The Brief
Science & Discovery 4 min read

9% Per Degree: What AI Found Inside Rare Rainfall

NAVION

Share

The most destructive weather events are often the ones communities have only experienced once. A flood that overwhelms infrastructure, a downpour that exceeds every historical record: these are the events that planning systems struggle to prepare for, precisely because there is so little data about them. A new study changes that picture, using generative AI to simulate thousands of possible climate futures and reveal how New Zealand’s rarest and most extreme rainfall events could intensify as the planet warms.

The Numbers That Most Climate Models Cannot See

Standard climate models divide the Earth into a grid. Each cell in that grid can be so large that an entire city, Auckland included, fits inside a single square. At that resolution, regional differences in rainfall disappear into averages. The study addressed this limitation by training an AI system on 140 years of regional climate simulations, covering both historical conditions and warmer projected futures. The AI learned to translate broad global weather patterns into local rainfall and temperature at a resolution of roughly 12 kilometres per grid cell.

The computational difference is striking. Conventional regional models would have required years to generate the volume of data the AI produced. After approximately one day of training, the system generated more than 15,000 simulated years in six hours, running on four specialised computer chips. Those simulations drew from 20 global climate models and covered many reruns of history alongside possible futures extending to 2100.

This is not a forecast of what will happen. It is a structured exploration of what could happen across a wide range of plausible futures, including both moderate and high emissions scenarios.

Rare Events Are Accelerating Faster Than Common Ones

Here is what most coverage of climate and rainfall misses: not all rain responds to warming at the same rate. The study found that heavier, rarer downpours intensify more sharply than the heavy rainfall communities already experience regularly.

In a future where greenhouse gas emissions continue rising, daily rainfall in a one-in-100-year event increased by about 9% per degree of warming, averaged across New Zealand and the simulated futures. For more frequent heavy rainfall, the increase was about 5.6% per degree. The rarest events, in other words, are pulling away from the rest.

A one-in-100-year event does not mean a flood happens once every hundred years on a fixed schedule. It means there is a 1% chance of rainfall exceeding record levels in any given year. As the climate warms, those odds shift. Events that were once genuinely rare can become more frequent, and their severity can exceed anything in living memory.

The study also found that warming widened the gap between milder and more severe possible futures. In the worst-case simulations, extreme rainfall increased substantially beyond the averages. The range of outcomes grew faster than the average intensity. This is the detail that matters most for infrastructure planning: it is not just that extremes get worse on average, it is that the uncertainty around those extremes expands.

The researchers used a rugby analogy to explain this dynamic. If a stronger team starts winning by more points on average, but their results also become far more variable, with unexpected losses alongside bigger victories, you cannot plan for outcomes by looking only at the average. The same logic applies to rainfall planning.

What This Means Beyond New Zealand

The significance of this research extends well beyond any single country’s flood risk. It demonstrates a method. Generative AI, trained on historical and projected climate data, can explore the statistical edges of possible futures at regional scales, at a speed and volume that conventional modelling cannot match. The researchers note that similar approaches have shown promise in other parts of the world, suggesting this could become a standard tool for regional climate risk assessment globally.

Work is already underway to coordinate methods across research groups and develop shared standards, a necessary step as more teams adopt AI for climate projection.

There are real limitations to acknowledge. AI systems can produce unreliable outputs when they encounter conditions outside their training data. The study addressed this by training on both historical and warmer climate conditions, and by applying a training strategy designed to produce realistic extreme events. Even so, more simulations cannot pinpoint exactly where or when the next record-breaking downpour will occur.

What they can do is sharpen the picture of what communities need to prepare for. Infrastructure designed around historical rainfall records may already be undersized for the futures now being simulated. The gap between what has been built and what the climate may deliver is precisely what this kind of research is designed to measure.

In Short

Generative AI allowed researchers to simulate more than 15,000 years of possible climate futures for New Zealand in six hours, a task that would have taken years with conventional models. The core finding: the rarest, most damaging rainfall events intensify faster than common heavy rain, at roughly 9% per degree of warming compared to 5.6% for more frequent events. Warming also widens the spread between best and worst outcomes, not just the average. The method itself may matter as much as the results: it offers a scalable way for regions around the world to understand climate risks that historical records alone cannot reveal.

Based on reporting from The Conversation.

Written by

NAVION