Why in news
Google DeepMind has introduced GenCast, a new machine-learning weather prediction model, which outperforms traditional forecasting systems in some situations.
This development is noted for its efficiency in generating forecasts with fewer resources and faster processing.
What is GenCast
GenCast is a machine-learning weather prediction model that uses a "diffusion model" approach, similar to AI image generators.
It generates multiple forecasts to capture complex atmospheric behaviors, producing more accurate predictions by averaging these forecasts.
It can predict the weather for 15 days in just 8 minutes.
How weather forecasts work
Weather forecasts are made by running simulations of the atmosphere based on different estimates of the current weather.
These simulations use a grid model of the atmosphere and solve equations about physical laws, producing forecasts for future conditions.
The process requires significant computational power and is usually done on high-performance supercomputers.
Machine-learning the weather
Recent efforts aim to use machine learning for weather prediction, often employing neural networks to learn from historical data.
These models traditionally produce smoother forecasts over time, which doesn’t always match the dynamic nature of real weather systems.
GenCast addresses this by generating multiple forecasts to reduce the smoothing effect and provide more detailed predictions.
Generative AI for weather
GenCast is trained using reanalysis data from past weather, corrected to match real historical observations.
It uses random noise to generate forecasts, with the model working much faster than traditional methods—forecasting up to 15 days in just 8 minutes on a single processor.
This is more efficient than traditional models, and machine learning-based forecasts could become more widespread.
What about the climate?
Machine learning weather forecasting systems are not suitable for long-term climate projections due to differences in data requirements and timescales.
Climate projections focus on long-term changes, like carbon emissions, and have less available data.
To address this, techniques like physics-informed neural networks and simplified models are being explored to use machine learning effectively for climate predictions
COMMENTS