AI model beats traditional forecasts

When Hurricane Isaias formed in early October, Google DeepMind's artificial intelligence model correctly predicted it would hit the western Florida Panhandle. Conventional models predicted it would head further west toward southeastern Louisiana.

Isaias made landfall as Category 2

Hurricane Isaias made landfall near Destin, Florida in early October as a Category two storm. The Google DeepMind model's accurate prediction stood out against traditional physics-based forecasting approaches.

Gold standard for hurricane forecasts

Michael Lowry, a longtime meteorologist and hurricane expert in Miami who has advised the Federal Emergency Management Agency and the National Hurricane Center, calls Google DeepMind the gold standard model.

“That model has been, by far and away, the best performing computer model that we've ever seen. Period.”

NHC adopted the model last year

The National Hurricane Center first started incorporating the Google DeepMind model into its official forecasts last year.

Unusually fast improvement for new model

New hurricane models typically take years to refine after introduction.Brian McNoldy, a hurricane researcher for the University of Miami, says the Google DeepMind model came out of the gate in first place, an impressive feat.

“It's something we're not used to, to have a model, the new kid on the block, very quickly becoming the best model out there.”

How traditional models work

Traditional physics-based hurricane models use real-time atmospheric data like temperature, humidity, pressure and wind shear measured by satellites and aircraft. They use complicated equations to predict storm path and strength.

AI models work differently

AI models rely on history rather than physics equations. They identify patterns by analyzing thousands of past storms along with atmospheric data. They don't lean on the governing equations of the atmosphere to make predictions.

The black box problem

Most people designing these AI models are software engineers developing neural networks, not meteorologists. The models can seem like a black box: forecasters don't know why they get predictions right or wrong.

“We don't know why they might get a forecast right. We also don't know why they might get a forecast wrong.”

Two proven AI models emerge

Not all AI weather forecasting models have performed well. The two most accurate so far are Google DeepMind and the European Centre for Medium-Range Weather Forecasts, a U.K.-based organisation, despite some pitfalls.

NHC blends multiple models

Forecasters at the National Hurricane Center use several proven models, both AI and traditional, when putting together official forecasts. Humans interpret the results before deciding what to tell the public.

“Every storm has different forecast challenges. That's why we take that consensus approach and use that blend of the best models.”

AI improving faster than traditional models

AI models update in months or weeks, unlike traditional models updated annually at best. Michael Lowry says AI is making advancements in hurricane prediction skill over a year that would have taken a decade or…

“They are making advancements in skill and hurricane predictions over the course of a year that would have taken us ten years, a decade or more, to make.”