How should we allocate resources for climate modelling if the goal is to improve climate-related decisions? Higher resolution, machine learning and/or storylines? A call for a deeper discussion on how we should develop the climate modelling toolbox.
Guest post by Marina Baldissera Pacchetti, Julie Jebeile and Erica Thompson
The need for “km-scale” models able to resolve fine detail at the scale of kilometres has been the subject of discussion both on this platform (see here and here) and in recent publications in peer reviewed journals (Slingo et al., 2022; Hewitt et al., 2022; Nature editorials; Stainforth and Calel, 2020). The core of the debate centres around the question of whether there should be large investments in improving the resolution of global models to the km-scale, and the extent to which these would benefit the societal response to a changing climate.
This debate has so far primarily touched on two things. Firstly, the benefits (or lack thereof) of having a more fine-grained representation of the physics. A km-scale model would have more topographical resolution and represent more physical processes, but key parameterizations are still necessary at the km-scale. Secondly, the technical feasibility of a very large model running on an exascale computer in a reasonable timeframe. This debate reflects what we call the current modelling paradigm, that focuses on increasing the resolution and complexity of GCMs in the hope that it will increase process understanding and, in time, lead to reliable fine-grained projections.
The assumption here is that more detailed GCM data will lead to better decisions, an assumption that has been disputed – especially in the context of climate services (Findlater et al., 2021) and assessing financial risks in the climate system. Moreover, it has been argued that the deluge of climate data already available needs to be assessed for quality, especially given the limitations that predicting climate change poses. Since funding on climate change research has so far been primarily channelled to the physical sciences (Overland and Sovacool, 2020), it is worth examining whether further investments in the hundreds of millions to build an exascale GCM would really allow for better decision-making.
We think that the current debate misses important fundamental aspects of how and why we do (climate) science. In a paper we recently published in BAMS (Baldissera Pacchetti et al. ,2024), we argue that funding a plurality of climate modelling strategies would better serve the decisions that need to be made in the face of climate change.
Different modelling strategies prioritise different methodological aims: for example, the “paradigm” approach described above prioritises empirical agreement with past data, realism of assumptions about the system represented, and comprehensiveness of the processes that are represented in a model, with the expectation that doing so will result in a model which is better able to predict physical climate outcomes.
But not all modelling approaches share these same aims, and different decision questions may require different types of information than physical climate outcomes. Diversifying modelling strategies therefore diversifies the type of information about the climate that is produced and the types of decision that it can support.
Machine Learning (ML) – one of the strategies we describe in our paper – for example, does not prioritise realism of assumptions, while still valuing empirical agreement, which could lead to a different perspective on model uncertainties and sensitivities to initial conditions and model structure. Users of ML approaches might find them more useful to inform decision questions relating to short-term trends and variability, and less useful for long-term physical transitions, extreme events, or the bounds of physical plausibility.
Storyline approaches – another of the examples we discuss — also differ in the aims they prioritise. By focusing on describing a causal chain of events, they prioritise intelligibility. Storylines have also been described as a “bottom up” approach to generating decision-relevant information, identifying relevant information by starting from real world events (what were the impacts? What meteorological and climatic conditions caused them?) and thereby also including the human dimension. While in many cases storylines still rely on GCM output, they can also be developed through expert elicitation. When using GCMs, there is an emphasis on better using the information that is already there with the aim of improving its intelligibility. Storyline approaches are suited to decision questions which require social or political approval or where decisions need to plan for robustness across a very wide range of possible outcomes.
In the paper, we use the metaphor of a toolbox. Machine Learning and storyline approaches are examples of modelling strategies that have different methodological aims than GCMs and also different strengths in terms of the decision questions they are suited to inform. We argue that since there are many different kinds of decisions that need to be taken in the light of climate change, providing a range of more diverse modelling tools will better address the informational needs of climate-related decision making.
Importantly, we are not proposing that GCM development should be stopped. Physical modelling of the system is needed, along with other modelling tools, because many decision questions would benefit from more accurate knowledge of future physical climate outcomes. We are also not proposing that funding should be divided equally among different approaches. Equitable funding would seek to support a range of approaches, acknowledging that some require greater investment than others, for example in computational resources. The current “paradigm” is also very resource-intensive; as such, it will be possible to achieve significantly improved diversity with relatively little diversion of funds. We are also not proposing that machine learning and storylines are the right alternatives to fund; these are simply two that we have chosen to highlight here. Ecological and sociopolitical models, Integrated Assessment Models, indigenous knowledge and climate literature are some further examples of modelling strategies which could be developed to inform different decision questions and which may communicate more effectively with different decision-making groups.
We would like to see a wider discussion about the quality and value (to different stakeholders) of different kinds of climate information, and for this to be used to support more careful decisions about what kinds of climate modelling strategies should be followed.
References
J. Slingo, P. Bates, P. Bauer, S. Belcher, T. Palmer, G. Stephens, B. Stevens, T. Stocker, and G. Teutsch, “Ambitious partnership needed for reliable climate prediction”, Nature Climate Change, vol. 12, pp. 499-503, 2022. http://dx.doi.org/10.1038/s41558-022-01384-8
H. Hewitt, B. Fox-Kemper, B. Pearson, M. Roberts, and D. Klocke, “The small scales of the ocean may hold the key to surprises”, Nature Climate Change, vol. 12, pp. 496-499, 2022. http://dx.doi.org/10.1038/s41558-022-01386-6
“Think big and model small”, Nature Climate Change, vol. 12, pp. 493-493, 2022. http://dx.doi.org/10.1038/s41558-022-01399-1
D.A. Stainforth, and R. Calel, “New priorities for climate science and climate economics in the 2020s”, Nature Communications, vol. 11, 2020. http://dx.doi.org/10.1038/s41467-020-16624-8
K. Findlater, S. Webber, M. Kandlikar, and S. Donner, “Climate services promise better decisions but mainly focus on better data”, Nature Climate Change, vol. 11, pp. 731-737, 2021. http://dx.doi.org/10.1038/s41558-021-01125-3
I. Overland, and B.K. Sovacool, “The misallocation of climate research funding”, Energy Research & Social Science, vol. 62, pp. 101349, 2020. http://dx.doi.org/10.1016/j.erss.2019.101349
M. Baldissera Pacchetti, J. Jebeile, and E. Thompson, “For a Pluralism of Climate Modeling Strategies”, Bulletin of the American Meteorological Society, vol. 105, pp. E1350-E1364, 2024. http://dx.doi.org/10.1175/BAMS-D-23-0169.1