🔗 Share this article

As Developing Cyclone Melissa swirled south of Haiti, weather expert Philippe Papin had confidence it was about to escalate to a major tropical system.

Serving as primary meteorologist on duty, he forecasted that in a single day the weather system would intensify into a category 4 hurricane and start shifting in the direction of the coast of Jamaica. Not a single expert had ever issued such a bold forecast for rapid strengthening.

But, Papin had an ace up his sleeve: artificial intelligence in the guise of Google’s recently introduced DeepMind cyclone prediction system – launched for the initial occasion in June. True to the forecast, Melissa evolved into a storm of astonishing strength that tore through Jamaica.

Increasing Dependence on Artificial Intelligence Predictions

Meteorologists are heavily relying upon Google DeepMind. On the morning of 25 October, Papin clarified in his official briefing that the AI tool was a key factor for his confidence: “Approximately 40/50 AI simulation runs show Melissa reaching a Category 5 storm. While I am not ready to predict that intensity at this time given path variability, that is still plausible.

“There is a high probability that a phase of quick strengthening is expected as the system drifts over exceptionally hot sea temperatures which is the highest marine thermal energy in the entire Atlantic basin.”

Outperforming Traditional Models

The AI model is the pioneer AI model focused on tropical cyclones, and currently the first to outperform standard weather forecasters at their own game. Across all 13 Atlantic storms so far this year, the AI is the best – even beating experts on path forecasts.

The hurricane eventually made landfall in Jamaica at maximum intensity, one of the strongest coastal impacts ever documented in almost 200 years of record-keeping across the Atlantic basin. Papin’s bold forecast likely gave people in Jamaica extra time to prepare for the disaster, possibly saving lives and property.

The Way Google’s Model Functions

Google’s model operates through spotting patterns that conventional lengthy physics-based weather models may miss.

“The AI performs far faster than their physics-based cousins, and the computing power is more affordable and demanding,” said Michael Lowry, a ex forecaster.

“This season’s events has proven in short order is that the recent AI weather models are competitive with and, in some cases, more accurate than the slower traditional forecasting tools we’ve relied upon,” he added.

Clarifying AI Technology

It’s important to note, Google DeepMind is an instance of machine learning – a method that has been used in data-heavy sciences like meteorology for years – and is distinct from generative AI like ChatGPT.

AI training takes mounds of data and extracts trends from them in a manner that its model only takes a few minutes to generate an result, and can do so on a desktop computer – in sharp difference to the primary systems that governments have used for decades that can require many hours to run and need some of the biggest high-performance systems in the world.

Expert Reactions and Future Developments

Nevertheless, the fact that the AI could outperform previous gold-standard traditional systems so quickly is truly remarkable to weather scientists who have spent their careers trying to predict the world’s strongest storms.

“It’s astonishing,” said James Franklin, a retired expert. “The sample is now large enough that it’s pretty clear this is not a case of chance.”

He noted that although Google DeepMind is outperforming all other models on predicting the future path of storms worldwide this year, similar to other systems it occasionally gets high-end intensity predictions inaccurate. It struggled with another storm earlier this year, as it was similarly experiencing quick strengthening to maximum intensity north of the Caribbean.

During the next break, Franklin said he intends to talk with the company about how it can enhance the DeepMind output more useful for forecasters by providing additional under-the-hood data they can use to assess the reasons it is producing its answers.

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

Wider Sector Trends

Historically, no a commercial entity that has developed a high-performance weather model which grants experts a view of its techniques – unlike nearly all systems which are offered at no cost to the public in their full form by the governments that designed and maintain them.

The company is not alone in adopting artificial intelligence to address difficult weather forecasting problems. The US and European governments are developing their own artificial intelligence systems in the development phase – which have also shown improved skill over earlier non-AI versions.

Future developments in AI weather forecasts appear to involve new firms taking swings at previously difficult problems such as long-range forecasts and better early alerts of tornado outbreaks and flash flooding – and they have secured US government funding to pursue this. One company, WindBorne Systems, is also launching its proprietary atmospheric sensors to address deficiencies in the national monitoring system.