VortexNet: Neural network based on fluid dynamics (github.com)

A new open-source repo implements toy versions of VortexNet — a neural architecture that embeds PDE-based vortex layers and other fluid-dynamics-inspired mechanisms into standard networks. The code provides runnable examples (vortexnet_mnist.py for MNIST and vortexnext_image.py for custom image sets) that build and train VortexNet autoencoders, with features such as data augmentation, latent-space interpolation, checkpointing, TensorBoard logging, and configurable training via YAML. The authors emphasize these are educational prototypes (not optimized or physically accurate fluid solvers); requirements include Python 3.8+, PyTorch, torchvision and common analytics packages. Why it matters: VortexNet demonstrates a concrete pathway for integrating continuous PDE-style dynamics as inductive priors inside deep models, offering a fresh structural bias for tasks like compression, reconstruction and latent-space manipulation. Key technical points include the use of vortex/PDE layers to propagate representations with fluid-like advection/diffusion behavior, autoencoder training on image domains, and extensible configuration for experimenting with latent dynamics and interpolation. For ML researchers this provides an accessible testbed to explore physics-inspired layers, hybrid PDE-neural models, stability/robustness tradeoffs, and how dynamical priors affect representation learning — all without needing a full numerical fluid solver.