The Possible Futures of Climate Modeling
Like most areas of science, progress in climate science is mainly driven by new capabilities. More powerful computers give us access to new classes of simulations, and new observing systems (satellites, buoys) give us new data to test our understanding. My favorite examples come from the development of atmospheric dynamics. In the late-70s and 80s, seminal papers by Brian Hoskins and others took advantage of computers being able to resolve mid-latitude storms (1000s of kilometers across) on spherical grids. By the early 2000s, the ability to run large suites of idealized general circulation models let us systematically probe parameter space and answer questions like how the Hadley circulation depends on planetary size, rotation rate, etc. Oceanography has always been more observation-limited – we are still learning so much from the ARGO program – but has certainly benefited from easier access to simulations as well.
Today, ever-growing compute power and new computational tools have sparked a debate about the future of climate model development. One side argues that “exascale” computing facilities should be used to increase the horizontal resolution of climate models down to the kilometer scale. These models would capture many of the unresolved processes that drive uncertainty in climate projections, especially those associated with precipitation, and produce more accurate forecasts. But simulations at this scale would require pooling resources across modeling institutions, and even countries, leading Tim Palmer and Bjorn Stevens to call for a “CERN for climate change”.
The other side of the debate prefers to keep using resolutions of 10-50km for operational climate simulations, but to harness machine learning (ML) to build more accurate models. This vision imagines data-driven models that are continuously learning from new data, while keeping moderately-high resolutions to generate large ensembles of simulations. Such ensembles are essential for quantifying uncertainties — both from model parameter and structural choices, and from the chaotic nature of the climate system.
There are several possible paths along this route. CliMA, the first group to take this approach, has focused on keeping traditional physics-based model parameterizations, but automating the tuning process, a contrast to the more heuristic way climate models have traditionally been tuned. Other groups, mostly based in big tech firms, are working to replace the parameterizations themselves with ML, or even to use ML to emulate entire models.
Another way of framing the debate is between an emphasis on a few “masterpiece” runs versus having the ability to simulate a broad distribution of futures. There are also interesting questions about equity and reproducibility. Running very high resolution simulations at a small number of centers could be viewed as overly concentrating power and resources. [Though in fairness, everyone agrees climate data should be made widely available]. On the other hand, ML-based models require their own specialized infrastructure and may have opaque architectures. The presence of tech firms in this space also brings the risk of proprietary code.
Neither approach is guaranteed to succeed. Kilometer-scale models may never fully converge, while data-driven models could hit stability walls or, more subtly, get trapped in local minima that prevent accurate long-term simulations. Meanwhile, the many modeling centers which traditionally participate in CMIP are exploring how and when to use ML, and pushing to higher resolution when possible, but otherwise proceeding more-or-less as before.
These philosophical questions come at an interesting moment, as we now have long enough records to really put models to the test. Early climate models produced surprisingly accurate forecasts of global-mean surface temperature, but at finer scales clear discrepancies are emerging: the muted warming of the eastern tropical Pacific compared to model projections and the over drying of the American Southwest (even when models are given historical SSTs), for example. These discrepancies matter not only for model credibility but also for future impacts: the waters of the east Pacific influence many aspects of the climate’s response to anthropogenic forcing. So just as we are re-thinking the climate model development process, we are getting new, more stringent tests for the models.
Funding-permitting, the diversity of climate modeling approaches seems like it will teach us a lot. Climate science has benefitted in the past from having several dozen modeling centers whose models could be compared. Now there is diversity along an even broader dimension of climate modelling approaches. And we are already learning things: At CFMIP in July there were several interesting talks looking at climate feedbacks in near-km scale versions of ICON. In the longer-term, if any of these approaches succeeds in producing qualitatively better climate models – even ones that look like “black boxes” – we would gain a true laboratory for the planet, one we could ask endless questions of to learn more about how the climate system works.
