PhiFlow Differentiable PDE Solving FrameworkNew

Concept

PhiFlow represents fields and simulations as native tensors in your DL framework of choice. That means autodifferentiation works end-to-end: you can backprop through the solver to optimize shapes, boundary conditions, or control inputs.

Features

• Stable-fluids smoke/fire simulation with API-level control of resolution
• Differentiable advection, pressure projection, viscosity
• PyTorch / TensorFlow / JAX backends behind one interface
• Designed for learned physics and inverse design (e.g. fluid obstacles)

Learning path

1. Run the smoke plume example; inspect the field tensors
2. Optimize an obstacle shape to minimize drag with autodiff
3. Add a learned correction term inside the solver step
4. Study the accompanying papers on differentiable fluid obstacles

Relation to physics-ML

PhiFlow sits at the intersection of classic graphics-style solvers and modern differentiable programming - useful when you want gradients of simulation output rather than only a learned surrogate.