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

People account for the forecasted weather in every aspect of their lives, from choosing an outfit to what to do in the event of a hurricane. Forecasting over a time frame that is typically three to seven days out is referred to as medium-range forecasting. Several sectors, like agriculture, construction, travel, etc., rely on “medium-range” weather forecasts for making decisions, which are offered up to four times daily by weather bureaus like the European Centre for Medium-Range Weather Forecasts (ECMWF).

There are two primary parts to medium-range weather forecasts, both 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 forecasts how weather-related variables will change over time; these models are typically built using numerical weather prediction (NWP).

However, traditional NWP-based forecasting models, which rely on computational clusters to execute simulations, can not scale efficiently due to the ever-increasing quantity of weather data. Their accuracy is dependent on the time-consuming and resource-intensive input of human specialists.

The new study by DeepMind and Google presents GraphCast, a machine-learning (ML) based weather simulator that scales well with data and can create a 10-day prediction in under 60 seconds. When compared to state-of-the-art ML-based benchmarks and the most accurate deterministic operational medium-range weather forecasting system in the world, GraphCast comes out on top. 

As mentioned in their paper “GraphCast: Learning Skillful Medium-Range Global Weather Forecasting,” GraphCast uses graph neural networks (GNNs) in an “encode-process-decode” arrangement to create an autoregressive model. According to the researchers, learning the intricate physics of fluids and other materials is ideally suited for GNN-based designs. In addition, the input graph structures can be used to simulate any spatial interaction pattern, as the input graph structures determine the interactions between portions of a representation. The team takes advantage of this GNN capability by developing a novel internal multi-mesh representation technique, which allows for long-range interactions with minimal message-passing overhead.

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

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

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

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

Check out the Paper. All Credit For This Research Goes To Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.