A message from our Principal Investigators |
In recent months, CliMA has achieved remarkable milestones. Our groundbreaking ocean model is setting new benchmarks in usability and computational efficiency, while our land model, trained with cutting-edge satellite data, is being integrated into a wide range of research initiatives at Caltech and beyond. Additionally, we are continuously running, testing, and refining global atmosphere-land climate simulations. All codes and documentation are available on GitHub, leaderboards are live, and we are focused on further model improvements and enhancements. Five years of dedicated development are now bearing fruit, not just in the 100 papers we have published or submitted, but more importantly, in establishing a climate model development process centered on clean code, modular design, and automated pipelines that enable efficient calibration and rapid iteration. We are at an exciting stage where our continued progress is limited not only by our ingenuity and team strength—which remain crucial—but also by available computational resources. Fortunately, we have secured significant allocations from NCAR and DOE, and a collaboration with Google is providing cloud computing resources.
The coming months and year promise to be even more exciting. We are poised to showcase the first climate simulations calibrated with Earth observations and accompanied by rigorously quantified uncertainties. In this newsletter, you will find snapshots of our recent progress, offering a glimpse into why we are so delighted about our current position and optimistic about the future.
Tapio & Raf
|
|
Tapio Schneider
Theodore Y. Wu Professor of Environmental Science and Engineering
|
Raffaele Ferrari
Cecil & Ida Green Professor of Oceanography
|
|
 |
The ocean’s surface speed in an eddying simulation with CliMA-Ocean at 1/12 degree resolution. |
How We Built the World's Fastest Ocean Model |
The MIT CliMA group has developed Oceananigans, a next-generation ocean model optimized for GPUs, which achieved a significant advance in ocean modeling. Traditional climate models have struggled with accurately simulating the transport of heat and carbon by ocean eddies and the resulting climate impacts. These eddies, with scales between 10 and 100 km, are smaller than the O(100 km) horizontal resolution used in current state-of-the-art models and must be parameterized. These parameterizations introduce significant uncertainties into climate projections. By leveraging GPU-optimized computing, Oceananigans is capable of achieving a resolution of 10 km for century-long simulations. At this resolution the model can adequately simulate ocean eddies, thereby eliminating the need for their parameterizations. A sample from a 10-km resolution simulation is shown in the figure: eddies can be seen as blue rings in the surface velocity field.
This past February, our group traveled to New Orleans to present Oceananigans at “Ocean Sciences,” the leading US conference for ocean-related research. Despite presenting on the last day, we had a full house; a fire marshal had to prevent more attendees from entering the room! Oceananigans achieves a breakthrough performance of 10 simulated years per wall clock day (10 SYPD) at 10 km resolution on 64 A100 GPUs. This enables us to run 100 century-long simulations in just 10 days on a supercomputer such as DOE’s Perlmutter, which houses around 7000 GPUs. This performance is essential for climate studies that require hundreds of century-long simulations to explore different future scenarios.
More details on how we achieved this performance are available in this blog post.
|
|
|
|
Example of ParamViz.jl results |
|
Democratizing Climate Modeling and Analysis: Land Model Web Applications |
At CliMA, we believe that climate modeling and analysis should be accessible to all. That's why we're developing a web application suite that allows anyone with an internet connection to explore and interact with components of our climate models. In that vein, CliMA Research Software Engineer Alexandre Renchon has been developing web applications integrated with CliMA’s land model.
One of Alexandre’s design objectives was to create an immersive experience exploring the land model. Users can interactively run simulations driven by site-specific parameters in a virtual climate laboratory. Indeed, the app suite allows users to quickly discover how different model configurations (e.g, different modules, resolutions, solvers, parameters) impact the land model results, visualize model outputs and observations, and delve into carbon-cycle dynamics.
Alexandre is passionate about democratizing modeling tools. “Ours,” he says, “are applications that run in a web browser rather than requiring installation on a computer. Running them this way makes them accessible to a much wider audience than traditional software since all that's needed is an internet connection.” Indeed, these web applications cater to diverse audiences and provide hands-on experiences for research and classroom use.
