Climate models

An AI solution to the gravity wave problem of climate models

Small waves, big impact

Global climate models require large amounts of computing power, such as that of the Summit supercomputer at the Department of Energy’s Oak Ridge National Laboratory. (Image credit: Oak Ridge National Laboratory)

Gravity waves are simply too small and short-lived to show up in models designed to cover the entire planet, just as fine detail is missing from low-resolution photographs. Higher resolution models can provide more detailed information, but running them on a global scale is computationally expensive for forecasts spanning more than a few weeks.

To account for smaller-scale processes like gravity waves without getting bogged down in calculations, scientists use simplified equations called “parametrizations,” which are informed by physics but don’t calculate wave oscillations and interactions. individual or do not even incorporate the limited available observational data. “We guessed what we think the gravity waves are doing at the average flow based on the variables the model can resolve,” Sheshadri said.

Even small changes in the approximations built into the gravity wave parameterizations can lead to very different regional climate projections. As a result, climate modelers are “tweaking” the parameterizations so that the results broadly resemble the climate observed today – leaving a cloud of uncertainty over how the circulation will react as people and industry add more dioxide. of carbon in the atmosphere.

Consideration of gravity waves thanks to AI

Sheshadri and Espinosa are among a growing number of researchers looking to machine learning and artificial intelligence techniques for a possible solution. “Parameterizations are a big computational sink for climate models, so if we can speed them up, that means we can increase the resolution of all kinds of things,” Espinosa said.

The researchers developed an AI-driven model, dubbed WaveNet, that can accurately mimic how dissipating gravity waves speed up and slow down atmospheric winds. The job involved building and training a set of artificial neural networks in the widely used Python programming language, then coupling them to a typical global climate model built decades ago in a 1950s language called Fortran.

Pressure balloons designed to provide internet service also collected data that researchers could use to calculate gravity wave movements. (Image credit: Loon)

The model passed two important tests. Trained on a single year of data, his predictions of how gravity waves would react to extremely high concentrations of CO2 over 800 years were similar to those produced by conventional parameterizations. And, based on a single phase of data, it accurately simulated a full two-phase cycle of the Quasi-Biennial Oscillation, a regular reversal of winds over the equator that affects surface weather patterns. and ozone depletion – and is driven by breaking gravity waves.

“WaveNet tells us nothing new about the gravity wave response to CO2. It simply does what the conventional gravity wave parametrization would have done in response to CO2 — at least, for now,” Sheshadri said.

The results are a promising first step towards the development of fully data-driven parameterizations of gravity waves, the focus of an international project that Sheshadri leads called DataWave. These parameterizations could be optimized for velocity and trained with data from high-resolution regional simulations, high-resolution but short-term global climate simulations, and a growing number of atmospheric measurements of pressurized balloons broadcasting on the internet. “I hope this will give us computational means to represent gravity waves in climate models that are physically meaningful and observationally limited,” she said. “That’s the ultimate goal of this project.”

Espinosa is now a doctoral student at the University of Washington. Stanford co-authors include Gerald R. Cain, associate professor in the Department of Computer Science. Other co-authors are affiliated with New York University, NASA’s Goddard Institute for Space Studies, and the Space Research Association of Universities.

This research was supported by the National Science Foundation, the NASA Postdoctoral Program at the Goddard Institute for Space Studies, and Eric and Wendy Schmidt on recommendation from the Schmidt Futures program.