Apply differentiable Earth system model components as instruments for gradient-based science.
Learn from data at scale: learn parameters, structures, and ML closures
Embed machine learning in physics: train learned components while coupled
Ask scientific questions in new ways: sensitivity, attribution, causality
tutorial/ contains a tutorial Jupyter Book for Differentiable Earth System Modeling with JAX, designed for scientists new to differentiable code. It covers a brief introduction to JAX applying jax.grad on a simple radiative-balance formula. We demonstrate applications including adjoint sensitivity maps, data assimilation, and hybrid modeling with online-training of ML closures on a 1-D energy balance model. We also introduce the more challenging aspects of using these tools in practice. All chapters are executed notebooks that run on a laptop.
Read it online at https://leap-stc.github.io/diff-earth/ (published by the deploy-book workflow; enable Settings → Pages → Source: GitHub Actions once).
Inspired by the M2LInES L96 demo book.