GRACE

**GRACE** (Graph Atomic Cluster Expansion) is a foundation model for interatomic potentials combining ACE features with graph neural network architecture. The GRACE-2L model (15.3M parameters) achieves state-of-art accuracy on OMat24 and…

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

Overview

**GRACE** (Graph Atomic Cluster Expansion) is a foundation model for interatomic potentials combining ACE features with graph neural network architecture. The GRACE-2L model (15.3M parameters) achieves state-of-art accuracy on OMat24 and MPTraj datasets.

Reference Papers

Reference papers are not yet linked for this code.

Full Documentation

Official Resources

  • Source Repository: https://github.com/ICAMS/grace-tensorpotential
  • License: Open source

Overview

GRACE (Graph Atomic Cluster Expansion) is a foundation model for interatomic potentials combining ACE features with graph neural network architecture. The GRACE-2L model (15.3M parameters) achieves state-of-art accuracy on OMat24 and MPTraj datasets.

Scientific domain: Graph ACE foundation model for universal potentials
Target user community: Researchers needing high-accuracy universal potential with ACE features

Theoretical Methods

  • Graph Atomic Cluster Expansion
  • ACE polynomial basis + GNN
  • Foundation model approach
  • 15.3M parameters (2L model)
  • Active learning support

Capabilities (CRITICAL)

  • Foundation model (89+ elements)
  • ACE + GNN architecture
  • 15.3M parameters (2L)
  • OMat24/MPTraj training
  • Active learning integration

Sources: GitHub repository, npj Comput. Mater. 11, 50 (2025)

Key Strengths

ACE + GNN:

  • Systematic ACE completeness
  • GNN flexibility
  • Best of both worlds
  • Physically motivated

Foundation:

  • 89+ elements
  • Pretrained model
  • Fine-tuning support
  • Active learning

Inputs & Outputs

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

Interfaces & Ecosystem

  • ASE: Calculator
  • LAMMPS: MD engine
  • Python: Core

Performance Characteristics

  • Speed: Fast (GPU)
  • Accuracy: State-of-art
  • System size: 1-10000+ atoms

Computational Cost

  • MD: ~1000x faster than DFT
  • Fine-tuning: Hours

Limitations & Known Constraints

  • New project: Still maturing
  • GPU required: For production
  • ICAMS maintained: Small team

Comparison with Other Codes

  • vs MACE: GRACE has ACE basis, MACE is pure equivariant
  • vs ACE1pack: GRACE is GNN+ACE, ACE1pack is pure ACE
  • Unique strength: Graph ACE foundation model combining systematic ACE completeness with GNN flexibility

Application Areas

Universal MD:

  • High-accuracy MD
  • Phase stability
  • Mechanical properties
  • Active learning workflows

Best Practices

  • Use GRACE-2L for best accuracy
  • Fine-tune for specific chemistry
  • Validate with DFT benchmarks

Community and Support

  • Open source
  • ICAMS maintained
  • Published in npj Computational Materials

Verification & Sources

Primary sources:

  1. GitHub: https://github.com/ICAMS/grace-tensorpotential

Confidence: VERIFIED

Verification status: ✅ VERIFIED

  • Source code: ACCESSIBLE (GitHub)
  • Specialized strength: Graph ACE foundation model combining systematic ACE completeness with GNN flexibility

Related Tools in 10.3 MLIPs DNN Universal