AOPP Seminar: Differentiable Modelling and GPU Acceleration for Atmospheric and Land Models

Two main ingredients of the recent AI revolution are the use of efficient GPU hardware and automatic differentiation (AD) to train immensely capable models. We can make use of these in Earth system modelling as well to tackle long standing issues with model calibration and biases and performance. Differentiable programming enables AD tools to compute gradients through code without manually defining derivatives. By formulating ESMs in a way that is differentiable via AD, we can naturally combine process-based and machine learning models and calibrate them jointly. In this talk, we will explore how differentiable and hybrid models benefit Earth system modelling, exemplified by ongoing work on the atmospheric GCM SpeedyWeather.jl and the land-surface model Terrarium.jl. We will do two deep dives: first on the way we optimized SpeedyWeather.jl’s numerics for performance across different hardware architectures. Then, we will talk about an important aspect of differentiability: How meaningful are gradients of chaotic systems? And how can we use gradients for system calibration? Recent approaches based on generalized fluctuation dissipation theorem allow us to calibrate a GCM like SpeedyWeather.jl even with respect to long-term averages and overcome previous limitations of gradient-based calibration methods. We’ll close with an outlook into future possibilities with differentiable models.