CFD Heat Control Playground (Hugging Face Space-ready)

An interactive 2D heat equation simulator with a PID baseline and a tiny RL-like linear policy. Runs on CPU and is designed to work out-of-the-box in Codespaces or Hugging Face Spaces.

Quickstart

pip install -r requirements.txt
python app.py

Open the Gradio link and press Run.

What it does

We solve the heat equation on a uniform grid with explicit Euler:

[ \frac{\partial T}{\partial t} = \alpha \nabla^2 T + q(x,y,t) ]

• BCs: Dirichlet (fixed to ambient) or Neumann (insulated).
• Control: one or more heater patches add a source term q (W/m²). Here we ship a single centered heater.
• Reward (for plots): tracks target temperature in a small region (upper-right box) and penalizes power.

Controllers

• PID: tuned conservatively; tries to keep the target region at target_T.
• RL-Linear: a tiny linear policy on error features (ships with weights in rl/policies/heat_linear.npz).

Note: You can replace the linear policy with your own or extend to Stable-Baselines3. This repo intentionally keeps dependencies light for Spaces.

Files

• app.py: Gradio UI.
• sim/heat2d.py: heat simulator.
• sim/numerics.py: Laplacian stencil, CFL, action clipping.
• sim/viz.py: PNG plots for fields and time series.
• baselines/pid_heat.py: PID controller.
• rl/rl_policy.py: simple linear-policy loader; weights in rl/policies/heat_linear.npz.

Limits

• Simplified units (rho*cp=1) for clarity.
• Coarse grids & explicit stepping; choose small dt for stability (cfl_heat <= ~0.25 is safe).
• Only heat equation (no Navier–Stokes) in this minimal working version.

License

MIT