Introducing MetNet-3: A New Weather Model for Accurate and High-Resolution Forecasts
Forecasting weather variables such as precipitation, temperature, and wind is crucial for various aspects of society. Accurate forecasts help in daily planning, transportation, energy production, and preparing for extreme weather events like floods, droughts, and heat waves. Today, we are excited to present MetNet-3, a new weather model developed by Google Research and Google DeepMind.
MetNet-3 builds upon the earlier models, MetNet and MetNet-2, and provides high-resolution predictions up to 24 hours ahead for a larger set of core variables. These include precipitation, surface temperature, wind speed and direction, and dew point. With lead time intervals of 2 minutes and spatial resolutions of 1 to 4 kilometers, MetNet-3 creates a temporally smooth and highly granular forecast.
In terms of performance, MetNet-3 outperforms traditional methods such as the High-Resolution Rapid Refresh (HRRR) and ensemble forecast suite (ENS) models for multiple regions up to 24 hours ahead.
MetNet-3’s capabilities have been integrated across various Google products and technologies to provide accurate and reliable weather information to people in multiple countries and languages. Currently, it is available in the contiguous United States and parts of Europe, with a focus on 12-hour precipitation forecasts.
Key Features of MetNet-3
Densification of Sparse Observations: Unlike many machine learning weather models that rely on atmospheric state generated by traditional methods, MetNet-3 uses direct observations of the atmosphere for training and evaluation. This approach offers higher fidelity and resolution. MetNet-3 includes point measurements from weather stations as inputs and targets, with the goal of making a forecast at all locations. The key innovation is a technique called densification, which merges data assimilation and simulation into a single pass through the neural network. This technique significantly improves forecast quality in areas with sparse coverage.
High Resolution in Space and Time: MetNet-3 utilizes high-resolution direct observations from weather stations and ground radar stations. Lead time conditioning allows the model to efficiently model the high temporal frequency of the observations, resulting in a fully dense 24-hour forecast with a temporal resolution of 2 minutes. MetNet-3 predicts a marginal multinomial probability distribution for each output variable and location, providing rich information beyond just the mean value.
Performance Comparison and Real-Time Delivery
MetNet-3’s forecasts have been compared with advanced probabilistic ensemble NWP models, and it consistently outperforms them in terms of accuracy. In addition, a real-time system has been developed to deliver MetNet-3’s precipitation forecasts every few minutes for the entire contiguous United States and 27 countries in Europe, with a lead time of up to 12 hours.
With MetNet-3, we aim to provide users with accurate and high-resolution weather forecasts to help them make informed decisions and stay safe. We will continue to enhance and expand the capabilities of MetNet-3 to bring the benefits of AI-powered weather forecasting to more regions around the world.
