PhaseNet

Overview

PhaseNet is a deep learning framework that identifies thermodynamic phases directly from molecular dynamics (MD) simulation trajectories. Instead of relying on hand-crafted order parameters, it learns discriminative features from raw atomic configurations.

Motivation

Traditional phase identification in MD relies on metrics like the Steinhardt bond-order parameters (Q₄, Q₆). These work well for simple systems but struggle with complex multi-component alloys or amorphous phases. PhaseNet learns these boundaries end-to-end.

Architecture

• Input: Per-atom neighbour environments (radial + angular fingerprints)
• Model: Graph Neural Network (GNN) + classifier head
• Output: Per-atom phase label (solid, liquid, gas, amorphous, BCC, FCC, HCP, ...)
• Training data: LAMMPS trajectories labelled via post-processing in OVITO

Performance

Tested on Cu-Zr binary system across 200–2000 K temperature range:

Inference is ~40× faster than OVITO's CNA on equivalent hardware.

Usage

from phasenet import PhaseClassifier

model = PhaseClassifier.from_pretrained('cu-zr-v1')
labels = model.predict('trajectory.dump')

References

• Steinhardt et al., Phys. Rev. B 28, 784 (1983)
• LAMMPS molecular dynamics simulator
• OVITO — Open Visualization Tool