This article is by Dr. Asmeret Asefaw Berhe, Director of the Office of Science at the U.S. Department of Energy; this story appeared on the DOE Office of Science website.
Combining Climate Research, Artificial Intelligence, Applied Math, and Supercomputing
At the Office of Science, we spend a lot of time thinking about the future, particularly future unknowns. It’s a strategy that is especially important as we focus on making the best possible climate change predictions we can to anticipate and mitigate the effects of global warming.
Considered broadly, climate change is made up of a range of interconnected parts, everything from changing weather to what is happening in oceans, rivers, and soil systems. It’s a complex array, with contributing social and cultural factors that we need to understand well enough to make accurate predictions. We develop and support climate models that tell us how fast things will change – at scales ranging from local to global levels.
To make predictions about the climate system, researchers must have a robust database of information to work from, and that database needs to be built and maintained. It’s a strong need that resonated recently with earth systems researchers around the world.
Nicki Hickmon, of Argonne National Laboratory along with a Department of Energy (DOE) multi-laboratory leadership team, reached out last year to colleagues across the globe. They wanted to discuss the use of AI and ML to better utilize the massive tranche of big data available for Earth systems modeling.
They were impressed and pleased with the almost-viral reaction.
It was a critical point to offer opportunities for communities with vision to come together to study, research, and review needed improvements to Earth system predictability.
More than 700 participants – public and private – attended the five-week virtual workshop, which produced 156 white papers from 640 authors across more than 170 institutions.
The result: the development of solid goals for a novel and informed framework involving nine science themes related to Earth system predictions and eight cross-cut themes related to computational science and methodology. Earth system predictability themes include watershed science; coastal dynamics, oceans, and ice; and climate variability and extremes.
The goal: to radically advance prediction capabilities by using the best-available data analytics and AI.
Charuleka Varadharajan, an Earth AI and data program researcher at Lawrence Berkeley National Laboratory, notes that we need new AI methodologies that incorporate physical laws to make timely and trustworthy predictions of Earth system processes that are relevant to society.
For example, where and when will extreme events occur? How much water will be available under future climate and land use scenarios?
The AI4ESP workshop, sponsored by the Office of Science Earth and Environmental Systems Division in the Biological and Environmental Research program, and the Advanced Scientific Computing Research program, set the stage to build out a large swath of foundational science for the smart use of AI and ML tools for the next generation of Earth System Models (ESMs).
Foundational Research is Key
This is the type of foundational research we are known for: providing the basic building blocks for researchers to take their investigations to the next level. The goal is a technology infrastructure for linking observations, data assimilation, and modeling that is efficient, accurate, strategic, and convenient. Such an infrastructure will help scientists make much better predictions than is possible today.
A common theme here is the use of surrogates, which are AI software that mimics the mathematical solutions produced by parts of complex models. Surrogates often speed up ESM simulations because they avoid performing many of the calculations involved in representing localized, small-scale, and complex physical processes.
ESMs tend to be multiscale models that couple complex phenomena at many scales. Surrogate AI- and ML-based models—while likely incapable of replacing years of climate model research—can help dramatically lower model run times or integrate more small-scale phenomena into existing models. This can lead to more accurate predictions with better quantification of uncertainties under different future scenarios.
It’s exciting, according to Gary Geernaert, director of the Office of Science Earth and Environmental System Sciences Division. He acknowledges that the ideas behind the AI4ESP concept are ground-breaking, timely, and will likely lead to a paradigm shift in how predictions are made across all science disciplines, especially climate research.
And we’re back to discovery science once more.
It simply would not have been possible without the innovative thinking and vision within DOE’s national labs, the contributors at the international AI4ESP virtual meeting, and the availability of the world’s first exascale computers for this type of work. It will have a huge impact on DOE’s investment strategy, Geernaert notes.
AI and ML are crucial to predicting Earth’s response to climate change. With the new tools and concepts that are described in the AI4ESP report, scientists will be able to greatly improve the prediction of extreme weather; provide interpretable, trustworthy AI; set common standards and benchmarks; and build out data-driven and hybrid models.