PET-MAD

**PET-MAD** is a lightweight universal interatomic potential trained on r²SCAN data. Covering 94 elements (102 in v1.5), it achieves competitive accuracy with only 2.8M parameters, making it efficient for production MD.

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

Overview

**PET-MAD** is a lightweight universal interatomic potential trained on r²SCAN data. Covering 94 elements (102 in v1.5), it achieves competitive accuracy with only 2.8M parameters, making it efficient for production MD.

Reference Papers

Reference papers are not yet linked for this code.

Full Documentation

Official Resources

  • Source Repository: https://github.com/lab-cosmo/pet-mad
  • License: Open source (MIT)

Overview

PET-MAD is a lightweight universal interatomic potential trained on r²SCAN data. Covering 94 elements (102 in v1.5), it achieves competitive accuracy with only 2.8M parameters, making it efficient for production MD.

Scientific domain: Lightweight universal potential on r²SCAN data
Target user community: Researchers needing efficient universal potential with meta-GGA accuracy

Theoretical Methods

  • Point Edge Transformer (PET) architecture
  • MAD dataset (r²SCAN-based)
  • Lightweight model (2.8M parameters)
  • 94-102 elements coverage
  • Metatrain fine-tuning

Capabilities (CRITICAL)

  • 94 elements (102 in v1.5)
  • r²SCAN training data
  • 2.8M parameters (lightweight)
  • Metatrain fine-tuning
  • ASE, i-PI, LAMMPS interfaces

Sources: GitHub repository, arXiv:2503.02089

Key Strengths

r²SCAN Training:

  • Meta-GGA quality data
  • Better than PBE-trained models
  • Higher accuracy reference
  • Chemical accuracy potential

Lightweight:

  • 2.8M parameters
  • Efficient inference
  • Fast MD
  • Low memory

Fine-Tuning:

  • Metatrain framework
  • Custom dataset adaptation
  • Transfer learning
  • Active learning

Inputs & Outputs

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

Interfaces & Ecosystem

  • ASE: Calculator
  • i-PI: MD interface
  • LAMMPS: MD engine
  • Metatrain: Fine-tuning

Performance Characteristics

  • Speed: Fast (lightweight)
  • Accuracy: r²SCAN-level
  • System size: 1-10000+ atoms

Computational Cost

  • MD: Very efficient
  • Fine-tuning: Minutes to hours

Limitations & Known Constraints

  • New project: Still maturing
  • r²SCAN data: Limited dataset size
  • Lightweight: May sacrifice some accuracy

Comparison with Other Codes

  • vs MACE-MP-0: PET-MAD is r²SCAN, MACE is PBE
  • vs CHGNet: PET-MAD is lighter, CHGNet has charge
  • Unique strength: Lightweight universal potential (2.8M params) trained on r²SCAN meta-GGA data

Application Areas

Production MD:

  • Efficient MD simulations
  • Meta-GGA quality at PBE cost
  • Structure relaxation
  • Energy screening

Fine-Tuning:

  • Custom chemistry
  • Active learning
  • Multi-fidelity

Best Practices

  • Use for r²SCAN-quality MD
  • Fine-tune with Metatrain
  • Compare with PBE-trained models

Community and Support

  • Open source (MIT)
  • Lab Cosmo maintained
  • Metatensor ecosystem

Verification & Sources

Primary sources:

  1. GitHub: https://github.com/lab-cosmo/pet-mad

Confidence: VERIFIED

Verification status: ✅ VERIFIED

  • Source code: ACCESSIBLE (GitHub)
  • Specialized strength: Lightweight universal potential (2.8M params) trained on r²SCAN meta-GGA data

Related Tools in 10.3 MLIPs DNN Universal