neuralgcm/dinosaur

What it solves

Dinosaur provides a differentiable dynamical core for global atmospheric modeling. It addresses the need for modern AI weather and climate models to integrate physical laws (like the shallow water and primitive equations) with machine learning, allowing for "online training" of hybrid models.

How it works

Written in JAX, the project uses spectral methods to solve atmospheric equations on sigma coordinates. Because it is built on JAX, it supports both forward- and backward-mode automatic differentiation and is optimized for acceleration on GPUs and TPUs, including multi-device parallelization.

Who it’s for

Researchers and developers working on AI-driven weather forecasting, climate modeling, and hybrid physics-AI systems.

Highlights

• Differentiable Dynamics: Supports auto-diff for training hybrid AI/physics models.
• Hardware Acceleration: Designed for efficient execution on GPUs and TPUs.
• Spectral Methods: Solves shallow water and primitive equations (moist and dry).
• Flexible Advection: Includes semi-Lagrangian advection options that allow for 6-12x longer time steps.