In 2023, when Storm Ciarán hit Northwest Europe, it left a trail of destruction. The storm’s intensity caught many off guard, exposing the limitations of current weather prediction models and highlighting the need for more accurate forecasting in the face of climate change.
In response to the limitations faced by advanced AI weather forecasting systems, a team of Microsoft researchers developed Aurora, an innovative AI foundational model designed to extract valuable insights from extensive atmospheric datasets.
Aurora offers a new approach to weather forecasting with the potential to revolutionise predictions and mitigate the impacts of climatic change events.
Trained on over a million hours of weather and climate simulations, Aurora boasts efficacy, enabling it to develop a comprehensive understanding of atmospheric dynamics. This allows the model to excel in various prediction tasks, even in regions with limited data or extreme weather conditions.
Operating at a spatial resolution of 0.1°, which measures to about 11 km at the equator, Aurora captures intricate details of atmospheric processes with detail, providing unprecedented accuracy in operational forecasts while significantly reducing computational costs compared to traditional systems.
Apart from accuracy and efficiency, Aurora stands out for its versatility. The model can forecast a wide array of atmospheric variables, including temperature, wind speed, air pollution levels, and concentrations of greenhouse gases. Its adaptable architecture is designed to handle diverse inputs and generate predictions at varying resolutions and levels of fidelity.
Aurora’s architecture incorporates a flexible 3D Swim Transformer with perceiver-based encoders and decoders, enabling it to process and predict a range of atmospheric variables across spatial and pressure levels.
Through pretraining on an extensive corpus of diverse data and fine-tuning for specific tasks, Aurora learns to capture intricate patterns and structures in the atmosphere, thereby excelling with limited training data during task-specific finetuning.
Meanwhile, Google has announced SEEDS, an AI technology aimed at enhancing weather forecasts through diffusion models. SEEDS also offers a decrease in computational expenses for producing ensemble forecasts and offers an enhanced depiction of rare or severe weather events.
Some of the other recent weather-related developments by Google include MetNET-3 and GraphCast.
MetNet-3 provides high-resolution forecasts up to 24 hours ahead, and GraphCast is a weather model capable of predicting conditions up to ten days in advance.
