Chinese researchers led by Congqi Cao from Northwestern Polytechnical University have developed a new artificial intelligence system that can generate a five-day weather forecast within seconds using medical technology. The team discovered that AI models originally designed to identify tumors and analyze brain scans excel at recognizing weather patterns across regions.
These medical imaging AI systems proved superior compared to weather-specific AI systems due to their ability to identify complex patterns in multidimensional data. The researchers developed a “cascade prediction” approach that breaks the five-day forecast into smaller chunks, with separate AI models trained to predict each segment while incorporating information from previous predictions. The team also introduced learnable “Gaussian noise”, a technique that allows the AI to learn how much randomness to add to different geographic locations during training.
This method of weather forecasting was tested using weather data from 2007 to 2016 across East Asia, including 70 different weather variables. Self-reported results from this study showed that the AI models outperformed traditional weather forecasting systems by 20%. For those in the weather verification business and number nerds like us, let’s break that number down. The highest-performing traditional weather AI model, FourCastNet, achieved a score of 0.2897, while the best medical imaging model, MISSFormer, reached 0.3147. The cascade prediction method achieved a score of 0.4048. When combining all elements, the final system reached a score of 0.4313, representing a nearly 20% improvement over baseline methods.
Not only does this method prove to be more accurate, but it also uses less computing power, costs less, and can produce forecasts in seconds versus hours or days. However, this technique has been limited to trials in East Asia only at this time, and has not yet been validated globally.