmax⁡=3L_{\max}=30, Lmax⁡=3L_{\max}=31, and Lmax⁡=3L_{\max}=32 K for nine zinc- and zirconium-based MOFs: ZIF-8, CALF-20, MOF-10, MOF-5, MIP-206, UiO-66, UiO-67, UiO-66-NHLmax⁡=3L_{\max}=33, and NU-1000 (Edwards et al., 28 Apr 2026).

Across all temperatures and all MOFs, the study reports for ORB-v3 an energy MAE of Lmax⁡=3L_{\max}=34 meV atomLmax⁡=3L_{\max}=35, a force MAE of Lmax⁡=3L_{\max}=36 meV ÅLmax⁡=3L_{\max}=37, a stress MAE of Lmax⁡=3L_{\max}=38 MPa, and a weighted loss of Lmax⁡=3L_{\max}=39, where the energy:force:stress weights are $512$0 (Edwards et al., 28 Apr 2026). Only fairchem OMAT performs better in that aggregate table, with $512$1 meV atom$512$2, $512$3 meV Å$512$4, $512$5 MPa, and $512$6, while fairchem ODAC23 and the MACE variants are substantially worse on the same benchmark (Edwards et al., 28 Apr 2026).

Near equilibrium, ORB-v3 is reported as one of the two strongest models tested. At $512$7 K, the paper explicitly gives ORB-v3 an energy MAE of $512$8 meV atom$512$9, a force MAE of $1024$0 meV Å$1024$1, and a stress MAE of $1024$2 MPa (Edwards et al., 28 Apr 2026). The same study concludes, however, that all tested universal models are unsuitable for simulating early-stage thermal decomposition without additional finetuning once the regime reaches $1024$3 K (Edwards et al., 28 Apr 2026).

The most consequential result concerns generative error. The benchmark runs $1024$4 ns of ORB-v3-driven molecular dynamics for each MOF, using ASE Langevin dynamics with a $1024$5 fs time step, a friction coefficient of $1024$6 fs$1024$7, $1024$8 K for $1024$9 ps, a linear ramp to Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,00 K over Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,01 ps at Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,02 K psFi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,03, and Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,04 K for Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,05 ps (Edwards et al., 28 Apr 2026). Frames sampled from those ORB-v3-generated trajectories are then rescored by DFT. The reported finding is that weighted loss at Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,06 K becomes Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,07–Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,08 larger than ORB-v3’s AIMD validation loss, and that the jump in loss correlates linearly with decreasing metal coordination number as bond breaking begins (Edwards et al., 28 Apr 2026).

The same paper identifies the principal failure mode as a softened high-energy potential-energy surface caused by sparse coverage of transition states, dissociative configurations, gas-phase fragments, and amorphous structures in universal-model training data (Edwards et al., 28 Apr 2026). For ORB-v3 specifically, the authors summarize the degradation as an increase in energy MAE by nearly a factor of Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,09 between Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,10 and Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,11 K, even though force MAE only approximately doubles over that temperature range (Edwards et al., 28 Apr 2026). A plausible implication is that static validation on near-equilibrium frames substantially understates the error one encounters in long reactive or decomposition MD, especially when the trajectory is generated by the model itself.

A recurrent source of ambiguity is that “ORB-v3” is also used informally in robotics and SLAM discourse to refer to ORB-SLAM3, the feature-based visual, visual–inertial, and multi-map SLAM system built around ORB features (Campos et al., 2020). That system is unrelated in domain and methodology to Orb-v3 interatomic potentials: it is a MAP-based SLAM framework for monocular, stereo, RGB-D, and inertial sensing, with keyframe-based tracking, local mapping, loop closing, and an Atlas multi-map architecture (Campos et al., 2020).

That naming ambiguity has broadened as later SLAM work treats ORB-SLAM3 as a baseline to be upgraded or specialized. “SuperPoint-SLAM3” replaces ORB with SuperPoint, uses adaptive non-maximal suppression, and in the evaluated configuration disables ORB-BoW loop closure because binary ORB descriptors and DBoW2 are incompatible with SuperPoint’s Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,12-D float descriptors (Syed et al., 16 Jun 2025). “VAR-SLAM” keeps the ORB-SLAM3 structure but adds a YOLOv4-based semantic keypoint filter and Barron’s adaptive robust loss to improve performance in dynamic environments while maintaining about Fi=−∇RiE,\mathbf{F}_i = -\nabla_{\mathbf{R}_i} E,13 FPS (Soares et al., 17 Oct 2025). A separate LiDAR-SLAM line applies standard ORB to rasterized LiDAR images and builds a three-thread tracking–mapping–loop-closing system over those features (Ali et al., 2021).

Earlier ORB-based visual odometry also uses the label only analogically. CFORB, for example, uses ORB as a detector with FREAK descriptors, circular matching, and stereo reprojection minimization, and has been described as conceptually resembling a next-generation ORB-based visual-odometry pipeline rather than a literal “ORB-v3” algorithm name (Mankowitz et al., 2015).

In formal bibliographic terms, however, the title “Orb-v3: atomistic simulation at scale” designates the interatomic-potential family introduced for universal atomistic simulation (Rhodes et al., 8 Apr 2025). Within current literature, that is the primary meaning attached to ORB-v3 as a named model family, while the SLAM usages remain informal or analogical.

Topic to Video (Beta)

No one has generated a video about this topic yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Sign Up to Generate

Follow Topic

Get notified by email when new papers are published related to ORB-v3.

Sign Up to Follow Topic by Email