Medium-range weather forecasts play a crucial role in agriculture, construction, travel and other industries. They also bring practical value to people’s daily lives, enabling us to plan outings and keeping us safe from extreme weather events. Traditional numerical weather prediction (NWP)-based forecasting models that run simulations on computing clusters however do not scale efficiently with today’s increasing weather data availability, and their accuracy relies on manual input from experts, which is time-consuming and cost inefficient.
In the new paper GraphCast: Learning Skillful Medium-Range Global Weather Forecasting, a research team from DeepMind and Google presents GraphCast, a machine-learning (ML)-based weather simulator that scales well with data and can generate a 10-day forecast in under 60 seconds. GraphCast outperforms the world’s most accurate deterministic operational medium-range weather forecasting system and all existing ML-based benchmarks.
The team summarizes their work’s key advances as follows:
- A novel multi-mesh GNN architecture for learned weather simulation.
- An autoregressive model that can be trained to generate forecasts on 0.25° latitude-longitude resolution and 37 levels of vertical resolution, for 40 or more steps.
- An evaluation protocol with comprehensive coverage of medium-range forecast variables.
- An ML-based forecasting model with greater skill than the best NWP-based deterministic model.
- The most accurate ML-based weather forecasting model.
GraphCast is an autoregressive model that employs graph neural networks (GNNs) in an “encode-process-decode” configuration. GNN-based architectures are well-suited for learning the complex physical dynamics of fluids and other materials. Moreover, their input graph structures can determine what parts of a representation interact with others, enabling the modelling of arbitrary patterns of spatial interactions over any range. The team exploits this GNN ability by introducing a novel internal multi-mesh representation approach to enable long-range interactions within few message-passing steps.
GraphCast’s simulation procedure comprises three steps: 1) The encoder maps input data from the original latitude-longitude grid into learned features on the multi-mesh, using a GNN with directed edges from the grid points to the multi-mesh; 2) The processor then uses a deep GNN to perform learned message-passing on the multi-mesh, where the long-range edges enable the information to be propagated efficiently across space; and 3) The decoder maps the final multi-mesh representation back to the latitude-longitude grid and combines this grid representation with the input state to output the final predictions.
The team evaluated GraphCast’s performance on a single Cloud TPU v4 device, where it was able to generate a 0.25° resolution, 10-day forecast in under 60 seconds, outperformed the European Centre for Medium-Range Weather Forecasts (ECMWF)’s high resolution (HRES) NWP-based deterministic operational forecasting system on 90 percent of the 2,760 variables, and beat the most accurate existing ML-based weather forecasting model on 99.2 percent of the 252 targets.
The researchers believe their work takes an important step forward in complementing and improving weather modelling with ML, opens new opportunities for fast and accurate weather forecasting, and could also advance the use of ML-based simulations in other areas of the physical sciences.
The paper GraphCast: Learning Skillful Medium-Range Global Weather Forecasting is on arXiv.
Author: Hecate He | Editor: Michael Sarazen
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.
Thanks for your sharing!