metadata
library_name: phase
license: mit
pipeline_tag: image-to-image
tags:
- magnetohydrodynamics
- neural-operator
- diffusion
- physics-informed-machine-learning
PHASE Kelvin-Helmholtz MR
Multi-regime PHASE model for the 2-D incompressible MHD Kelvin-Helmholtz instability on t=[0,5], with gated Re/Rm adapters and residual diffusion.
- Channels: [ux, uy, Bx, By]
- Reynolds numbers: Re=Rm in {80, 200, 400, 650, 1000, 1500, 2050, 2750, 3600, 4500}
- Conditioner:
conditioner/model.safetensors - Diffusion model:
diffusion/model.safetensors - Normalization:
normalization/
POSEIDON dependency
git clone https://github.com/camlab-ethz/poseidon.git external/poseidon
git -C external/poseidon checkout b8fa28f59bd7f7673323f28d11a12c6f3a215c61
pip install --no-deps -e external/poseidon
Load
import torch
from phase import PHASEPipeline
model = PHASEPipeline.from_pretrained(
"phaseMHD/PHASE-KH-MR",
device="cuda" if torch.cuda.is_available() else "cpu",
)
# Physical initial condition: [batch, channels, x, y]
# The KH model was trained on 101 frames from t=0 to t=5.
times = torch.linspace(0.0, 5.0, 101, device=model.device)
prediction = model.predict(initial_fields, times, re=1000, seed=0)
# prediction: [batch, channels, time, x, y]
Install the runtime from PHASE:
pip install "phase[hub,dino,scot] @ git+https://github.com/PHASE-MHD/PHASE.git"
The package contains model-only weights, exact inference configs, and training-derived normalization statistics. See pipeline.yaml and provenance.json for the artifact mapping.