Alexandre’s web applications integrate with CliMA’s Julia codebase, which is open-source and well documented. Thus, not only can users employ the applications, but they can also modify and extend them to meet their needs. By making our models and data accessible to everyone, we hope to foster a more collaborative and interdisciplinary approach to climate research.
|
|
Improving Climate Modeling with a Unified Cloud Scheme |
|
“Clouds hold a central but enigmatic role in the climate sciences,” says CliMA Senior Research Scientist Zhaoyi Shen. Their intricate interactions with radiation and dynamics help control Earth's temperature, precipitation patterns, and overall climate. Yet, capturing their behavior and feedback in climate models remains a daunting challenge, presenting one of the most significant uncertainties in climate projections.
That's where the EDMF (Eddy-Diffusivity/Mass-Flux) scheme comes in. The EDMF scheme is a cutting-edge turbulence, cloud, and convection parameterization scheme that we developed over the past years, building on earlier work at the European Centre for Medium-Range Weather Forecasts and at JPL. Our EDMF scheme, with a single set of parameters, captures many of Earth’s turbulence, convection, and cloud regimes, from stable boundary layers to deep tropical convection. In a unified physical framework in which we have embedded machine-learned components, it addresses several shortcomings of traditional parameterization schemes, for example, in not assuming scale separation between resolved scales and convective scales in a climate model. The EDMF scheme has the potential to produce more accurate, reliable, and robust climate projections. In local tests targeted at traditionally error-prone cumulus and stratocumulus regimes, we have seen up to 10x reduced errors in integrated cloud water content relative to state-of-the-art schemes. We are currently running and refining global simulations and addressing remaining issues such as speeding up the numerics of the scheme.
|
|
Cloud liquid water and total water flux profiles, predicted by EDMF and compared against large eddy simulations in a warmer climate. |
|
So, what makes EDMF so significant? Traditional climate models typically rely on separate schemes for boundary layer turbulence, shallow convection, and deep convection. This compartmentalized approach requires an assortment of trigger functions and semi-empirical closures, leading to compensating errors among the schemes, biases in climate models, and ultimately uncertainties in climate projections. By contrast, the EDMF scheme is a unified parameterization scheme that seamlessly integrates these processes, providing a more holistic representation of turbulence, convection, and clouds. Moreover, uncertain closure functions in the EDMF scheme are built to be calibrated with climate data. Graduate student Costa Christopoulos and former student Ignacio Lopez-Gomez have shown how machine-learned functions such as neural networks and linear regression models can be embedded in the scheme and trained end-to-end to match high-fidelity data, further enhancing its performance and reducing uncertainties.
By reducing uncertainties in climate models and providing more reliable climate projections, EDMF can help us better understand the risks and impacts of climate change, enabling us to make informed choices that safeguard our planet and ensure a sustainable future.
|
|
Pictured from left to right: Julia Sloan, Polina Khapikova, and Ronak Patel at WCRP OSC 2024 |
|
CliMA at World Climate Research Programme Open Science Conference in Rwanda |
|
Caltech graduate students Ronak Patel and Polina Khapikova and Schmidt Software Engineering Scholar Julia Sloan traveled to present at the World Climate Research Programme (WCRP) Open Science Conference in Kigali, Rwanda this past fall.
Polina presented “Global bias corrected and downscaled projections of humid heat stress.” She enjoyed the opportunity to interact with and learn from researchers across a wide range of disciplines and from a much more diverse range of countries than at most other conferences. “The discussions about the role of climate scientists in informing the response to climate change worldwide inspired me to think of new research and future career directions.”
Ronak found the experience fascinating, hearing scientists and practitioners from dozens of countries around the world discuss the incorporation of local knowledge and beliefs in the climate services value chain. so that climate information can actually be used by those most acutely affected by climate change. He says this is especially relevant to him as he crafts research questions for his thesis and thinks about future career directions. He presented his research on "Anthropogenic fingerprints of observed temperature extremes."
Julia received a WCRP outstanding poster award for her presentation "Calibration of CliMA's land model." She says, “For me personally, it was really valuable to learn about the range of environmental research going on globally, especially hearing input from members of different communities about what they think is important for us to focus our efforts on. I also got a lot of feedback on CliMA's particular calibration pipeline, which is useful for us to keep in mind as we further develop and use those methods.”
After the conference, Julia, Polina, and Ronak toured Rwanda and neighboring Kenya parks and nature reserves and shared the beautiful photography with us.
|
|
Hiking Mt. Bisoke, Rwanda |
|
|
|
|
|
|
