Earth System Numerical Prediction Model Improves Forecasts

To forecast weather, climate and extreme events with greater accuracy, the China Meteorological Administration Earth System Modeling and Prediction Center (CEMC) in Beijing has developed an Earth system numerical prediction model, a "digital twin" super-simulator tailor-made for Earth.

It divides the Earth system components — the atmosphere, ocean, land, cryosphere and biosphere — into grid cells, then feeds observational data, including temperature, wind, air pressure and humidity, into the system.

Then using supercomputing power, the model runs calculations to predict how winds will blow, clouds form and rain fall based on the laws of physics.

Data assimilation

How can the huge amount of fragmented meteorological observation data transmitted by satellites, radars and buoys be transformed into accurate temperature, wind fields and air pressure values at the grid points of a numerical prediction model? The answer is data assimilation.

"Data assimilation is like putting together a three-dimensional panoramic picture," Han Wei, deputy chair of the scientific committee of CEMC, said. The model background field provides the base map, the observations are the scattered pieces, and the assimilation algorithm is the deft hand that stitches them together using mathematical methods to estimate the atmospheric state as accurately as possible.

Previously, the conventional data assimilation techniques used by the research team had a major limitation: They assumed that atmospheric errors were static and homogeneous.

A change in data assimilation technology was imperative. After extensive discussions, the team decided to pursue an approach known as ensemble four-dimensional variational (En4DVar) data assimilation.

But the massive amount of error information introduced by multiple data streams caused severe noise artifacts. As the computational load increased exponentially, the noise overwhelmed the genuine signals, leaving them virtually undetectable.

To counteract this, the team launched a cross-disciplinary effort, bringing in spectral filtering and wavelet-threshold denoising techniques from the field of signal processing. It was like filtering out the deafening din of a rock concert using specialized algorithms to isolate a faint but clear human voice.

Coupling Earth system components

In the past, atmospheric scientists treated the ocean surface as a fixed boundary, while ocean scientists regarded the atmosphere as little more than a secondary disturbance.

This compartmentalized approach often proved inadequate when dealing with extreme rainstorms.

The key challenge in improving forecast accuracy lies in achieving all-weather data assimilation. In conventional satellite data processing, observations from cloudy and rainy areas are often treated as noise and discarded.

"That's because cloudy and rainy areas contain non-spherical particles such as ice crystals and snowflakes. When radar and microwave signals pass through these areas, the reflected signals become highly complex and difficult to interpret," Han explained.

To overcome this challenge, the team innovatively introduced a superellipsoid model, using mathematical equations to simulate ice crystals in a wide range of shapes and orientations, from needle-like and plate-like forms to hexagonal shapes.

Data that had once been discarded as noise have now been transformed into high-resolution probes for revealing the internal structure of typhoons.

Next, the team turned their attention to atmospheric chemistry. Black carbon aerosols in the atmosphere are not merely pollutants; they are also active agents that influence radiation and cloud formation.

Led by Shen Xueshun, an academician of the Chinese Academy of Engineering, the team worked with external experts to achieve, for the first time globally, strongly coupled four-dimensional variational data assimilation of meteorological variables and black carbon aerosols.

Innovating computing

Solving the problem of accuracy is only half the battle. The forecast needs to be made faster as well.

As more components of the Earth system, including the atmosphere, land and oceans, are incorporated, the volume of data processed by the system is growing from the terabyte (TB) to the petabyte (PB) scale. If the traditional central processing unit (CPU) architecture was retained, a single global data update would take several hours, making timely forecasting virtually impossible.

The researchers moved to graphics processing units (GPUs). But the problem was that weather forecasting models are not large AI models, so they cannot be simply loaded onto GPUs, Jiang Qingu, a member of the team responsible for high-performance computing, said.

The code of numerical weather prediction models contains numerous conditional branches — instructions in the form of "if ... then ...". This logic runs efficiently on CPUs. But when transferred to a massively parallel GPU architecture, even a single branch decision can cause the entire computing queue to stall.

There was no ready-made solution to this problem. The team took the opposite approach and embarked on an intensive process of breaking down the code. They replaced every "if ... then ..." statement with a continuous mathematical function.

The logic that had previously branched into different paths was thus converted into straightforward matrix multiplication and addition operations. This enabled GPUs to execute the entire model efficiently, much like they process neural networks.

Looking ahead, Shen is confident about the future. "We will continue to pursue scientific and technological innovation, making weather forecasts faster and more accurate, and bringing every pulse of the Earth into sharper focus than ever before."

Source: Science and Technology Daily