As smoke from the record-breaking West Coast wildfires pours across the globe and tropical storms continue to form at unprecedented rates, the state of the global climate is once again looming in the public eye. Owing to the massively complex and interconnected nature of climate systems, much of the work to understand what’s happening, why it’s happening, and how to stop it is conducted on supercomputers – but limitations in computing power have historically bottlenecked the resolution of these crucial simulations. Now, a team of climate scientists at ETH Zurich has applied supercomputing power to generate ultra-high-resolution climate simulations covering all of Europe and the central Atlantic Ocean.

In climate and weather analysis, coarse resolutions can pose serious problems, inhibiting the modeling of crucial bodies like clouds and storms. The coarseness of features like these, in turn, produce ripples in the simulations that lead to enormous uncertainties. Estimates of future temperature increases if atmospheric CO2 doubled, for instance, range between 1.5°C and 4.5°C – a huge range that Christoph Schär, a climate science professor at ETH Zurich, says “is mostly due to the low resolution of current climate models,” the best of which often still operate with broad 12- to 50-kilometer grid spacing.

Schär and his colleagues set out to shift this paradigm. They worked with the Swiss National Supercomputing Centre (CSCS) and MeteoSwiss (Switzerland’s national meteorology office) to adapt the popular COSMO atmospheric model (previously exclusive to CPUs) for use on GPUs – a move that Schär said “renders the calculations more efficient, faster and lower cost.” 

With the code in hand, the scientists received a supercomputer time allocation from the Partnership for Advanced Computing in Europe (PRACE): specifically, CSCS’ Piz Daint system. Piz Daint, a Cray System, contains 5,704 XC50 nodes (each with an Intel Xeon E5-2690 CPU and an Nvidia Tesla P100 GPU) in addition to 1,813 XC40 nodes (each with dual Intel Xeon E5-2695 CPUs). Piz Daint delivers 21.2 Linpack petaflops, placing it just inside the top ten most powerful publicly ranked supercomputers according to the most recent Top500 list. 

Using COSMO and Piz Daint, the researchers achieved an impressive leap: they generated climate projections of most of Europe (including areas in Scandinavia, the Mediterranean and Africa) at an ultra-precise 2.2-kilometer resolution. This finer resolution allowed the researchers to make new determinations – for instance, that hourly precipitation events will increase in intensity by 7 percent for each degree of temperature increase.

“In Europe, we are primarily interested in short-term heavy rainfalls, as they frequently occur in summer,” said Schär. Based on the results, he said, “the water infrastructure should therefore be adapted to more frequent and more violent rainfalls.”

Another crucial element of the new simulations was a more accurate estimate of albedo, a property that describes how much a surface reflects sunlight. The new simulations’ more accurate representation of clouds (which are relatively high-albedo) helped the researchers more closely measure their total albedo – and demonstrate that even small differences in those values can have enormous effects.

The difference in cloud cover representation between a 12km simulation (left) and a 2km simulation (right). Image courtesy of the researchers.

“This means that we must understand even the subtlest changes in tropical cloud-cover and cloud properties in order to generate accurate projections,” said Schär. “With the ability to use kilometre-scale models on much larger areas, we have come a lot closer to reaching this goal.”

Nevertheless, Schär said, “we still have a long way to go.” Even with the scientists’ leap in resolution, the simulations were limited to Europe – a lens that will need to broaden to include the globe if a truly accurate global climate model is to be created.

The post Swiss Supercomputer Enables Ultra-Precise Climate Simulations appeared first on HPCwire.