🔬 Overview
Foundation machine learning interatomic potentials (MLIPs) are trained on overlapping chemical spaces, yet their latent representations remain model-specific and incommensurable. This work introduces the Platonic representation—a unified framework that aligns embeddings from diverse MLIPs into a common geometric space.
Key Contributions:
🎯 Universal alignment framework: Projects 7+ MLIP models into shared latent space using anchor-based transformation 📊 Model interoperability: Enables direct cross-model comparison via optimal transport and similarity metrics 🔧 Practical applications: Zero-shot model stitching, fine-tuning diagnostics, and symmetry-breaking detection 🧪 Physical validation: Correlates geometric distortions with prediction failures (phonon dispersions, rotational invariance)
🚀 Quick Start
- Load a foundation model
- Extract embeddings from structures
- Initialize Platonic transform with K=100 anchors
- Transform to unified space
- Compare models
- Compute similarity
📊 Reproducing Paper Results
Zenodo (DOI: 10.5281/zenodo.17721681)
🧮 Already convergence-tested Models
| Model | Architecture | Training Data | Status |
|---|---|---|---|
| MACE-MP-0-large | E(3)-equiv. GNN | MPtrj | ✅ Supported |
| MACE-MP-0-medium | E(3)-equiv. GNN | MPtrj | ✅ Supported |
| MACE-MP-0-small | E(3)-equiv. GNN | MPtrj | ✅ Supported |
| MACE-OMat | E(3)-equiv. GNN | OMat24 | ✅ Supported |
| SevenNet-OMat | E(3)-equiv. GNN | OMat24 | ✅ Supported |
| Orb-v3-conservative | GNN | OMat24 | ✅ Supported |
| Orb-v3-direct | GNN | OMat24 | ✅ Supported |
| NequIP-OAM-L | E(3)-equiv. GNN | OMat24, Alexandria, MPtrj | ✅ Supported |
| NequIP-MP-L | E(3)-equiv. GNN | OMat24, Alexandria, MPtrj | ✅ Supported |
| MACE-MPA-0 | E(3)-equiv. GNN | MPTrj + sAlex | ✅ Supported |
Made with ❤️ by the Walsh Materials Design Group
🗺️ Roadmap v1.0 (Current): Core framework + paper reproduction Contributions toward these goals are welcome!Last Updated: January 2025