Machine learning models have been increasingly used to forecast the weather, predicting anything from quick rain showers to century-level climate patterns. Google DeepMind has been at the forefront of these advancements, with their AI models making predictions based purely on data, without any understanding of the physics behind weather patterns.
One of DeepMind’s models, known as “nowcasting,” utilizes precipitation maps and doppler radar data to predict immediate weather conditions, such as the likelihood of rain while walking to the store. Despite lacking knowledge of weather physics, the model has proven to be accurate in low-stakes situations. Another model, MetNet-3, looks 24 hours into the future using data from a larger area, enabling predictions for more complex weather phenomena.
Furthermore, GraphCast is a pioneering model that forecasts weather conditions up to 10 days in advance with unprecedented accuracy and speed. Covering the entire planet at a high resolution, GraphCast simulates major weather patterns and is particularly useful for predicting large-scale weather events such as storms. The model’s efficiency is also noteworthy, as it can make predictions in less than a minute using a single Google compute unit.
In addition to providing quick and accurate predictions, machine learning models like GraphCast and the ClimSim project at AI2 aim to complement and improve traditional weather forecasting methods. These models look at weather data as an interconnected vector field and are significantly more cost-effective than physics-based models. While climate scientists remain skeptical, the potential of machine learning in long-term weather and climate predictions is undeniable, and the need for these tools continues to grow.