Source-linked AI summary
Orb-v3: atomistic simulation at scale
Benjamin Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, Mark Neumann
TL;DR
Universal interatomic potentials must combine accuracy with scalability for large-system and high-throughput simulation. Orb-v3 explores conservatism, roto-equivariance, neighbor limits, and datasets across a performance-speed-memory frontier, finding that lightweight direct models can retain strong physical-property accuracy. The released family therefore spans workflow-dependent trade-offs while targeting faster, larger-scale atomistic simulation.
Problem
Universal MLIPs need to model realistic large systems while maintaining accurate predictions across diverse chemical systems and physical properties.
Method
Orb-v3 varies conservatism, maximum neighbor limits, training datasets, and rotational regularization to construct models across the performance-speed-memory Pareto frontier.
Results
Orb-v3 direct models can match or exceed expensive physically constrained models on established physical-property benchmarks while offering substantially improved speed and memory profiles.
Takeaways & Limitations
The model family supports workflow-dependent choices between direct and conservative potentials for high-throughput and mesoscale all-atom simulations.
Takeaways & Limitations
Neighbor limits improve efficiency but can penalize some calculations, likely because they create discontinuities in the potential energy surface.
Abstract
from arXiv · showhide
We introduce Orb-v3, the next generation of the Orb family of universal interatomic potentials. Models in this family expand the performance-speed-memory Pareto frontier, offering near SoTA performance across a range of evaluations with a >10x reduction in latency and > 8x reduction in memory. Our experiments systematically traverse this frontier, charting the trade-off induced by roto-equivariance, conservatism and graph sparsity. Contrary to recent literature, we find that non-equivariant, non-conservative architectures can accurately model physical properties, including those which require higher-order derivatives of the potential energy surface. This model release is guided by the principle that the most valuable foundation models for atomic simulation will excel on all fronts: accuracy, latency and system size scalability. The reward for doing so is a new era of computational chemistry driven by high-throughput and mesoscale all-atom simulations.
Orb-v3 Models
Orb-v3 models share Orb-v2’s architecture and diffusion pretraining while varying conservatism, neighbor limits, and training dataset to navigate the performance-speed-memory frontier.
- Orb-v3 retains Orb-v2’s basic architecture and diffusion pretraining scheme.
- The main design variables are whether forces and stress derive from energy gradients, the maximum neighbors per atom, and the training dataset.
- Direct and conservative models occupy different workflow-dependent trade-offs, so Orb-v3 releases both types.
- Equigrad regularizes the energy gradient with respect to an identity rotation to encourage rotational invariance in conservative models.
- Neighbor limits can penalize some calculations, likely because they introduce discontinuities in the potential energy surface, but remain useful for another Pareto-frontier region.
- OMat-trained models use only OMat24’s AIMD subset because rattled structures harmed performance on out-of-distribution hetero-diatomic systems.
Speed and Memory
Orb-v3 direct models target the latency and memory bottlenecks of long molecular-dynamics trajectories and large periodic systems through faster inference and efficient graph processing.
- Molecular dynamics requires millions of sequential MLIP calls because femtosecond time steps are needed to reach nanosecond-scale phenomena.
- Hundreds rather than tens of forward passes per second allow orb-v3-direct-* models to exceed one million steps per hour for small systems.
- At 100,000 atoms, orb-v3-direct-20 uses 32.8GB of GPU memory and completes in under half a second, while baselines and conservative models encounter OOM errors.
- Expensive graph construction can dominate state-of-the-art MLIP runtime, motivating brute-force and GPU-accelerated nearest-neighbor solutions.
Benchmark Results
Orb-v3 is evaluated across geometry optimization, phonon, thermal-conductivity, molecular-dynamics, and mechanical-property workflows, showing competitive accuracy alongside substantially improved efficiency.
- Orb-v3 benchmarks cover geometry optimization, phonon calculations, and molecular dynamics across established computational workflows.
- Rotationally non-invariant direct models achieve competitive thermal-conductivity results while supporting second- and third-order potential-energy derivatives.
- Speed and Memory: Figure 2 compares relative runtime and maximum GPU memory on an NVIDIA H200 as atom count varies, including both inference and graph construction.
- Physical Property Predictions: Table 2 spans phonon frequencies, vibrational entropy, free energy, heat capacity, and mechanical properties from elastic-constant benchmarks.
- Physical Property Predictions: orb-v3-conservative-inf-omat achieves the highest accuracy for almost all Table 2 metrics while being faster than the best-performing literature models.
