How does Google’s GenCast AI, which predicts the weather, work?

Sub: Sci

Sec: Awareness in IT

Context:

Google DeepMind introduced GenCast, an advanced AI model designed to improve weather forecasting. According to a paper published in the journal Nature, GenCast outperforms existing models in predicting the weather, offering more accurate forecasts, especially for extreme weather events.

How Does the Weather Forecasting Work?

Numerical Weather Prediction (NWP):
- Weather forecasts are traditionally made by running multiple simulations of the atmosphere using supercomputers.
- These simulations are based on equations that model the physical laws of nature, but they can only predict the weather accurately for about a week.
- Ensemble forecasting was introduced in the 1990s, where several predictions are made using different starting conditions to account for uncertainty.

How GenCast Performs?

AI-Driven Ensemble Forecasting:
- Unlike traditional models, GenCast generates forecasts using AI instead of NWP simulations.
- Trained on 40 years of reanalysis data (1979–2019), GenCast provides predictions with greater accuracy than the ECMWF’s ensemble (ENS) model, which is regarded as one of the best in NWP.
- Key Performance Metrics:
  - Outperforms ENS on 97.2% of 1,320 targets.
  - More accurate than ENS on 99.8% of the targets when predicting weather more than 36 hours in advance.
  - Better predictions for extreme weather, tropical cyclones, and wind power production.

How GenCast Works?

AI Model Structure:
- GenCast uses a neural network with 41,162 nodes and 240,000 edges to process data.
- The model “de-noises” input data through a series of 30 refinements to generate accurate weather predictions.
- This process is similar to how diffusion models work in AI (e.g., generating images or videos).
- The model generates 50+ forecasts simultaneously, each covering 15 days with a 12-hour temporal resolution and spatial resolution of 0.25° x 0.25°.
Speed:
- GenCast can produce forecasts in 8 minutes using a single TPU v5 unit, a significant improvement over the hours required by supercomputers for traditional NWP.

Will GenCast Replace NWP?

Probabilistic Forecasts:
- Unlike deterministic NWP models, GenCast provides probabilistic forecasts, such as the chance of rain rather than exact measurements.
- Experts believe probabilistic forecasts are more useful for predicting extreme weather events, offering better preparation time.
Complementary, Not Replacing NWP:
- While GenCast’s performance shows promise, traditional NWP models remain essential. They provide critical initial weather data and training data for AI models like GenCast.
- Both NWP and GenCast rely on fundamental weather data governed by the laws of physics.

Other AI Models in Development:

DeepMind is also working on GraphCast for medium-range deterministic forecasts.
Google Research is developing NeuralGCM to combine AI with NWP for deterministic predictions.
Other companies, like Huawei and Nvidia, are developing AI models that outperform traditional NWP models in speed and extreme weather prediction.

Source: TH