AeroJAX – Real-time, differentiable 2D CFD in Jax (github.com)

AeroJAX, a newly announced JAX-native computational fluid dynamics (CFD) framework, offers real-time capabilities for interactive flow research, control, and inverse design. Designed primarily for generating synthetic data to support Neural Operators, AeroJAX integrates classical numerical analysis with modern machine learning workflows, enabling researchers to modify physical parameters and solver settings on the fly. Its unique features include optimization-ready capabilities with end-to-end differentiability for adjoint-based design, an interactive physics sandbox for immediate visual feedback, and high-performance computations on standard CPUs without the need for GPUs. Significantly, AeroJAX allows for live experimentation with fluid dynamics, offering tools like freeform shape drawing and toggling between various numerical techniques while observing results in real-time. The framework supports advanced techniques for flow control and provides mechanisms for direct manipulation of simulation parameters, crucial for iterative design processes. By optimizing the architecture for CPU performance and using shared-memory buffers for efficient data transfer, AeroJAX stands out as a powerful platform for both academic research and the development of physics-consistent datasets for machine learning applications in fluid dynamics.