- Physical Property Predictions: orb-v3-direct-20-omat is comparable in accuracy to current literature state of the art and about 30 times faster than SevenNet.
Equigrad - Learned Rotational Invariance
Equigrad measures and regularizes rotational sensitivity through the energy gradient with respect to an identity rotation, improving invariance and robustness in symmetry-dependent workflows.
- Equigrad quantifies rotational invariance by differentiating predicted energy with respect to an identity rotation inserted at the model input.
- Invariant models have Δrot = 0, whereas augmented non-invariant models have small nonzero ||Δrot|| values indicating hypothetical rotational energy changes.
- For conservative models, Δrot can be L2-regularized during training without additional cost because forces and stress already require a backward pass.
- Equigrad improves Orb-v3 rotational invariance by approximately 5x and improves robustness in crystal-symmetry-based thermal-conductivity workflows.
Uncertainty Estimates
Orb-v3 adds an intrinsic confidence head that predicts binned per-atom force errors from final node representations. The predicted confidence correlates with atom-level Force MAE, including on out-of-distribution datasets.
- Orb-v3’s confidence head predicts binned atomic force errors from final per-atom node representations.The head is included in all Orb-v3 models.
- The confidence procedure discretizes force-error magnitudes into bins ranging from 1 to 50.The supplied algorithm passage lists the bin sequence and per-atom confidence computation.
- Force-error targets are generated online during training, with errors capped at 0.3 Å to calibrate confidence near converged-model predictions.Detached node representations ensure only the confidence head is trained from this signal.
- Confidence bins correlate well with atom-level Force MAE even for out-of-distribution datasets.Figure 4 samples MP Traj validation systems, small-molecule optimization trajectories, and 233 relaxed IZA zeolite structures.
Conclusion
Orb-v3 expands universal interatomic potentials across accuracy, speed, memory, and system-size scalability. The release combines lightweight direct-force models with conservative alternatives and demonstrates stable simulations beyond the training distribution, while identifying graph sparsity and energy conservation as practical boundaries.
- Conclusion: Orb-v3 models redefine the performance-speed-memory Pareto frontier, with lightweight potentials matching or exceeding expensive physically constrained models.The conclusion specifically compares their accuracy with MACE and SevenNet families.
- Conclusion: The release spans conservative, highly roto-equivariant models and non-conservative models with sparser graph featurization, covering different speed-memory-accuracy trade-offs.The models are positioned at multiple points on the performance-speed-memory frontier.
- Conclusion: Neighbor limits provide speed and memory benefits but can introduce performance loss or discontinuities, motivating future work on processing fewer edges without losing information.The authors identify adaptive graph construction and potentially edgeless architectures as future directions.
- The New Frontier: Meso-scale All-atom Simulations: Orb-v3-direct-inf-omat maintained stable, near-reference structure in a fully solvated enzyme simulation exceeding 20,000 atoms despite being extremely out-of-distribution.The simulation used Langevin dynamics at 300 K without physical constraints, and additional validation remains necessary.
- Appendix B: Lessons from Orb-v2: Workflow choice matters: only orb-v3-conservative-inf rigorously satisfies continuous conservative-force requirements for constant-energy dynamics.Non-conservative models may be unsuitable for NVE workflows because they do not conserve energy.
Appendix C: Orb-v3 modelling updates
Orb-v3 introduces modeling and systems updates that span speed, memory, graph construction, conservatism, and physical-property workflows. The models trade architectural constraints and neighbor limits against scalability and reliability in different simulation settings.
- Model updates: Increasing MLP width from 512 to 1024 enables a five-layer model with approximately the same ∼25M parameters as Orb-v2 but 2−3× faster.The change prioritizes width over depth in the GNS backbone.
- Conservatism: Orb-v3 releases direct and conservative models, with conservative variants computing forces and stress as energy gradients.Conservative models impose energy-force and stress consistency through backpropagation.
- Graph sparsity: Controllable neighbor limits reduce graph-construction and forward-pass costs but introduce PES discontinuities and modest penalties in some workflows.Unlimited-neighbor and 20-neighbor variants occupy different points on the performance-speed-memory frontier.
- Model updates: Orb-v3 adds a confidence head and compiles models by default in PyTorch while supporting dynamic graph sizes.Compilation requires torch==2.6.0 because earlier versions contain an RMSNorm-related compilation bug.
