PoissoNet

PoissoNet

A physics-constrained U-Net surrogate for fast 2-D pressure Poisson solves

PoissoNet is a deep learning model that replaces the classical pressure Poisson solve in incompressible CFD simulations. Trained entirely on synthetic data generated with complex obstacle masks, the network learns to infer the pressure field that enforces incompressibility in a velocity projection step.

This model performs inference on 2-D grids (e.g. 256×256) and supports mask-based obstacle geometry, RHS divergence sources, and optional divergence penalty.

---

🧠 Model Details

• Architecture: 5-level U-Net (64 → 1024 channels), GroupNorm + GELU
• Input channels:
  • rhs – divergence field
  • mask – binary fluid/solid geometry
  • u_star, v_star – intermediate velocity components
• Output: pressure field consistent with divergence-free velocity
• Training: 40 epochs on synthetic Poisson solves using PyTorch + mixed precision on BU SCC GPUs

---

🔎 How to use

from huggingface_hub import hf_hub_download
import torch

# Download the checkpoint
ckpt_path = hf_hub_download("obelisk2u/PoissoNet", "poissonet_fp16.pt")

# Load the model (must match architecture used during training)
from scripts.train_model import UNet
model = UNet(in_ch=2, base=64)
model.load_state_dict(torch.load(ckpt_path, map_location="cpu"))
model.eval()

---

🛠️ Training Code & Data

The full training pipeline, data generation scripts, and visualization tools are available here:

👉 github.com/obelisk2u/PoissoNet

---

📄 License

This model is licensed under the MIT License.
Please cite or link back if you use it in a publication or project.

@misc{poissonet2025,
title = {PoissoNet: A Physics-Constrained U-Net for Fast Pressure Projection},
author = {Stout, Jordan},
year = {2025},
url = {https://huggingface.co/obelisk2u/PoissoNet}
}