🔗 Share this article

When Developing Cyclone Melissa was churning south of Haiti, weather expert Philippe Papin felt certain it would soon escalate to a monster hurricane.

Serving as lead forecaster on duty, he forecasted that in just 24 hours the weather system would intensify into a severe hurricane and start shifting in the direction of the Jamaican shoreline. Not a single expert had ever issued this confident forecast for quick intensification.

However, Papin had an ace up his sleeve: AI technology in the form of Google’s new DeepMind cyclone prediction system – released for the first time in June. And, as predicted, Melissa evolved into a storm of astonishing strength that ravaged Jamaica.

Growing Dependence on AI Predictions

Meteorologists are heavily relying upon the AI system. On the morning of 25 October, Papin explained in his public discussion that the AI tool was a key factor for his confidence: “Roughly 40/50 Google DeepMind simulation runs show Melissa becoming a most intense hurricane. While I am unprepared to forecast that intensity yet due to path variability, that is still plausible.

“It appears likely that a phase of rapid intensification is expected as the storm drifts over very warm ocean waters which represent the most extreme oceanic heat content in the entire Atlantic basin.”

Outperforming Traditional Models

Google DeepMind is the pioneer AI model focused on hurricanes, and now the initial to outperform traditional weather forecasters at their specialty. Across all 13 Atlantic storms so far this year, the AI is top-performing – surpassing experts on path forecasts.

The hurricane eventually made landfall in Jamaica at category 5 strength, one of the strongest coastal impacts ever documented in nearly two centuries of data collection across the region. Papin’s bold forecast probably provided people in Jamaica additional preparation time to get ready for the catastrophe, potentially preserving people and assets.

The Way Google’s Model Works

The AI system works by spotting patterns that traditional lengthy scientific prediction systems may miss.

“They do it much more quickly than their traditional counterparts, and the computing power is more affordable and demanding,” said Michael Lowry, a former forecaster.

“This season’s events has proven in quick time is that the newcomer artificial intelligence systems are competitive with and, in some cases, more accurate than the less rapid physics-based weather models we’ve relied upon,” he said.

Understanding Machine Learning

It’s important to note, Google DeepMind is an example of AI training – a technique that has been employed in research fields like weather science for years – and is distinct from creative artificial intelligence like ChatGPT.

AI training processes mounds of data and pulls out patterns from them in a such a way that its system only takes a few minutes to generate an result, and can do so on a standard PC – in sharp difference to the flagship models that authorities have utilized for years that can require many hours to run and need some of the biggest high-performance systems in the world.

Professional Reactions and Future Advances

Still, the fact that Google’s model could exceed earlier gold-standard legacy models so rapidly is truly remarkable to weather scientists who have dedicated their lives trying to forecast the most intense weather systems.

“I’m impressed,” commented James Franklin, a former expert. “The data is now large enough that it’s pretty clear this is not a case of chance.”

Franklin noted that although Google DeepMind is outperforming all other models on forecasting the trajectory of storms worldwide this year, similar to other systems it sometimes errs on high-end intensity predictions wrong. It had difficulty with another storm earlier this year, as it was also undergoing rapid intensification to category 5 above the Caribbean.

In the coming offseason, he said he intends to talk with Google about how it can make the DeepMind output even more helpful for experts by offering additional internal information they can use to assess exactly why it is producing its answers.

“A key concern that nags at me is that while these predictions seem to be highly accurate, the output of the model is essentially a black box,” said Franklin.

Broader Industry Developments

There has never been a private, for-profit company that has developed a high-performance weather model which allows researchers a peek into its techniques – unlike nearly all other models which are offered at no cost to the general audience in their entirety by the governments that created and operate them.

Google is not alone in adopting AI to address challenging meteorological problems. The authorities also have their own AI weather models in the development phase – which have also shown improved skill over previous non-AI versions.

The next steps in AI weather forecasts seem to be startup companies taking swings at previously difficult problems such as long-range forecasts and improved early alerts of severe weather and flash flooding – and they are receiving US government funding to do so. One company, WindBorne Systems, is even launching its proprietary weather balloons to fill the gaps in the US weather-observing network.