New AI Model Beats Traditional Weather Forecasts

Recently, a groundbreaking machine learning weather forecast model named GenCast has shown it can outperform traditional systems in certain situations. Developed by researchers at Google DeepMind, this model uses a diffusion approach similar to AI image generators to create multiple forecasts quickly and efficiently. Traditional weather forecasts work by running multiple simulations of the atmosphere, starting from slightly different weather estimates. These simulations use a grid-based model of the atmosphere and require substantial computing power. However, they don't account for the exact current weather everywhere due to lack of global sensors. GenCast, on the other hand, employs machine learning to generate an ensemble of forecasts that better mimic the complexity of natural weather systems. By averaging these forecasts, it produces more accurate predictions than the best numerical weather prediction system in the world.The model is trained on reanalysis data spanning 1979 to 2018, which is produced by general circulation models corrected to match historical weather observations. GenCast predicts various weather variables like temperature, pressure, humidity, and wind speed across a global grid. It uses a neural network to interpret random noise and create forecasts, enabling it to run predictions up to 15 days in just 8 minutes on a single processor. However, while machine learning forecasts could become more common, they still rely on traditional models for initial conditions and fine-tuning. Additionally, these models aren't suitable for long-term climate projections, which require different considerations like changes in oceans and sea ice, future carbon emissions, and statistical climate trends. The future of weather forecasting and climate projections will likely involve a blend of machine learning and fundamental physics. While machine learning can handle vast amounts of weather data, physical principles are crucial for understanding slower climate phenomena.