AeroJAX: JAX-native CFD, differentiable end-to-end. ~560 FPS at 128x128 on CPU [P]

2026-04-29 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Advanced, quick

Summary

AeroJAX is a new JAX-native Computational Fluid Dynamics (CFD) framework designed for end-to-end differentiable Navier-Stokes simulations, specifically for integration into machine learning loops. It supports inverse design and learned closures by maintaining differentiability throughout the solver stack, including velocity, pressure, and vorticity fields. The framework features a CPU-first, vectorized implementation with no external dependencies, offering both Navier-Stokes (projection method) and Lattice Boltzmann Method (LBM D2Q9) support. It incorporates Brinkman-style forcing with smooth masks for geometry handling. Performance benchmarks show approximately 560 FPS at 128x128 resolution and 300 FPS at 512x96, with differentiable flow fields and hooks for neural operators.

Key takeaway

For AI Engineers developing physics-informed machine learning models, AeroJAX offers a critical tool to overcome the "black box" limitation of traditional CFD solvers. You can now integrate fluid dynamics simulations directly into optimization and learning pipelines, enabling gradient-based inverse design and the training of learned closures or neural operators within the solver loop, significantly enhancing hybrid model development.

Key insights

AeroJAX enables differentiable CFD simulations for seamless integration into ML optimization and learning pipelines.

Principles

• Maintain full solver differentiability.
• Integrate physics into ML loops.
• Prioritize CPU-first vectorized design.

Method

AeroJAX uses a projection method for 2D incompressible Navier-Stokes and an LBM solver, both fully differentiable, to allow gradient propagation through flow fields for inverse design and learned closures.

In practice

• Perform inverse design with gradient feedback.
• Learn turbulence closures directly in solver.
• Generate differentiable CFD data for ML.

Topics

• JAX
• Computational Fluid Dynamics
• Differentiable Simulation
• Navier Stokes
• Lattice Boltzmann Method

Best for: AI Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Build Accelerated, Differentiable Computational Physics Code for AI with NVIDIA Warp — NVIDIA Warp accelerates GPU-native physics simulations and spatial computing with Python, enabling differentiable workflows.
• Advances in Scientific Machine Learning for Coupled Fluid Flow and Transport — SciML offers efficient surrogate models for complex, computationally expensive coupled fluid flow and transport systems.
• Building a Navier-Stokes Solver in Python from Scratch: Simulating Airflow — Implementing a Navier-Stokes solver from scratch clarifies fluid physics and numerical simulation mechanics.
• The Bug That Ruined Game Physics For Decades — A new fluid simulation method prevents volume loss and handles complex dynamics by design, making advanced theory practical.
• Assessing Emulator Design and Training for Modal Aerosol Microphysics Parameterizations in E3SMv2 — Emulator design choices, including scaling and network complexity, critically influence aerosol microphysics emulation accuracy in global atmospheric models.
• Operator learning for the 2D incompressible Navier-Stokes equations: a conformal prediction approach in the data-scarce regime — The perturbation-based conformal prediction framework quantifies uncertainty in neural operators efficiently, especially in data-scarce regimes, by comparing noisy and original label predictions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.