It is now possible to model the climate system at the kilometre scale, but more work and resources are needed to harvest the full potential of these models to resolve long-standing model biases and enable new applications of climate models.
Since the mid-twentieth century, numerical models have played an instrumental role for climate sciences, for both forecasting purposes and to systematically study the mechanisms that control weather and climate. Recent years have seen the emergence of exciting new avenues, as more and more processes, such as biogeochemical cycles or ice dynamics, are included to move from pure climate models to comprehensive Earth system models. Similarly, new data-driven- and machine-learning-based methods are being used in combination with these more sophisticated models. In light of these developments, it might come as a surprise that some of the most important advances are currently seen from a seemingly ordinary evolution: the increase in spatial resolution in models.
While the resolution of climate models has gradually increased over time, recent advances in computing power now allow us to simulate the climate system at the kilometre scale. This leads to more accurate simulations, as would be expected from any increase in resolution, but also enables us to explicitly model physical processes that were only poorly represented in previous models. Doing so has the potential to overcome substantial biases that have limited the capabilities of models and to discover the mechanisms of how different components interact with each other by explicitly resolving their boundaries.
One example is the representation of clouds and convective systems in models. As shown by Julia Slingo and colleagues in a Comment in this issue of Nature Climate Change, even newer generations of climate models struggle to simulate tropical precipitation, as they are not able to properly represent mesoscale convective systems (see image). These systems are a key driver of extreme precipitation and influence weather conditions throughout the world. Hence, an accurate representation of tropical convection is crucial for our assessment of future climate impacts. The first prototypes of kilometre-scale models show their capability of reducing these biases, suggesting that kilometre-scale models will lead to a more appropriate assessment of extreme precipitation changes.
It is not only atmospheric processes that require greater spatial resolution in order to understand processes. High-resolution ocean models are equally important in order to understand changes in key ocean circulation systems and coastal impacts, as argued by Helene Hewitt and colleagues in another Comment in this issue. Mesoscale eddies determine many processes in the oceans, such as heat and nutrition transport, especially in marginal ocean areas with more complex geographies and bathymetry. Explicit representation of these systems allows reassessment of how ocean currents change with human emissions and how this will affect, for example, surface water ecosystems, carbon cycling or coastal ice shelves.
As these examples show, kilometre-scale modelling is needed for a better qualitative and quantitative understanding of our climate system. Still, it cannot be seen in isolation from other advances in modelling. Fine scales are particularly important at the boundaries between the different components of the climate system, such as the interactions between oceans and coastal ice shelves. Therefore, resolving small scales is most beneficial in comprehensive Earth system models that also explicitly include other components of the climate system such as the biosphere1 or coastal ice shelves2.
There are challenges to overcome, as modelling ever more variables at ever higher resolution will result in an unprecedented data avalanche, requiring substantial advances in both computing power and data storage and analysis. Machine-learning methods could become crucial in processing these data, but could also play an important role in accelerating the model efficiency and computing power3. At the same time, some methods, such as machine-learning-based subgrid parametrizations of small-scale processes, rely on the output from high-resolution models for training, as current observations are not sufficiently long and do not cover possible future climates for these methods to work reliably4. Therefore, a leap to kilometre-scale models is not only essential to overcome some of the most notorious biases of current climate models but will also be crucial to enable other new methods to fulfil their full potential.
The scientific need for global kilometre-scale models is clear, but we are still far from reaching this goal. So far, kilometre-scale models are mainly deployed at a regional scale or over very short time periods. The next big step will be to expand this to global models and to run these models under full emission scenarios.
Efforts are already underway, such as the European Destination Earth (https://digital-strategy.ec.europa.eu/en/policies/destination-earth) and the US Energy Exascale Earth System Model (E³SM; https://e3sm.org/) programmes, but the computational power needed to run these models and to deal with the data generated by them currently exceeds the capabilities of most modelling centres. This is particularly the case when aiming for larger ensembles of models that are needed to capture the full variability of the climate system. As Slingo and colleagues point out in their Comment, this requires not only technological improvements of single models, but a new degree of collaboration between modelling centres. Resolving the climate system at the kilometre scale is now within reach, but a renewed effort is needed to get over the line.
Couespel, D., Lévy, M. & Bopp, L. Biogeosciences 18, 4321–4349 (2021).
Rackow, T. et al. Nat. Commun. 13, 637 (2022).
Bauer, P. et al. Nat. Comput. Sci. 1, 104–113 (2021).
Yuval, J. & O’Gorman, P. A. Nat. Commun. 11, 3295 (2020).
Rights and permissions
About this article
Cite this article
Think big and model small. Nat. Clim. Chang. 12, 493 (2022). https://doi.org/10.1038/s41558-022-01399-1