Engineers at Rice University have created a self-learning “capsule neural network.”

Researchers at Rice University have developed a computer system that taught itself how to accurately predict extreme weather events such as heatwaves and cold spells up to 5 days in advance.

Interestingly, the so-called “capsule neural network” requires very little information about current weather conditions. Instead, it uses a somewhat antiquated (analog) method of weather forecasting that became obsolete in the 1950’s due to the advent of computers.

What an unlikely combination! A method from seventy years ago combined with modern-day, high-tech artificial intelligence.

How this extreme weather forecasting works

The capsule neural network is trained to examine hundreds of pairs of maps. Each of these maps reflects surface temperatures and air pressures at the height of 5 km. Its pair shows the same conditions but several days later. The system also received input on conditions that led to extreme weather such as extended hot and cold spells, which ultimately produced devastating heatwaves and winter storms.

The network, after training, could look at previously unseen maps and generate five-day forecasts of possible extreme weather with 85% accuracy.

The online American Geophysical Union’s Journal of Advances in Modeling Earth Systems published a study about the system this week.

Limitations of the current day, computer-based, numerical weather prediction (NWP)

Computer-based numerical weather prediction (NWP) commenced in the 1950s, but even with advanced modeling and supercomputers, it failed to predict many extreme climatic events.

By looking instead at maps, the prediction model relies more on pattern recognition rather than numerical computation.

Rice’s Pedram Hassanzadeh, a co-author of the study, said an indication on the weather map usually preceded extreme conditions. Such indications could be, for example, weird behavior in the jet stream or abnormally large waves. It could even be a big high-pressure system that is not moving at all.

“It seemed like this was a pattern recognition problem,” he said. “So we decided to try to reformulate extreme weather forecasting as a pattern-recognition problem rather than a numerical problem.”

Using capsule neural networks also contributed to the accuracy of the results. Capsule neural networks can recognize relative spatial relationships, especially significant in understanding weather patterns. “The relative positions of pressure patterns, the highs, and lows you see on weather maps, are the key factor in determining how the weather evolves,” Hassanzadeh said.

What next?

Though there is a lot of work left to be done before the system can be used for real-life weather forecasting, the researchers’ immediate goal is to extend the forecast time to beyond ten days.

That’s the period beyond which NWP weather forecasting models show weaknesses.

Related Story:  AI Can Now Predict Lightning Strikes