tum-pbs/PhiFlow

Supports multiple discretization methods (grids, unstructured meshes, particles) and integrates seamlessly with PyTorch, JAX, TensorFlow, and NumPy through a backend-agnostic architecture. Enables end-to-end differentiable simulation by exposing automatic differentiation throughout the solver stack, allowing gradient-based optimization of both physics parameters and neural network models jointly. Implements classical numerical schemes (finite volume, SPH, FLIP) alongside modern techniques for inverse problems and physics-informed learning.

1,835 stars and 991 monthly downloads. Actively maintained with 1 commit in the last 30 days. Available on PyPI.