Towards a differentiable Earth system model in Julia: Progress and perspectives

Speaker: Brian Groenke, Maximilian Gelbrecht, Niklas Boers, Potsdam Institute for Climate Impact Research

Abstract

Earth system models (ESMs) are crucial tools for understanding past, present, and future changes to our planet’s climate system. However, ESMs continue to suffer from systematic biases arising from poorly constrained parameterizations and unresolved physical processes. Purely data-driven weather prediction models have seen great success in recent years but remain difficult to apply to climate simulations due to the problem of distribution shift in past and future climates. Hybrid modeling, i.e. the incorporation of statistical and machine learning (ML) based subcomponents into a physics-based model (or vice versa), has long promised to bridge the gap by combining the advantages of both modeling approaches. A key challenge, however, in the integration of physics-based models into data-driven workflows is the need for efficient optimization (or sampling) of free parameters in the learned components. Most traditional numerical models (often written in Fortran/C++) provide limited, if any, support for automatic differentiation (AD), thereby making it difficult to train high-dimensional ML components “online” in the simulations. Thus, there is a need for a new generation of ESMs that support both AD and GPU acceleration to enable fully integrated hybrid modeling at scale. In addition, ESMs pose fundamental challenges for the AD community due to the chaotic nature of their dynamical cores, which make gradient calculations over long trajectories highly unstable. In this talk, I will give a brief overview of recent progress towards solving these grand challenges in differentiable Earth system modeling, with a focus on recent developments in the Julia programming language using Enzyme.jl. I will also discuss current roadblocks and outline a roadmap for future development.