AI for climate modelling

Machine learning for better and faster climate models. Global climate models aim to represent the key physical, chemical, and biological processes of Earth's climate. As such, these models are essential tools to conduct policy-informing climate change simulations and to understand paleo-climates. A caveat is that climate models are extremely computationally expensive and therefore require the fastest available supercomputers. However, for many types of simulations, even those supercomputers are still not powerful enough to globally resolve several important climate processes (e.g. convection, clouds, atmospheric chemistry). The relatively coarse spatial resolution of climate models (typically between 0.5 to 2.5º lat×lon) also limits their usefulness in informing societies on truly regional to local changes in climate.

Over the past two decades, an increasing number of climate models have incorporated complex chemistry solvers and chemistry–climate coupling, but doing so comes at a high computational cost. Despite advances in computational capacity, comprehensive atmospheric chemistry schemes remain too expensive for many applications, such as large ensemble simulations or convection‑permitting climate modeling. As a result, around two‑thirds of climate models participating in CMIP6 still lack an interactive representation of ozone. In this context, computationally inexpensive schemes that still allow adaptive and interactive representation of chemical components are extremely valuable.

Machine learning techniques have the potential to make climate models better, faster, more energy‑efficient, and potentially more accurate. Our group develops new machine-learning-based parameterizations to replace computationally expensive but important chemical components. The principal idea is to formulate the underlying process dependencies as a regression problem, enabling ozone on the current day to be parameterized from the climate state on the previous day. Figure 1 from Nowack et al., ERL (2018) sketches such an approach, where we proposed a ML-based ozone parameterization that generates three-dimensional daily ozone fields for climate sensitivity simulations (piCTRL and abrupt-4xCO2) using temperature as the sole input. Despite its simplicity, the approach achieved accurate and robust offline predictions owing to the strong links between ozone and atmospheric temperature. Nowack et al. (2019) further demonstrated that the parameterization can be transferred across different climate models with only straightforward data transformations.

Building on these developments, Ma et al. JAMES (2026) implemented the parameterization interactively in two climate modeling frameworks: the UK Earth System Model (UKESM) and the Icosahedral Nonhydrostatic Model (ICON). Figure 2 illustrates the mloz scheme, which parameterizes ozone in both the stratosphere and troposphere. For each ozone grid point, an independent ML function—using ridge regression—is trained to map single‑column temperatures at a given timestep to ozone concentrations at the next timestep. This linear ML formulation enables fast, accurate, and stable ozone simulations across a wide range of climate scenarios. mloz also represents ozone interactively, including two‑way ozone–circulation coupling and radiative feedbacks—an important aspect because changes in the ozone layer influence atmospheric temperatures and circulation. With atmospheric temperature profile information as the only input, mloz produces stable ozone predictions ∼31 times faster than the chemistry scheme in UKESM, contributing less than 4% of the respective total climate model runtimes. Importantly, mloz trained on UKESM data also performed well in ICON, showing it can be transferred across different models. This approach allows climate models that lack atmospheric chemistry schemes to include a fast and interactive ozone representation, supporting broader use of interactive chemistry in Earth system modeling and helping policymakers to better understand how climate may evolve.

In addition, we work on machine learning-based climate emulators.

At least for certain variables, these emulators can learn the behaviour of entire climate model systems. Such emulators offer great opportunities when it comes to accelerating the calculation of ensembles of climate model simulations, for example to explore the implications of a wide range of greenhouse gas emission scenarios following different climate policy options.

This is illustrated in the figure on the right (from Kaltenborn et al. NeurIPS 2023): first, an Earth system model, or climate model, simulates several climate change scenarios under varying scenario-dependent inputs of changes in greenhouse gas and aerosol emissions. The climate model will estimate the corresponding impacts on key variables such as regional surface temperature and precipitation (part A). In B, we then aim to avoid further expensive simulations run on supercomputers by learning the relationships between inputs and outputs from these already existing simulations stored in big data climate modelling archives. The resulting emulator systems allow for the calculation of many climate policy scenarios and their implications, far more than is possible with state-of-the-art climate models.

Selected publications:
(1) Nowack et al. Using machine learning to build temperature-based ozone parameterizations for climate sensitivity simulations. Environmental Research Letters (2018).
(2) Nowack et al. Machine learning parameterizations for ozone: climate model transferability. Conference Proceedings of the 9th International Workshop on Climate Informatics (2019).
(3) Ma et al. mloz: A highly efficient machine learning-based ozone parameterization for climate sensitivity simulations. Journal of Advances in Modeling Earth Systems, 18, e2025MS005459. (2026).
(4) Kaltenborn et al. ClimateSet: A large-scale climate model dataset for machine learning. Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track (2023).
(5) Mansfield et al. Predicting global patterns of long-term climate change from short-term simulations using machine learning. npj Climate and Atmospheric Science (2020).
(6) Watson-Parris et al. ClimateBench v1.0: A Benchmark for Data-Driven Climate Projections. Journal of Advances in Modeling Earth Systems (2022).
(7) Nowack et al. A large ozone-circulation feedback and its implications for global warming assessments. Nature Climate Change (2015).