README.md

metadata

title: Crystal Embedder
emoji: 🔮
colorFrom: purple
colorTo: blue
sdk: docker
pinned: false
license: mit

Crystal-Embedder

Compute Tri-Fusion physics embeddings for crystal structures.

What it does

This service takes a crystal structure in CIF format and returns a 2738-dimensional embedding vector that captures:

Component	Dimensions	Description
Orb-v3	1792	Force field features (PyTorch)
l-MM	758	Electronic structure features (TensorFlow/MEGNet)
l-OFM	188	Orbital field matrix features (TensorFlow/MEGNet)
Total	2738	Concatenated, L2-normalized

API

POST /embed

Compute embedding for a CIF structure.

Request:

{
  "cif": "data_Si\n_cell_length_a 5.43..."
}

Response:

{
  "vector": [0.123, -0.456, ...],
  "dims": 2738
}

GET /health

Health check endpoint.

Response:

{
  "status": "healthy",
  "models_loaded": true,
  "vector_dims": 2738
}

Example Usage

import httpx

cif_content = open("structure.cif").read()

response = httpx.post(
    "https://hafnium49-crystal-embedder.hf.space/embed",
    json={"cif": cif_content},
    timeout=60
)

embedding = response.json()["vector"]
print(f"Embedding shape: {len(embedding)}")  # 2738

Performance

• Latency: ~15 seconds per structure (CPU-only)
• Memory: ~8GB peak during inference
• Cold start: ~2 minutes (model loading)

Local Development

# Build
docker build -t crystal-embedder .

# Run
docker run -p 7860:7860 crystal-embedder

# Test
curl http://localhost:7860/health

Powered By

• MatterVial - Physics embeddings
• pymatgen - Crystal structure parsing
• FastAPI - Web framework