DeepMind and Google Introduce GraphCast: A Fast and Scalable Machine Learning Weather Simulator

People consider the weather forecast in all aspects of their lives, from choosing an outfit to what to do in the event of a hurricane. Forecasts over a period of generally three to seven days are called medium-range forecasts. Several sectors, such as agriculture, construction, travel, etc., rely on “medium-range” weather forecasts to make decisions, which are offered up to four times a day by meteorological offices such as the European Center for Medium-Range Weather Forecasts (ECMWF).

Medium-range weather forecasts have two main parts, both of which are simulated using massive high-performance computing (HPC) clusters. The first part is “data assimilation”, which is the method of forecasting weather conditions by analyzing current and historical data collected by satellites, weather stations, ships, etc. The second is a model that predicts the evolution of time-related variables over time. ; these models are usually built using numerical weather prediction (NWP).

However, traditional forecasting models based on numerical weather prediction, which rely on computational clusters to run simulations, cannot scale effectively due to the ever-increasing amount of weather data. Their accuracy depends on the time-consuming and resource-intensive contribution of human specialists.

The new study from DeepMind and Google showcases GraphCast, a machine learning (ML)-based weather simulator that adapts well to data and can create a 10-day prediction in less than 60 seconds. Compared to state-of-the-art ML-based benchmarks and the world’s most accurate deterministic operational medium-range weather forecasting system, GraphCast comes out on top.

As mentioned in their article “GraphCast: Learning Skillful Medium-Range Global Weather Forecasting”, GraphCast uses graph neural networks (GNN) in an “encode-process-decode” arrangement to create an autoregressive model. According to the researchers, learning the complex physics of fluids and other materials is a perfect fit for GNN-based designs. Additionally, input graph structures can be used to simulate any spatial interaction model, because input graph structures determine the interactions between parts of a representation. The team takes advantage of this GNN capability by developing a new internal multi-mesh representation technique, which enables long-range interactions with minimal message-passing overhead.

The three-step simulation process in GraphCast is as follows:

1. GNN with edges directed from the grid points to the multi-mesh is used to map the original latitude-longitude grid input data into the learned features on the multi-mesh
2. A deep GNN is used to perform learned message passing over the multi-mesh, where the long-range edges allow information to be propagated efficiently through space.
3. The decoder maps the final multi-mesh representation onto the latitude-longitude grid and does whatever is necessary.

The team tested GraphCast on a single Cloud TPU v4 device. Their findings show that GraphCast can produce a 10-day forecast with 0.25° resolution in less than 60 seconds. GraphCast’s performance outperforms the European Center for Medium-Range Weather Forecasts (HRES) High-Resolution NWP-based deterministic operational forecasting system on 90% of 2,760 variables. It also outperforms the most accurate ML-based weather forecast model available on 99.2% of 252 targets.

This study advances the use of ML-based simulations in other areas of physical science. The team believe their work will open up new possibilities for fast and accurate weather forecasting.

Check Paper. All credit for this research goes to the researchers on this project. Also don’t forget to register. our Reddit page and discord channelwhere we share the latest AI research news, cool AI projects, and more.