As scientists work to forecast climate patterns from years to as much as a decade in advance, the National Center for Atmospheric Research (NCAR) has created a vast new set of computer simulations to help identify the types of events that are most predictable.

For example, an early analysis of the data set, which contains a staggering 24,800 simulated years of climate information, finds there is potential to predict sea surface temperatures in the North Atlantic, which are tied to climate conditions across Europe and Asia. The analysis also finds that multi-year precipitation anomalies — wetter- or drier-than-average conditions — over parts of Africa, including the Sahel, as well Europe and Eurasia, may be predictable as well.

The NCAR scientists who led the effort, called the Decadal Prediction Large Ensemble (DPLE), expect other areas of predictability to emerge as experts from across a wide range of fields begin to dig into the data set, which is freely available to the research community. For example, the DPLE contains data on ocean biogeochemistry, which will allow scientists to search for predictability in aspects of the ocean that affect fisheries and the global carbon cycle.

The massive new data set, which is described this month in the Bulletin of the American Meteorological Society, was created using the NCAR-based Community Earth System Model (CESM). For each year between 1954 and 2015, the scientists used historical observations to create initial conditions for 40 model simulations and then allowed the simulations to run forward 10 years. The bulk of the simulations were run on the Cheyenne system at the NCAR-Wyoming Supercomputing Center.

"There are groups all over the world working on using observations to initialize decadal predictions," said NCAR scientist Stephen Yeager, the lead author of the paper. "What makes our effort unique is the number of model simulations and the sheer computational resources it took to generate this data set. Having a data set this size allows you to explore a lot of unanswered questions in this field that no one else can."

The creation of the data set was funded by National Science Foundation, which is NCAR's sponsor, the National Oceanic and Atmospheric Administration, and the U.S. Department of Energy.

Finding signals in the noise

In recent decades, scientists have improved short-term, localized weather forecasts that go out about a week or a little longer, but the natural chaos of the atmosphere makes it impossible to predict precise weather conditions at a particular time and place more than about two weeks in advance.

Scientists, however, are interested in whether regional climate patterns — from unusually cold winters in Europe, for example, to multi-year droughts in northern Africa — might be predictable months, years, or even a decade in advance. Such long-term forecasts would be invaluable for farmers, utility managers, and many industries that could benefit from extended planning.

The DPLE offers scientists a new tool, unprecedented in scope, to examine the possibility of making these kinds of long-term predictions on time scales of a few years to a decade out. The DPLE comprises 62 individual ensembles (or sets of simulations), one for every year between 1954 and 2015. Each ensemble consists of 40 members running forward a decade in simulated time.

The ensembles were kicked off using real-world historical observations for Nov. 1 of the year they were started, but the initial temperature conditions for each of the 40 members were tweaked ever so slightly — by less than a trillionth of a degree. Those tiny differences create striking variety across the ensemble, representing the natural variability in the climate system. By averaging together the individual members of the ensemble, however, the scientists can also pick out signals from the noise. Those signals point to areas of potential long-term predictability.

For example, if scientists analyze the average of the 40 ensemble members from one particular year and find a trend — perhaps drier-than-average conditions over a particular land area for multiple years — they can then look to see if the trend actually occurred in the historic record.

"The 40 ensemble members evolve differently, producing a spread of results over time, due to the chaotic nature of the climate system," Yeager said. "But when you take the ensemble mean, you average out all those chaotic variations and what you are left with is the signal. If that signal corresponds with observed reality, you have a skillful prediction."

Aside from its incredible size (the DPLE clocks in at roughly 600 terabytes), it has another unique advantage compared to other decadal prediction efforts. The DPLE builds on and complements another big data project at NCAR known simply as the Large Ensemble. The Large Ensemble, which also uses CESM and is based on the same protocol, consists of a single 40-member ensemble stretching from 1920 to 2100. Unlike the DPLE, it is not initialized using real-world conditions. Instead, its ensemble mean “signal” reflects the impacts of human-caused climate change as well as variations in the amount of solar radiation driving the climate system.

Being able to compare the DPLE and the Large Ensemble allows researchers to determine whether apparently skillful DPLE predictions are connected to an actual ability to forecast natural, internal cycles — such as changes to large ocean currents that transport heat around the globe — or rather to changes related to human-caused climate change, including a general warming of global surface temperatures.

"No one else in the world has these two large ensembles that can be compared and contrasted in this way," Yeager said.

See all News