PhiFlow

Why we included this project

PhiFlow is for people who want to train models that actually obey physics rather than just fitting data. You write the simulation as a differentiable Python program on grids, meshes, or particles, and run the same code through NumPy, PyTorch, Jax, or TensorFlow, so gradients flow through the solver and back to your network. That makes it practical for physics-informed learning, inverse problems, and trajectory optimization, where you need a forward simulator that is aware of gradients. It also ships working demos like smoke plumes, wake flows, and a learning-to-throw example, so you can start from something that already runs instead of a blank file. If your work is scientific machine learning, this toolkit spares you most of the wiring between a neural network and the equations.