The new AI model from Dale Durran, University of Washington professor of atmospheric and climate science, and graduate student Nathaniel Cresswell-Clay, simulates up to 1000 years of the current climate using less computing power than conventional methods. It captures atmospheric conditions like the low pressure system over the central US pictured above.NASA Earth Observing System/Interdisciplinary Science (IDS) program under the Earth Science Enterprise (ESE)
So-called "100-year weather events" now seem almost commonplace as floods, storms and fires continue to set new standards for largest, strongest and most destructive. But to categorize weather as a true 100-year event, there must be just a 1% chance of it occurring in any given year. The trouble is that researchers don't always know whether the weather aligns with the current climate or defies the odds.
Traditional weather forecasting models run on energy-hogging supercomputers that are typically housed at large research institutions. In the past five years, artificial intelligence has emerged as a powerful tool for cheaper, faster forecasting, but most AI-powered models can only accurately forecast 10 days into the future. Still, longer-range forecasts are critical for climate science - and helping people prepare for seasons to come.
In a new study published on Aug. 25 in AGU Advances, University of Washington researchers used AI to simulate the Earth's current climate and interannual variability for up to 1,000 years. The model runs on a single processor and takes just 12 hours to generate a forecast. On a state-of-the-art supercomputer, the same simulation would take approximately 90 days.
"We are developing a tool that examines the variability in our current climate to help answer this lingering question: Is a given event the kind of thing that happens naturally, or not?" said Dale Durran, a UW professor of atmospheric and climate science.
Durran was one of the first to introduce AI into weather forecasting more than five years ago when he and former UW graduate student Jonathan Weyn partnered with Microsoft Research. Durran also holds a joint position as a researcher with California-based Nvidia.
"To train an AI model, you have to give it tons of data," Durran said. "But if you break up the available historical data by season, you don't get very many chunks."
The most accurate global datasets for the daily weather go back to roughly 1979. Although there are plenty of days between then and now that can be used to train a daily weather forecast model, the same period contains fewer seasons. This lack of historical data was perceived as a barrier to using AI for seasonal forecasting.
Counterintuitively, the Durran group's latest contribution to forecasting, Deep Learning Earth SYstem Model, or DLESyM , was trained for one-day forecasts, but still learned how to capture seasonal variability.
The model combines two neural networks: one representing the atmosphere and the other, the ocean. While traditional Earth-system models often join atmospheric and oceanic forecasts, researchers had yet to incorporate this approach into models powered by AI alone.
"We were the first to apply this framework to AI and we found out that it worked really well," said lead author Nathaniel Cresswell-Clay, a UW graduate student in atmospheric and climate science. "We're presenting this as a model that defies a lot of the present assumptions surrounding AI in climate science."
Because the temperature of the sea surface changes slower than the air temperature, the oceanic model updates its predictions every four days, while the atmospheric model updates every 12 hours. Cresswell-Clay is currently working on adding a land-surface model to DLESyM.
(a) a low pressure system simulated by the model in the winter of 3016, (b) an observed low pressure system in March 2018. The black lines show pressure and color indicates wind speed. Comparing the images reveals the model's accuracy.Created by Nathaniel Cresswell-Clay
"Our design opens the door for adding other components of the Earth system in the future," he said, especially components that have been difficult to model in the past, such as the relationship between soil, plants and the atmosphere. Instead of researchers coming up with an equation to represent this complex relationship, AI learns directly from the data.
The researchers showcased the model's performance by comparing its forecasts of past events to those generated by the four leading models from the sixth phase of the Coupled Model Intercomparison Project, or CMIP6, all of which run on supercomputers. Climate predictions of future climate from these models were key resources used in the last report from the Intergovernmental Panel on Climate Change (IPCC).
DLESyM simulated tropical cyclones and the seasonal cycle of the Indian summer monsoon better than the CMIP6 models. In mid-latitudes, DLESyM captured the month-to-month and interannual variability of weather patterns at least as well as the CMIP6 models.
For example, the model captured atmospheric "blocking" events just as well as the leading physics-based models. Blocking refers to the formation of atmospheric ridges that keep regions hot and dry, and others cold or wet, by deflecting incoming weather systems. "A lot of the existing climate models actually don't do a very good job capturing this pattern," Cresswell-Clay said. "The quality of our results validates our model and improves our trust in its future projections."
Neither the CMIP6 models nor DLESyM are 100% accurate, but the fact that the AI-based approach was competitive while using so much less power is significant.
"Not only does the model have a much lower carbon footprint, but anyone can download it from our website and run complex experiments, even if they don't have supercomputer access," Durran said. "This puts the technology within reach for many other researchers."
Other authors include Bowen Liu, a visiting UW doctoral student in atmospheric and climate science; Zihui (Andy) Liu a UW doctoral student in atmospheric and climate science; Zachary Espinosa, a UW doctoral student in atmospheric and climate science; Raúl A. Moreno, a doctoral student in atmospheric and climate science and Matthais Karlbauer, a postdoctoral researcher in neuro-cognitive modeling at the University of Tübingen in Germany.
This work was funded by the U.S. Office of Naval Research, the U.S. Department of Defense, the University of Chinese Academy of Sciences, the National Science Foundation of China, Deutscher Akademischer Austauschdienst, International Max Planck Research School for Intelligent Systems, Deutsche Forschungsgemeinschaft, U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research and the NVIDIA Applied Research Accelerator Program.
