Google DeepMind touts AI model for 'better' global weather forecasting

Google DeepMind researchers claim they’ve used machine learning to devise a model that can deliver better 15-day weather forecasts and requires only modest quantities of compute resources to produce its predictions.

The model, dubbed GenCast, provides a probabilistic ensemble weather forecast – a distribution of probable weather – rather than a deterministic numerical weather prediction (NWP) calculation that points to a specific outcome – at risk of error.

One of the paper’s co-authors, DeepMind researcher Ilan Price, told The Register: "GenCast is a machine learning-based weather model, which learns directly from historical weather data. This is in contrast to traditional models, which make forecasts by solving physics equations.”

"One limitation of these traditional models is that the equations they solve are only approximations of the atmospheric dynamics. GenCast is not limited to learning dynamics/patterns that are known exactly and can be written down in an equation.

It has the opportunity to learn more complex relationships and dynamics directly from the data

"Instead it has the opportunity to learn more complex relationships and dynamics directly from the data, and this allows GenCast to outperform traditional models."

In a DeepMind write-up, Price and co-author Matthew Wilson rate the model “a critical advance in AI-based weather prediction that builds on our previous weather model [GraphCast], which was deterministic, and provided a single, best estimate of future weather. By contrast, a GenCast forecast comprises an ensemble of 50 or more predictions, each representing a possible weather trajectory."

There's more to this than Googlers boasting about their work on the company blog. GenCast is also the subject of a a paper published in the journal Nature that argues the model "significantly outperforms the top operational ensemble NWP model, ENS."

The European Centre for Medium-Range Weather Forecasts (ECMWF) produces ENS, an ensemble of 51 NWP-based forecasts that together represent a range of possibilities and is relied upon by 23 member states and 12 co-operating states in Europe.

According to the DeepMind paper, "GenCast … generates global 15-day ensemble forecasts at 0.25° resolution, which are more accurate than the top operational ensemble system, ENS of ECMWF."

The paper claims GenCast beat the ENS on 97.2 percent of the evaluated targets. It also does better than ENS in predicting the path of tropical cyclones, according to the researchers.

• AWS says AI could disrupt everything – and hopes it will do just that to Windows
• Amazon promises 4x faster AI silicon in 2025, turns Trainium2 loose on the net
• NASA's Solar Dynamics Observatory datacenter flooded, offline until 2025
• IBM sued again in storm over Weather Channel data sharing

However GenCast gets the job done, better forecasts are welcome as weather events have significant socio-economic consequences. According to a report last month commissioned by the International Chamber of Commerce, "climate-related extreme weather events have cost the global economy more than $2 trillion over the past decade."

The DeepMind researchers’ work therefore has the potential to help mitigate the consequences of severe weather by assisting people and businesses better prepare for adverse conditions. It may also help with renewable energy planning, through improved wind-power forecasting, for example.

And it appears GenCast does so quite affordably. NWP compute costs can be significant – potentially hundreds of thousands of dollars annually in compute and data egress charges, depending on the scenario. ENS ensemble forecasts at a resolution of 0.2° or 0.1°, "take hours on a supercomputer with tens of thousands of processors," observe Price and Wilson.

"It takes a single Google Cloud TPU v5 just 8 minutes to produce one 15-day forecast in GenCast’s ensemble, and every forecast in the ensemble can be generated simultaneously, in parallel," they say.

Google Cloud bills Cloud TPUv5 instances at around $1-3 per chip-hour.

Asked to provide a more specific figure, Price told The Register, "We don't have a concrete comparison in terms of monetary cost or energy, but the amount of computation necessary differs by orders of magnitude."

Price and Wilson say that DeepMind has released the GenCast model code and weights to help accelerate research and development for the weather and climate community. And the release of real-time and historical forecasts from GenCast and prior models is planned. ®