Learning skillful medium-range global weather forecasting

The problem

Forecasts for the coming week or so are produced by numerical weather prediction: the atmosphere is divided into a grid, the equations describing its motion and thermodynamics are stepped forward on high-performance computers, and the result is a forecast. Building and running such a system is a long engineering effort, and every forecast costs a substantial amount of computing. GraphCast asks a different question: instead of solving the physical equations, can a model learn from recorded weather how the atmosphere evolves, and do it well enough to compete with the operational systems used in practice?

What the model does

GraphCast is a learned global weather model built on graph neural networks. Rather than treating the atmosphere as a regular grid on which equations are integrated, it represents the atmospheric state as a graph of points and learns from data how that state develops over time. According to the paper's abstract, the model predicts hundreds of weather variables over 10 days at 0.25 degree resolution globally in under one minute ^[1].

The model was trained on four decades of weather reanalysis data from the ECMWF's ERA5 dataset ^[2]. Reanalysis is a record of the past atmosphere produced by running a fixed forecast model together with a data assimilation system over historical observations; because the same system is used throughout, it gives a consistent, gridded history suitable for learning from.

Results

The paper states that GraphCast "significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets" ^[1]. Those targets are the individual variables, regions and forecast lead times on which forecasts are scored, so the claim covers a broad range of conditions rather than a single headline number.

The accompanying blog post puts the comparison in more concrete terms: GraphCast predicts weather conditions up to 10 days in advance more accurately and much faster than HRES, the ECMWF system ^[2]. On speed, it states that making 10-day forecasts with GraphCast takes less than a minute on a single Google TPU v4 machine ^[2], in contrast to the computing resources operational numerical weather prediction consumes.

The blog post also gives a specific case: the model predicted about nine days in advance that Hurricane Lee would make landfall in Nova Scotia ^[2].

Authorship and publication

The paper was first submitted to arXiv on 24 December 2022 as arXiv:2212.12794, and it has 18 authors, starting with Remi Lam ^[1]. A DeepMind blog post dated 14 November 2023 announced publication in the journal Science and gives the paper's DOI as 10.1126/science.adi2336 ^[2]. The blog post lists co-authors including Alvaro Sanchez-Gonzalez, Oriol Vinyals, Shakir Mohamed and Peter Battaglia ^[2].

Why it matters

The comparison that matters here is with deterministic operational systems — the conventional physics-based forecasts that produce a single best estimate for each variable, as opposed to ensembles that sample a range of possible outcomes. Matching and beating such systems on the majority of verification targets, while producing a 10-day global forecast in under a minute, shows that a learned model trained on reanalysis data can be a practical alternative to a purely physics-based pipeline for medium-range forecasting. It also changes the cost profile of forecasting: forecasts that require a large high-performance computing run can be generated on a single accelerator.

The approach has clear boundaries. A model of this kind learns from the record it is trained on, so its skill depends on the quality and consistency of that reanalysis dataset, and its behaviour in conditions poorly represented in that record is not guaranteed by the training procedure. The reported comparison is against deterministic systems; it is not a claim about ensemble prediction. What the paper establishes is narrower and more concrete: a learned model can produce global medium-range forecasts that are both faster and, on most of the measures the authors checked, more accurate than the operational deterministic baseline.