“`html

These so-called “100-year weather occurrences” now appear nearly routine as floods, storms, and wildfires persist in establishing new records for magnitude and devastation. To classify weather phenomena as authentic 100-year events, there must merely be a 1% likelihood of occurrence in any year. The challenge, however, is that scientists frequently lack certainty about whether the weather conforms to the present climate or contravenes the odds.

Conventional weather prediction models rely on energy-intensive supercomputers typically located at large research facilities. Over the last five years, artificial intelligence has served as an efficient mechanism for faster, more cost-effective forecasting, yet a majority of AI-driven models can only predict accurately up to 10 days into the future. Nonetheless, long-term forecasts are vital for climate science and aiding individuals in preparing for forthcoming seasons.

In a recent study published on August 25 in AGU Advances, researchers from the University of Washington utilized AI to simulate the existing climate of Earth and interannual variability for up to 1,000 years. The model operates on a single processor, needing only 12 hours to produce a forecast. In contrast, the same simulation on an advanced supercomputer would take around 90 days.

“We are creating a tool that investigates the variability within our current climate to assist in addressing the enduring question: Is a specific event the type that happens naturally, or not?” remarked Dale Durran, a UW professor specializing in atmospheric and climate science.

Durran was among the pioneers to integrate AI into weather forecasting over five years ago when he collaborated with former UW graduate student Jonathan Weyn and Microsoft Research. Durran also has a concurrent role as a researcher with Nvidia, headquartered in California.

“To educate an AI model, it requires extensive data,” Durran stated. “However, if you segment the available historical data by season, you don’t have many segments.”

The most precise global datasets for daily weather extend back to approximately 1979. Although there are numerous days in the interim that can assist in training a daily weather forecasting model, the same timeframe encompasses fewer seasons. This scarcity of historical data was considered a hindrance to employing AI for seasonal forecasting.

Ironically, the Durran group’s latest advancement in forecasting, Deep Learning Earth System Model, or DLESyM, was trained for one-day predictions, yet it still acquired the ability to capture seasonal variability.

The model integrates two neural networks: one representing the atmosphere and another for the ocean. While traditional Earth system models typically merge atmospheric and oceanic predictions, researchers had not yet adopted this methodology in models driven solely by AI.

“We were the first to implement this framework in AI, and we discovered it functions exceptionally well,” stated lead author Nathaniel Cresswell-Clay, a UW graduate student in atmospheric and climate science. “We are presenting this as a model that challenges many of the current assumptions surrounding AI in climate science.”

Since the temperature of the sea surface fluctuates more slowly than air temperature, the oceanic model revises its predictions every four days, whereas the atmospheric model updates every 12 hours. Cresswell-Clay is presently focused on incorporating a land-surface model into DLESyM.

“Our design paves the way for integrating additional elements of the Earth system in the future,” he remarked, particularly components which have been challenging to model previously, such as the interplay between soil, vegetation, and the atmosphere. Rather than researchers formulating an equation to encapsulate this intricate relationship, AI assimilates insights directly from the data.

The researchers demonstrated the model’s proficiency by contrasting its predictions of past events with those produced by the four leading models from the sixth phase of the Coupled Model Intercomparison Project, or CMIP6, all of which depend on supercomputers. Climate predictions generated by these models were essential resources employed in the recent report from the Intergovernmental Panel on Climate Change (IPCC).

DLESyM effectively simulated tropical cyclones and the seasonal rhythm of the Indian summer monsoon better than the CMIP6 models. In mid-latitudes, DLESyM captured the variations in weather patterns month-to-month and interannually at least as adeptly as the CMIP6 models.

For instance, the model accurately represented atmospheric “blocking” events in a manner comparable to leading physics-based models. Blocking involves the formation of atmospheric ridges that maintain some areas hot and dry while others remain cool or wet by redirecting incoming weather systems. “Many of the existing climate models really struggle to capture this phenomenon,” Cresswell-Clay noted. “The quality of our results affirms our model and enhances our confidence in its future predictions.”

Neither the CMIP6 models nor DLESyM claim to be infallible, but the fact that the AI-driven methodology performed competitively while consuming significantly less energy is noteworthy.

“This model not only boasts a considerably reduced carbon footprint, but it is also accessible for download from our website, permitting anyone to conduct intricate experiments, even without access to supercomputers,” Durran stated. “This renders the technology attainable for numerous other researchers.”

Additional authors comprise Bowen Liu, a visiting UW doctoral student in atmospheric and climate science; and Zihui (Andy) Liu, a UW doctoral student.

“““html

A student in atmospheric and climate studies; Zachary Espinosa, a doctoral candidate at UW specializing in atmospheric and climate studies; Raúl A. Moreno, a PhD candidate in atmospheric and climate studies; and Matthais Karlbauer, a postdoctoral fellow in neuro-cognitive modeling at the University of Tübingen in Germany.

This research received funding from 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.

For further details, reach out to Nathaniel Cresswell-Clay at [email protected] or Dale Durran at [email protected]

“`