Google DeepMind has launched its cutting-edge GenCast AI model, designed to provide more accurate and longer-term forecasts than current weather forecasting techniques.
GenCast
- GenCast is an advanced weather forecasting system, based on machine learning and trained on historical data from 1979-2018.
- Methodology: Uses diffusion models, which are based on techniques similar to AI image generation.
Ensemble Forecasting
- Starts with random noise.
- Refines it through neural networks.
- Combines multiple forecasts to give a best estimate and estimate uncertainty.
Features of GenCast
Accuracy and Long-Term Forecasting
- Uses machine learning techniques to provide more accurate and longer-term weather forecasts than traditional models.
- AI-based ensemble forecasting: Unlike traditional numerical weather prediction (NWP) models, GenCast uses an ensemble of AI-generated forecasts trained on 40 years of reanalysis data.
Performance
- More accurate than traditional numerical weather prediction systems.
- Outperforms the system of the European Centre for Medium-Range Weather Forecasts (ECMWF).
- Produces forecasts of atmospheric variables such as temperature, pressure, humidity, and wind speed.
- Provides data at the surface and 13 different heights.
Capacity
- Produces up to 15 days of forecasts in just 8 minutes on a Tensor Processor Unit (TPU).
- Much faster than traditional circulation models.
- Model training completed in 5 days on 32 TPUs.
How GenCast works
Data training
- Trained on 40 years of reanalysis data from 1979-2019.
- Blends historical data and modern forecasts.
Technical structure
- Neural network consists of 41,162 nodes and 240,000 edges.
- Nodes process data and edges connect them.
- Diffusion model: Refines noisy data in 30 steps to improve accuracy.
Ensemble forecasting
- Produces about 50 forecasts at a time.
- Provides probabilistic forecasts (e.g., probability of rainfall), not exact quantities.
Capacity and speed
- Produces forecasts in 8 minutes using a TPU v5 unit.
- Compared to traditional NWP models, which take several hours, it is much faster.
Existing forecasting models
Numerical weather forecasting
- Based on solving physical equations.
- Requires high computational power.
- Provides only deterministic forecasts.
- Huawei’s Pangu-Weather: Produces weekly weather forecasts faster than NWP models.
- Significance: Performs better than current leading systems. Forecasts a wide range of weather scenarios. Provides a more comprehensive and accurate picture of upcoming conditions.