Official Resources
- Source Repository: https://github.com/torchmd/torchmd-net
- Documentation: https://torchmd-net.readthedocs.io/
- Conda: https://anaconda.org/conda-forge/torchmd-net
- License: Open source (MIT)
Overview
TorchMD-NET is a PyTorch-based neural network potential for molecular dynamics. It implements multiple architectures including TorchMD-Net, PaiNN, and Transformer-based models, with pretrained models and efficient MD integration.
Scientific domain: Neural network potentials for molecular dynamics
Target user community: Researchers running ML-driven molecular dynamics
Theoretical Methods
- TorchMD-Net architecture
- PaiNN (polarizable atom interaction)
- Transformer-based potentials
- Tensor field networks
Capabilities (CRITICAL)
- Multiple architecture support
- Pretrained models
- TorchMD integration for MD
- ASE calculator
- Conda installable (790K downloads)
Sources: GitHub repository, arXiv:2202.02541
Key Strengths
Multi-Architecture:
- PaiNN, TorchMD-Net, Transformers
- Easy architecture switching
- Benchmarking framework
MD Integration:
- TorchMD for simulations
- ASE calculator
- Efficient inference
Popular:
- 790K+ conda downloads
- Active development
- Well documented
Inputs & Outputs
- Input formats: Molecular structures
- Output data types: Energies, forces
Interfaces & Ecosystem
- TorchMD: MD engine
- ASE: Calculator
- PyTorch: Backend
Performance Characteristics
- Speed: Fast (GPU)
- Accuracy: Architecture-dependent
- System size: Molecular
- Automation: Full
Computational Cost
- Inference: Milliseconds
- Training: Hours on GPU
Limitations & Known Constraints
- Molecular focus: Primarily non-periodic
- No universal model: Need training
- TorchMD dependency: For MD
Comparison with Other Codes
- vs SchNetPack: TorchMD-NET is MD-focused, SchNetPack is broader
- vs MACE: TorchMD-NET is molecular, MACE is universal
- Unique strength: Multi-architecture NN potential with integrated MD (TorchMD)
Application Areas
Molecular MD:
- ML-driven molecular dynamics
- Property prediction
- Conformational sampling
Architecture Research:
- Benchmarking architectures
- PaiNN vs Transformer
- Model comparison
Best Practices
- Start with PaiNN for balanced performance
- Use pretrained models when available
- Validate with QM benchmarks
Community and Support
- Open source (MIT)
- Conda installable
- TorchMD team maintained
- ReadTheDocs documentation
Verification & Sources
Primary sources:
- GitHub: https://github.com/torchmd/torchmd-net
Confidence: VERIFIED
Verification status: ✅ VERIFIED
- Source code: ACCESSIBLE (GitHub)
- Conda: AVAILABLE (790K downloads)
- Specialized strength: Multi-architecture NN potential with integrated MD (TorchMD)