Artificial Intelligence (AI) is making strides in weather forecasting, promising more accurate, faster, and reliable predictions. A new AI model developed by Deepmind, a subsidiary of Google, is showing potential in enhancing weather forecasts. According to a study, this model could provide better predictions, especially for extreme weather events, the path of tropical cyclones, and wind strength development.
An expert from the German Weather Service (DWD), Roland Potthast, emphasizes that while AI models hold significant potential, they should be seen as complementary to traditional, physics-based models rather than replacements. Potthast, who leads Numerical Weather Prediction at DWD, describes the study published in the journal Nature as a “significant step” forward. He believes the approaches by Google and other tech companies can complement, inspire, and advance weather services, ultimately providing the public with improved forecasts and warnings.
The AI-based weather forecasting method, named Gencast, was developed by a team led by Ilan Price from Deepmind in London. The study was conducted solely by Deepmind employees and reviewed by independent experts. The findings suggest that Gencast surpasses the best conventional medium-range weather forecasts. The AI was trained using analysis data from weather events over 40 years (1979 to 2018) and tested on its ability to predict weather for 2019.
Gencast is capable of producing global 15-day forecasts within eight minutes. Traditionally, the European Centre for Medium-Range Weather Forecasts (ECMWF) has been regarded as the most accurate for such predictions. However, Gencast outperformed in over 97% of cases when predicting wind speeds, temperatures, and other atmospheric features, according to the developers. Gencast generates predictions multiple times per forecast, increasing the likelihood of accuracy. The system promises higher precision, efficiency, and accessibility across a wide range of scenarios.
The DWD is currently testing its own AI model, and more are in development to complement existing methods. Potthast explains that combining physics-based models and AI models in the DWD’s forecasting chain aims to provide the best predictions for various time scales and forecast variables, such as precipitation, temperature, winds, pressure, humidity, gusts, and more.
Despite the advancements, Potthast clarifies that AI does not render human input obsolete. In fact, more human effort is required to maintain the reliability of traditional systems. AI models cannot yet match the quality, breadth, diversity, and reliability of physics-based systems but show faster or better results in certain variables or scores. There is, however, a steep learning curve in this field.
AI models differ from traditional weather models as they do not inherently adhere to the laws of nature. Weather results from interconnected processes, and physics-based models align with these natural laws, making their predictions consistent and understandable. In contrast, machine learning models focus on predicting individual values accurately without directly considering these laws. This can lead to forecasts that appear accurate but may not be entirely correct, especially when the weather becomes complex. Physics-based models are designed to maintain these connections from the outset, providing more reliable predictions.