- Efficient graph construction: Adaptive graph construction is needed across system sizes: GPU overhead dominates small systems, brute-force GPU routines suit somewhat larger systems, and memory-efficient GPU methods become important at very large sizes.The preferred implementation can therefore change with system scale.
- Efficient graph construction: Dynamic supercell construction computes the minimum unit-cell tilings needed for correct periodic graph construction instead of always using a fixed 3×3×3 supercell.The fixed approach depended on neighbor limits, radius cutoffs, and minimum unit-cell dimensions.
- Energy conservation: Only orb-v3-conservative-inf rigorously conserves energy in constant-energy dynamics; neighbor limits, non-conservative forces, and discontinuities produce energy drift in other variants.This model is recommended when physical properties are calculated from NVE dynamics.
- Higher-order derivatives: Thermal-conductivity error depends on both Phonopy displacement step size and floating-point precision in force evaluation.A mixed-precision experiment separates low-precision geometry relaxation from high-precision finite-difference force evaluations.
Appendix G: MDR benchmark and mechanical properties
The benchmark protocols evaluate phonon and mechanical properties after geometry and cell optimization, using specified finite-difference and strain settings. These procedures combine FIRE-based optimization with ASE cell filtering and fixed numerical perturbations.
- Phonon MDR: For the phonon MDR, Phonopy generates displacements and second-order force constants after FIRE and FrechetCellFilter optimization.The protocol uses 0.01 Å displacements, is_plusminus=True, and a default q-mesh of [20].
- Mechanical properties: Mechanical-property evaluation sub-samples 1,000 materials and optimizes positions and cell components before applying strain displacements.The procedure evaluates bulk and shear moduli using normal strains of ±0.05 and ±0.1 and off-diagonal strains of ±0.01 and ±0.02.
- Mechanical properties: After strain is applied to the optimized unit cell, atomic positions are optimized at fixed cell, following the original Materials Project protocol.The selected strain magnitudes were chosen for agreement with PBE reference values, though possible error cancellation is acknowledged.
Appendix H: Distillation for direct models
Direct Orb-v3 models can struggle with higher-order energy derivatives when trained on MPA, motivating distillation from conservative models. The report presents this as a dataset-dependent issue and notes that the distillation method remains basic.
- Distillation approach: Distillation from orb-v3-conservative-inf-mpa resolves the direct-model derivative problem by generating static energy, force, and stress predictions across MPA.The conservative teacher supplies targets for direct-model training.
- Dataset dependence: The need for distillation appears dataset-dependent: it is observed for MPA but not when direct models are trained on OMAT.The report attributes the difference to a possible quirk in the MPA force distribution, whose exact nature remains unresolved.
- Open questions: The distillation procedure is basic and does not use newer Hessian-based methods for MLIPs.Whether similar degradation occurs across other downstream fine-tuning datasets remains an open question.
Appendix I: Effect of filtering OMat24
Filtering OMat24 addresses undesirable out-of-distribution behavior on homo-nuclear diatomics, but only removing all non-AIMD data was completely effective. The role of rattled systems and architecture remains uncertain.
- Observed behavior: All Orb-v3 models showed undesirable out-of-distribution behavior on homo-nuclear diatomics when trained on full OMat24.The issue was not general across small bulk crystals and emerged specifically in a two-atom, one-edge-per-atom setting.
- Filtering strategies: Removing all non-AIMD data was the only tested filtering strategy that completely eliminated the observed diatomic pathology.Filtering outlying energies, forces, and stresses helped but left large kinks in the energy surface.
- Open limitation: It remains unclear whether OMat24’s 45% rattled systems are generally beneficial for universal MLIPs or whether the observed problems are specific to more unconstrained architectures.The authors describe discarding valid DFT data as dissatisfying because retaining training data is generally desirable.
Appendix J: Compatibility between VASP pseudo-potentials
The appendix examines compatibility between OMat24 and MP-Traj despite their incompatible pseudopotentials. Similar gradient fields yield similar optimizations, while composition-dependent energy shifts account for much of the benchmark variation.
- Incompatible PBE pseudopotentials generate a dataset-transfer issue between OMat24 and MP-Traj.OMat24 uses PBE 52, whereas MP-Traj uses PBE 54.
- OMat-only models successfully optimize WBM test structures when evaluated by RMSD.
- A composition-based least-squares energy reinitialization substantially improves F1 while marginally changing RMSD.
- A constant shift in atomic energies explains 70% of the F1 change between the datasets.
- Combining incompatible datasets can incentivize short benchmark-specific finetuning that improves static scores without producing more effective models.