E(n) Equivariant Graph Neural Network
pretrained weights for 3D particle dynamics
1 min readupdated
This model implements an E(n) Equivariant Graph Neural Network for molecular conformations, particle systems, and Hamiltonian dynamics. The weight package includes a geometric message-passing inference engine built with an architecture of four equivariant layers. Weighing just 80.6 KB, the weights are delivered in the native safetensors format. The model is useful for researchers and developers working with 3D simulations and molecular modeling.
Key features
- Node Features $d_h$: 16 invariant scalar features (mass, charge, atom type)
- Edge Features $d_m$: 32-dimensional radial message representation
- Equivariant Layers: 4 continuous coordinate message-passing blocks
- Cutoff Radius: radial graph cutoff radius of 2.8 Å
What you can do
- Predict future coordinates and velocities for particle systems using a Python library
- Use the weights to model 3D molecular conformations and physical systems
- Test the engine using the accompanying interactive 3D Studio space
Example
import json
import numpy as np
from safetensors.numpy import load_file
from egnn_3d import EGNN3dDynamicsEngine
tensors = load_file("model.safetensors")
with open("config.json", "r") as f:
config = json.load(f)
engine = EGNN3dDynamicsEngine(tensors, config)
N = 8
h = np.random.randn(N, 16).astype(np.float32)
x = np.random.randn(N, 3).astype(np.float32)
x_next, v_pred = engine.predict_displacements(h, x)
print("Updated coordinates:", x_next.shape)Read more
DetailsRelative rotational residual error: < 10^{-6}%
Activation: SiLU / Swish
Weight format: model.safetensors (80.6 KB)










Comments