ORB

**ORB** (Open Reusable Bindings) is an open-source universal interatomic potential covering 117 elements with 25M parameters. Trained on multiple datasets, it achieves competitive accuracy on Matbench Discovery with efficient inference.

10. NICHE & ML 10.3 MLIPs DNN Universal VERIFIED
Back to Directory Official Website

Overview

**ORB** (Open Reusable Bindings) is an open-source universal interatomic potential covering 117 elements with 25M parameters. Trained on multiple datasets, it achieves competitive accuracy on Matbench Discovery with efficient inference.

Reference Papers

Reference papers are not yet linked for this code.

Full Documentation

Official Resources

  • Source Repository: https://github.com/orbital-materials/orb-models
  • License: Open source (Apache-2.0)

Overview

ORB (Open Reusable Bindings) is an open-source universal interatomic potential covering 117 elements with 25M parameters. Trained on multiple datasets, it achieves competitive accuracy on Matbench Discovery with efficient inference.

Scientific domain: Universal potential with broadest element coverage (117)
Target user community: Researchers needing widest element coverage in a universal potential

Theoretical Methods

  • Graph neural network architecture
  • 117 elements coverage
  • 25M parameters
  • Multi-dataset training
  • Efficient inference

Capabilities (CRITICAL)

  • 117 elements (broadest coverage)
  • 25M parameter model
  • Matbench Discovery benchmarked
  • ASE calculator
  • Efficient GPU inference

Sources: GitHub repository, arXiv:2502.20851

Key Strengths

Broadest Coverage:

  • 117 elements
  • Most of periodic table
  • Rare earth elements
  • Actinides

Performance:

  • 25M parameters
  • Competitive accuracy
  • Efficient inference
  • Matbench Discovery tested

Inputs & Outputs

  • Input formats: Structures (ASE)
  • Output data types: Energies, forces, stresses

Interfaces & Ecosystem

  • ASE: Calculator
  • PyTorch: Backend

Performance Characteristics

  • Speed: Fast (GPU)
  • Accuracy: Competitive
  • System size: 1-10000+ atoms

Computational Cost

  • MD: ~1000x faster than DFT
  • Inference: Milliseconds

Limitations & Known Constraints

  • New project: Still maturing
  • GPU required: For production
  • Limited MD integration: ASE only
  • Documentation: Growing

Comparison with Other Codes

  • vs MACE-MP-0: ORB covers 117 elements, MACE covers 89
  • vs CHGNet: ORB has broader coverage, CHGNet has charge
  • Unique strength: Broadest element coverage (117) with 25M parameter universal potential

Application Areas

Wide-Coverage MD:

  • Rare earth materials
  • Actinide simulations
  • Multi-element alloys
  • Unexplored chemistries

Screening:

  • High-throughput energy evaluation
  • Structure relaxation
  • Stability prediction

Best Practices

  • Use for systems with rare elements
  • Validate against DFT
  • Compare with other UIPs

Community and Support

  • Open source (Apache-2.0)
  • Orbital Materials maintained
  • GitHub repository

Verification & Sources

Primary sources:

  1. GitHub: https://github.com/orbital-materials/orb-models

Confidence: VERIFIED

Verification status: ✅ VERIFIED

  • Source code: ACCESSIBLE (GitHub)
  • Specialized strength: Broadest element coverage (117) with 25M parameter universal potential

Related Tools in 10.3 MLIPs DNN Universal