New Google AI weather model beats most reliable forecast system

• Previous AI models only resulted in a take-it-or-leave-it projection (also known as a deterministic model), rather than a range of outcomes.

Driving the news: Google DeepMind researchers told reporters on Monday that the model shows skill in anticipating extreme events outside the bounds of what occurred during the training period.

• That indicates it may accurately predict unprecedented events that climate change makes more likely and severe.

The big picture: AI-based weather forecasting is about to be further integrated into government weather and climate agencies along with private companies, experts told Axios.

• However, it won't replace human forecasters or negate the need for people with scientific expertise.
• University of Oklahoma meteorologist Aaron Hill, who wasn't part of the new research, said human forecasters have "unparalleled" abilities "to parse through complex weather prediction model output and observations."

Models like the new "GenCast" are also unlikely to fully displace the current generation of weather models.

• Rather, they will be used as another forecasting tool in meteorologists' toolbox when predicting day-to-day weather, including extreme events.
• "Their experience looking at forecast output and observations daily gives them the upper hand on forecast models: they know when they are good, and they know when they are bad, and they can use those things to their advantage to make improved forecasts," Hill told Axios via email.

Zoom in: The new model is detailed in a study published in the journal Nature on Wednesday.

• It finds the model has greater skill than the extremely reliable ensemble run by the European Center for Medium-Range Weather Forecasts (ECMWF) on 97.2% of 1,320 forecast metrics they evaluated.
• This includes predicting extreme weather events, the tracks of tropical cyclones and wind power output.
• GenCast's machine learning approach used weather data from 1979 to 2018 for training purposes.

The intrigue: Today, top forecast centers such as ECMWF in the U.K. and NOAA prepare their forecasts using ensemble methods rooted mainly in physics-based models.

• Such systems use mathematical equations that describe how the atmosphere functions, with calculations for everything from how moisture is transported to how the jet stream's winds can fuel and steer storm systems.
• A massive haul of observational weather data gets swept into these models to produce projections for hours to days to come.

• This is computationally heavy, requiring supercomputers running for hours to accomplish each model run. Forecast improvements have been iterative in recent years, rather than revolutionary.

The new-generation of AI weather models that are rolling out from tech companies are trained on past weather data and rely on machine learning techniques.

• They can be run on cloud processing systems in just minutes. GenCast, for example, takes just 8 minutes to generate an ensemble forecast.

Friction point: Some outside meteorologists told Axios that GenCast's output and skill have gaps.

• They contend it doesn't generate enough insights to be an off-the-shelf, complete forecast model.
• For example, currently it shows projections for every 12 hours, out to 15 days, missing potentially crucial and impactful weather in between the time steps.

Zoom out: GenCast is one of a slew of AI-driven weather models unveiled in recent years from companies such as Nvidia and Microsoft, each of which has shown promise in improving forecast accuracy.

What they're saying: "It is exciting that machine-learning will change the game when it comes to probabilistic forecasting," Marshall Shepherd, a research meteorologist at the University of Georgia who didn't take part in the new study, told Axios.

• As we embraced ensemble forecasting over deterministic approaches, this is just an inevitable next step in the process."