Kelvin K. Droegemeier on ‘Working Toward a Computational Framework to Support Earth System Science and Climate Projection’

Editor’s note: NCSA is cohosting an International Summit on a Computational System for Frontier Earth System Science and Climate Simulation & Projection September 29 through October 2, 2024 at the University of Illinois Urbana-Champaign.

Toward a Computational Framework for Earth System Models at Kilometer Resolution to Support Earth System Science and Climate Projection

By Kelvin K. Droegemeier, Professor of Atmospheric Science and Special Advisor to the Chancellor for Science and Policy at the University of Illinois Urbana-Champaign

Almost daily, and often without realizing it, we rely upon computers to guide our decisions. Whether finding the fastest route using GPS on our smartphone or selecting the lowest-cost item when shopping online, computers help us make informed choices. Likewise, computers are essential in weather and climate applications, where estimating the future and knowing the likelihood of outcomes can mean the difference between life and death.

For many decades, both numerical weather prediction (NWP) models and their more complex Earth system model (ESM) cousins have been among the world’s foremost drivers of high-performance computational systems. Today, NWP models are run at operational centers across the globe, producing forecasts valid for a few hours out to two weeks on scales ranging from local to global. Such models also are used routinely for academic and commercial purposes.

Similarly, extremely sophisticated ESMs – usually developed by large teams or the broader community and made available as community resources – are now run by some 50 groups globally to understand the behavior of our planet and to project future climate states under a wide variety of scenarios regarding human and natural activity. Importantly, such work is foundational to major climate assessments in the United States and internationally.

Contemporary NWP models and ESMs share many characteristics, including requiring the most powerful computers available and needing to quantify uncertainties inherently present in their solutions. They also differ in several ways. Whereas both solve complex equations governing atmospheric motion, thermodynamics, cloud and precipitation processes, radiation physics, land-surface/atmosphere interactions, vegetation, soil moisture and groundwater, ocean circulations and sea ice, ESMs also include slowly changing processes associated with atmospheric gases and particulates, biogeochemical cycles, ice sheets and marine ecosystems. ESMs also accommodate natural forcings such as volcanic eruptions and variations in solar output, as well as anthropogenic forcings, such as human-induced greenhouse gas emissions and land use and land cover changes.

Because NWP models forecast weather for around two weeks using domains ranging from regional to global, they can employ fine spacings in their computational meshes. Conversely, because climate models project future climate states many decades to a century or more in the future, and because they also involve additional physical processes, they are forced to use much coarser meshes, leading to poorer representation of coastal processes, land-atmosphere processes, tropical cyclones, topographic forcing, cloud and precipitation processes, and extreme events. These limitations sometimes manifest as significant biases and misrepresentations of key factors governing the weather and climate system.

That is not to say ESMs are somehow horribly flawed. Indeed, they are elegantly sophisticated and from them we have learned a great deal about the behavior and global and regional trends of the atmosphere, oceans and other elements of the Earth system. The sad truth, however, is that ESMs are like sleek Formula 1 racecars, which, owing to current computational limitations, can only be driven in a small parking lot. Artificial intelligence (AI) and machine learning (ML) are fundamentally transforming our ability to forecast weather, mostly because of the ability to train ML-based models using decades of observations and forecasts. Although such advances are evolving more slowly in climate projection, they are playing increasingly important roles in ESMs and are an integral part of an international summit coordinated by the University of Illinois Urbana-Champaign this fall.

Kelvin K. Droegemeier, Professor of Atmospheric Science and Special Advisor to the Chancellor for Science and Policy

The above discussion leads to a particularly important question. If we were somehow able to run ESMs at global spatial resolutions like those of today’s NWP models – resolutions that would unlock the full capability of ESMs and uncover masked limitations – would this yield significant improvements in our understanding of climate change and the ability to project future climate states? Interestingly and fortuitously, a similar question began to be addressed a few decades ago for NWP when model resolutions were taken experimentally from a few tens of kilometers down to a few kilometers and the models initialized with fine-resolution data, especially from Doppler weather radars (e.g., Lilly 1991, Droegemeier 1997). Subsequent results were stunning (e.g., Youssof et al. 2016), leading some to characterize the resulting transition of such models to operations as a revolution (The Royal Society 2021). Equally stunning is that theory suggested events such as convective storms were numerically predictable for only an hour or so (Lorenz 1969). Today, such weather is being forecast several hours in advance using operational NWP models, transforming our ability to anticipate, manage and recover. Could such a revolution occur in Earth system modeling?

