China said it can now predict weather 10 days in advance at a cutting-edge resolution of 5km (3.1 miles), finer than most major systems in routine use around the world.
The advance was achieved by a Chinese-developed weather forecasting model running on Sugon 8,000, a massive home-grown computing system built entirely with domestically made artificial intelligence (AI) chips.
Most major global forecasting systems have resolutions of roughly 5km to 10km, according to the China Meteorological Administration and Beijing-based computing company Sugon, which jointly announced the results on Sept 16.
For instance, the Global Forecast System – the main global weather model run by the US National Weather Service – has a maximum resolution of about 13km.
The team has already tested the system at an even finer global resolution of 3km, according to Sugon, bringing it closer to a goal set out in China’s latest five-year meteorological plan: 1km forecasts nationwide and resolution down to 100m (328 feet) in key areas by 2030.
Numerical weather models work by dividing the atmosphere into a three-dimensional grid, then calculating how wind, temperature, air pressure and moisture change across it.
Smaller grid cells allow forecasters to see weather systems in greater detail, but at a steep computational cost. For instance, halving the distance between grid points can demand an eightfold increase in computing power.
The Chinese model is known as the Mesoscale Convective Vortex. To handle the huge workload, it was adapted to run on Sugon 8,000, a computing cluster unveiled in July that links 100,000 powerful computing units that can work on calculations simultaneously.
What sets Sugon 8,000 apart is its ability to handle different kinds of computing on the same system, according to the company. Traditional supercomputers are built for the highly precise calculations used in scientific simulations, while AI systems typically trade some precision for the ability to process huge amounts of data quickly.
Sugon 8,000 was designed to do both, using domestically developed chips, storage and networking technology. In the latest test, it produced a 10-day global forecast in about an hour.
That calculation time is not unusually fast by itself. During tests, the US Global Forecast System has completed a 10-day forecast in about 45 minutes. But Sugon said its one-hour turnaround met the requirements for routine weather forecasting.
Getting results within that window could give forecasters more time to assess approaching typhoons, heavy rain and other severe weather, it said.
Sugon 8,000 has also been used for large-scale simulations in areas including protein research, molecular dynamics and turbulence.
According to the company, the system has been adapted for nearly 200 applications and more than 500 mainstream AI models across over 30 research and industrial fields. – South China Morning Post
