Innovative GenCast Model Unveiled by DeepMind for Advanced Weather Prediction Using Diffusion Technique – Achieves 15-Day Forecast

DeepMind has unveiled a new weather forecasting model called GenCast, developed from the GraphCast model from 2023. GenCast is a generative model that utilizes diffusion algorithms familiar in image, audio, and video generation models. It can forecast detailed weather conditions (at a resolution of 0.25° latitude and longitude, approximately 28×28 kilometers) up to 15 days in advance (compared to GraphCast’s 10 days) and boasts greater accuracy than the current best model, the ECMWF’s ENS model.

Unlike GraphCast’s deterministic forecasting approach, which seeks a single possible answer for future weather conditions, GenCast aims to explore over 50 possible scenarios. Its fundamental concept involves viewing the world as a spherical space and using diffusion techniques to generate various possible patterns. Training data for GenCast is derived from 40 years of ECMWF historical weather data, with DeepMind using data up to the year 2018 to test forecasting against weather conditions from 2019 onwards. The results are impressive, with GenCast outperforming the ENS model by 97.2% in all tests and achieving a 99.8% success rate for forecasts beyond 36 days.

Running the GenCast model to predict weather conditions 15 days in advance takes only 8 minutes of processing time on Google Cloud TPU v5, a significant reduction from the ENS model’s requirement for large-scale supercomputers and hours of processing time. An example of GenCast’s forecasting success is highlighted during the prediction of Typhoon Hagibis in Japan in 2019, where the GenCast prediction (blue line) closely aligns with the actual path of the typhoon (red line).

Published in the prestigious scientific journal Nature and made available as an open-source model by Google, GenCast can be accessed on GitHub for download. Future plans include releasing past and real-time forecasting data to empower weather forecasting communities with increasingly accurate models over time.

TLDR: DeepMind introduces the GenCast weather forecasting model, surpassing current models in accuracy and forecasting capabilities, providing a more nuanced view of future weather patterns.