2026-08-14 –, Room 6
Climate models rely on parameterizations that are traditionally tuned manually. We present a differentiable calibration framework using Enzyme.jl to compute exact reverse mode gradients of energy-balance diagnostics in SpeedyWeather.jl. By batching single- timestep gradients across chaotic dynamics, we enable systematic, reproducible optimization of shortwave radiation parameters, establishing an extensible workflow for objective calibration in Earth system models.
Climate models simulate many atmospheric processes such as radiation, convection, clouds, and turbulent fluxes. As these processes are not explicitly resolved, they are represented through so-called parameterizations, which fit hundreds of parameters. Traditionally, tuning these parameters is a manual, expert-driven process guided by physical intuition and iterative experimentation given the computational cost of climate models. In this talk, we present a different approach: With automatic differentiation via Enzyme.jl we calibrate the parameters in SpeedyWeather.jl towards the observed Earth’s global energy budget. We differentiate individual timesteps in reverse mode but batch to stabilise gradients across the chaotic weather time scales. Given a target energy budget, we train on data that is continuously simulated and therefore allow slow processes to adapt despite single timestep gradients, which bypasses the need for checkpointing. Exact gradients are computed of physical diagnostics, such as top-of-atmosphere radiation and surface fluxes, with respect to model parameters that we chose for tuning. Our work focuses on the shortwave radiation, including cloud and surface albedos as the primary source of energy in the Earth system. Enzyme propagates the loss back through the full atmospheric physics, including radiative transfer and surface processes. We use Julia's optimization ecosystem for systematic calibration. SpeedyWeather is a large code base with parameters placed in most branches of large nested structs. For this we implemented a parameter handling scheme that allows convenient parameter updates and model reconstruction despite a missing central model configuration interface. The result is not only improved agreement with observed energy balance, but a reproducible and extensible workflow for parameter optimization in Earth system models. Beyond climate science, this work demonstrates how Julia enables new paradigms for scientific model development leveraging automatic differentiation for objective calibration.
Authors:
Niklas Viebig (1,2), Milan Klöwer (1), Maximilian Gelbrecht (3,4) , Brian Groenke (3), Gregory Munday (1)
- University of Oxford, UK
- ETH Zürich, Switzerland
- Potsdam Institute for Climate Impact Research, Germany
- Technical University of Munich, Germany
Master’s student in Physics at ETH Zurich, currently completing my Master’s thesis in the climate modeling group at AOPP, University of Oxford. Im researching differentiable programming and systematic parameter calibration for Earth system models, with interests in exoplanet climates, high-performance computing, and scientific software engineering.
Milan Klöwer is a NERC Independent Research Fellow at the University of Oxford. He did his postdoc at the Massachusetts Institute of Technology (MIT) working on climate model development in Julia. He started SpeedyWeather.jl, a global atmospheric model designed as a research playground to develop prototype ideas on machine-learned representations of climate processes and computationally efficient climate models. He also works on low precision computing, data compression and information theory, predictability of weather and climate, and software engineering.
Researching differentiable programming and machine learning for Earth system models and dynamical systems
- The GPU acceleration of SpeedyWeather.jl, the friendly and flexible climate model
- SpeedyWeather.jl: Towards a differentiable and GPU-capable general circulation model
- A learned surface roughness scheme for climate prediction in SpeedyWeather.jl
- Terrarium.jl: Fully differentiable and GPU-accelerated land modeling at all scales in Julia
DPhil Research Student at the University of Oxford. Interested in all things hybrid climate modelling.
I am a postdoctoral researcher at the Potsdam Institute for Climate Impact Research in Potsdam, Germany. My primary research interests are in applications of differentiable and probabilistic programming, uncertainty quantification, and scientific machine learning to geophysical modeling of Earth systems.
In my PhD, I worked on probabilistic inverse modeling of subsurface heat transfer in terrestrial permafrost. Prior to that, I worked on the application generative deep learning to statistical downscaling of climate and weather variables from coarse scale model outputs.
My industry background consists primarily of software engineering, both front-end and back-end development, with a wide range of frameworks and languages.