Real-Time AI Surrogate for Fluid-Thermal Dynamics

Zero-Install Browser Inference via Fourier Neural Operators (FNO)

Interactive CFD Book Documentation on GitHub
1.5 m/s
293.15 K
Velocity X (m/s)
Default Velocity X
-0.05 m/s 0.80 1.65 m/s
Velocity Y (m/s)
Default Velocity Y
-1.40 m/s -0.65 0.05 m/s
Temperature (K)
Default Temperature
293.15 K 301.65 310.15 K

Simulation Architecture & Topology

  • Physical Topology: A 2D cross-section (3.2m width × 3.2m height) of an HVAC-ventilated room bounded by no-slip walls. Features a 0.4m high-wall cool air inlet (top-left), a 0.4m low-wall exhaust (bottom-right), and a constant 37°C solid obstacle simulating a human heat load.
  • End-to-End SciML Pipeline: First, a custom numerical CFD solver was engineered from scratch to resolve the Navier-Stokes equations. This engine generated a proprietary dataset of 500 unique fluid-thermal scenarios via Latin Hypercube Sampling. Finally, a Fourier Neural Operator (FNO-2D) was trained on this specific data to deeply understand and predict fluid-thermal behavior.
  • Interactive Inference: Adjust the velocity and temperature sliders above. The model utilizes a highly optimized PyTorch FastAPI backend to bypass traditional CFD solvers, visualizing complex non-linear physics in time.