GenCast - Google DeepMind's AI weather forecasting model

What is GenCast?

GenCast is a revolutionary AI weather forecasting model from DeepMind, based on diffusion modeling technology, providing global weather forecasts up to 15 days in advance. GenCast outperforms the world's leading medium-range weather forecasting system, ENS, in 97.2% of forecasting tasks, with particularly outstanding performance in predicting extreme weather events. Compared to traditional models, GenCast can generate forecasts in 8 minutes, significantly improving forecasting efficiency. GenCast is open source, including its code and model weights, supporting a wider range of weather forecasting communities and research.

GenCast's main functions

• Medium and long-term weather forecastsIt provides global weather forecasts for up to 15 days, updated every 12 hours.
• High-resolution predictionThe model runs at a 0.25° latitude-longitude resolution, providing high-resolution predictions.
• Ensemble predictionUnlike traditional single forecasts, GenCast generates 50 or more sets of forecasts, each representing a possible weather trajectory and expressing uncertainty.
• Extreme weather forecastGenCast excels at predicting extreme weather events, such as heat waves, strong winds, and tropical cyclones.
• Quick predictionOn Google Cloud TPU v5, GenCast can generate a 15-day weather forecast collection in just 8 minutes.

GenCast's technical principles

• diffusion modelBased on a diffusion model, this is a generative AI model used to generate new samples in fields such as images, videos, and music. The model generates samples of future states from noise through an iterative refinement process.
• Earth's spherical geometry adaptationGenCast is particularly well-suited to the Earth's spherical geometry and can learn to accurately generate complex probability distributions of future weather scenarios.
• Conditional probability distributionThe GenCast model represents the conditional probability distribution of future weather conditions, given the current and previous weather conditions.
• Neural Network ArchitectureGenCast uses a neural network architecture that includes an encoder, processor, and decoder. The encoder maps the input from a latitude and longitude grid to an internally learned representation, the processor (graph transformer) focuses on its neighboring nodes, and the decoder maps the result back to the original grid.
• Training and PredictionGenCast is trained using 40 years of historical weather data from the ECMWF's ERA5 archive to learn global weather models and generate 15-day ensemble forecasts autoregressively.
• Parallel generationSince each time step is initialized with noise, a prediction set is generated by repeating the generation process with different noise samples.

GenCast's project address

• Project official website:gencast-predicts-weather
• GitHub repository:https://github.com/google-deepmind/graphcast
• arXiv technical paper:https://arxiv.org/pdf/2312.15796

GenCast Application Scenarios

• Disaster Prevention and ManagementGenCast can predict extreme weather events in advance, helping governments and relief organizations reduce disaster losses.
• Energy ManagementIn the fields of wind and solar power generation, GenCast provides accurate forecasts that optimize energy production predictions and improve energy efficiency.
• Agricultural PlanningGenCast helps farmers avoid risks caused by weather changes and make reasonable arrangements for agricultural production activities.
• Transportation and LogisticsGenCast provides critical weather information for aviation and maritime transport, ensuring safe and efficient transportation.
• Urban planning and constructionGenCast helps city planners consider climate change and build infrastructure that is more adaptable to weather changes.