JAX-Fluids
JAX-Fluids is a fully auto-differentiable CFD solver built on Google JAX, focused on end-to-end differentiable fluid simulation. Unlike traditional solvers (e.g. OpenFOAM), it back-propagates gradients through simulation parameters (geometry, boundary conditions, properties), naturally fitting Physics-Informed ML, optimization and neural-network-coupled simulation.
Core features
- Fully differentiable: derivatives w.r.t. all simulation parameters, no manual adjoint equations
- Multi-GPU: JAX's SPMD abstraction enables multi-GPU/TPU with zero code changes
- High order: WENO-JS scheme, 3rd~5th order
- Multiphase: two-phase flow (VOF)
- NN coupling: serves as the physics forward solver for Neural ODE / PINN
Use cases: Shape optimization (wing, pipe, bluff body) with gradients back to design params; PINN/PIAN forward solver; data-driven turbulence closure learning; end-to-end simulation-control co-learning.
Note: Marked MIT, but the original CPC paper part is GPLv3 - use as algorithm reference, not embedded in closed-source products. Best for research; cross-validate with OpenFOAM/ANSYS for industry.
- JAX-CFD · Google Research's differentiable JAX-based fluid solver enabling end-to-end grad
- Translating CUDA Tile Operations from Python to Rust Using Agentic AI · cuTile Rust (cutile-rs) is a tile-based system for safe, idiomatic GPU kernel au
- Accelerating Dropless MoE Training in JAX with NVIDIA Transformer Engine · NVIDIA Transformer Engine introduces JAX-optimized kernels for Mixture of Expert
- DeepXDE v1.15.0 · DeepXDE v1.15.0 released with L-LAAF activation function support for JAX backend
- AIPerf: LLM Inference Benchmarking at Scale · AIPerf is an LLM inference benchmarking tool from NVIDIA designed to evaluate th