MetNet-3, jointly developed by Google and DeepMind, is a model that accurately predicts core weather variables (such as precipitation, surface temperature, wind speed and direction, and dew point) for the next 24 hours. Accuracy surpasses current state-of-the-art physics-based weather prediction models.
MetNet-3 has a spatial resolution of only 1 to 4 kilometers and can refresh the forecast data every 2 minutes.
Key features of MetNet-3:
1. High-resolution forecasting: MetNet-3 can provide high-resolution weather predictions with time intervals of 2 minutes and spatial resolution of 1 to 4 kilometers.
This means that the MetNet-3 model provides a new weather forecast for a specific location every 2 minutes. For example, if you want to know how the weather will change in the next hour, MetNet-3 will provide you with 30 consecutive forecasts, each 2 minutes apart.
2. Surpassing traditional models: MetNet-3 surpasses the most advanced physics-based weather prediction models in predicting core weather variables for the next 24 hours, such as precipitation and surface temperature.
3. Use of direct observation data: MetNet-3 uses direct observation data for training and evaluation, which often has higher fidelity and resolution.
Unlike traditional weather prediction models, MetNet-3 directly uses observational data from the atmosphere for training and evaluation, which may come from ground-based weather stations, satellites, etc.
4. Data Intensification Technology: A key innovation of MetNet-3 is data intensification technology, which combines traditional data assimilation and simulation processes into a single process. This means that the model can directly use observational data to make predictions without data assimilation and simulation first.
5. Learn from sparse observations: It can learn and predict from sparse weather observation data. “Sparse observations” mean that the observation data is not continuous in space or time, and there may be many gaps or missing places. Despite this, this AI model is still able to effectively learn from this incomplete data and provide accurate predictions for future weather conditions.
In short, it is an advanced AI weather prediction model that can learn from incomplete or sparse weather data.
MetNet-3 has been integrated into multiple Google products and technologies, and is currently operational in 48 states in the U.S. and Europe, providing real-time 12-hour precipitation forecasts for Google’s various products and technologies, such as search, and transforming the model’s rich outputs into actionable information to serve millions of users.
Details: https://blog.research.google/2023/11/metnet-3-state-of-art-neural-weather.html
Paper: https://arxiv.org/abs/2306.06079