Numerous studies have made clear (e.g., Palmer and Stevens 2019) that the broad dimensions of climate change (e.g., global warming due to increases in greenhouse gas concentrations) have long been known and do not require sophisticated ESMs run at high resolution. However, such models run at high resolution are needed to address sources of model bias, which in some cases is significant; the pace of warming; local and regional impacts; and details of clouds and precipitation, including severe storms, hurricanes and flash floods (Palmer and Stevens 2019). Although such resolutions are not yet available globally, they have been achieved regionally via dynamical downscaling, which involves embedding a fine-scale ESM computational mesh within a coarser global mesh. Such experiments have yielded tremendous insight into local and regional impacts. Unfortunately, they are, in most cases, unable to capture the upscale and downscale interactions which connect local and regional processes with global ones, and often do not provide data for regions that are most vulnerable to climate change impacts such as the Global South.

Reflecting on the importance of model resolution and how lessons learned from NWP might be leveraged in climate modeling, The Royal Society (2021) noted that “…the lack of capacity to simulate [physical processes] in fine detail accounts for the most significant uncertainties in future climate, especially at the regional and local levels.” They go on to say, “Meanwhile, over those same 30 years of IPCC reports, there has been a ‘quiet revolution’ in weather forecasting. Weather models are now delivering global predictions at 10-kilometre resolution, and regional forecasts at the kilometre-scale. So why hasn’t climate modelling followed the same path? Quite simply it is because the scale of computing power required to perform multi-century global simulations for multiple scenarios at pace eclipses what is needed to make operational weather forecasts for the next few weeks.”

It is clear the value of fine resolution for NWP is no longer debated and has at least partially inspired approximately a dozen groups around the world to establish ESMs that can be run at global kilometer mesh spacing. Many of these groups are incorporating AI and ML in creative ways to improve accuracy and performance (e.g., Kochkov et al. 2024), recognizing that numerous model runs are needed in an ensemble context to quantify uncertainty (e.g., Bouallegue et al. 2024). Indeed, fine-resolution ESMs are not a “silver bullet” and should be utilized in tandem with AI/ML tools to achieve desired goals. Yet, the need for fine-resolution ESM codes is only part of the story. The other part involves the computational systems on which they are to be run and leads to the following question: Will the evolution of computational systems and AI/ML capabilities, as presently envisioned, be sufficient to support large ensembles of kilometer-scale global climate simulations soon enough to help answer key scientific questions, to train ML-based models of the climate system and to support local and regional decision-making regarding climate change?

The answer would appear to be an emphatic “no!” And with that answer comes the imperative of charting a clear course going forward. Two principal options exist, the first being to continue down the projected path of computational capability, utilizing regional high-resolution climate simulations, advances in AI/ML and dynamically downscaled coarser-grid global simulations on general-purpose machines to understand climate change and its related regional and local impacts. Indeed, such work must continue because its value is indisputable and because advances in AI will continue to improve overall capabilities. The second option is to dramatically disrupt the current technology trend by creating a specialized computational system for Earth system science research and climate projection that accommodates both ESMs and new approaches that couple or tether such models to AI.

The Department of Climate, Meteorology & Atmospheric Sciences has partnered with the National Center for Supercomputing Applications to begin addressing the second option by holding an International Summit on a Computational System for Frontier Earth System Science and Climate Simulation & Projection on its Urbana campus from September 29 through October 2, 2024. Sponsored by a grant from the National Science Foundation, this summit will bring together thought leaders from academia, industry, government and non-profit organizations to examine the need for a computational system and modeling framework that accommodate ESMs run at kilometer resolution globally. Additionally, the summit will address how a globally engaged community framework could be structured to meet this need.

The summit is being organized by the following international team of experts who provided thoughtful comments on this article: Francina Dominguez (University of Illinois Urbana-Champaign), Kelvin K. Droegemeier (University of Illinois Urbana-Champaign), Barb Helland (U.S. Department of Energy, Retired), Ruby Leung (Pacific Northwest National Laboratory), Maria Molina (University of Maryland College Park), Andreas Prein (National Center for Atmospheric Research), Dan Reed (University of Utah), John Shalf (Lawrence Berkeley National Laboratory), Bjorn Stevens (Max Planck Institute for Meteorology) and John Towns (University of Illinois Urbana-Champaign).

This invitation-only event is expected to gather 150 experts in person with more than 50 online. The summit also will be streamed live globally on NCSA’s YouTube channel. More information can be found on the summit webpage.