AI reveals how a warming climate will intensify NZ’s most extreme downpours

The most damaging weather events are the record-breaking “one-in-100-year” disasters communities have usually only seen once.

But understanding how these rare events could change in the future is hard. With few examples in historical records, it is difficult to estimate how severe they could become and how outcomes might vary across possible futures.

Recent research shows the weather systems bringing heavy rain to New Zealand each year are expected to become more intense as the climate warms.

In our new study, we used generative artificial intelligence (genAI) to simulate hundreds of ways this century might unfold and to explore how extreme rainfall might change across these possible futures.

We know that a warmer atmosphere can hold more water vapour, allowing storms to produce heavier rainfall. But much of our understanding comes from global models whose grids are so coarse that entire region, such as Auckland, can fall within a single grid cell.

By using genAI to explore these changes at finer scales across New Zealand, we found the rarest downpours become more intense as the planet warms, with larger increases than for the heavy rainfall we experience more often.

In a future where greenhouse gas emissions keep rising, the amount of rain falling per day in a one-in-100-year event increased by about 9% per degree of warming, averaged across New Zealand and the possible futures we simulated. For the heavy rainfall we experience more often, the increase was about 5.6% per degree.

The heavier the daily downpour, the greater the potential for flooding and damage.

A one-in-100-year event means there is a 1% chance of rainfall exceeding record amounts in any year. But these events do not happen neatly once a century. They can happen close together, and their odds change as the climate warms.

The average increases we describe tell only part of the story. Our study found warming also widened the gap between milder and more severe possible futures. In the more severe futures, extreme rainfall increased substantially more than the averages above.

Understanding the range of possible outcomes, including the best and worst cases in our simulations, gives us a clearer picture of the rare, damaging rainfall events we need to prepare for.

How AI makes this possible

Conventional climate models use the laws of physics to simulate the climate. Detailed regional models require powerful computers, limiting how many versions of the future researchers can explore.

Our team trained the AI on 140 years of regional climate simulations covering historical and warmer future conditions. It learned how global weather patterns translate into local rainfall and temperature.

After about a day of training, the AI generated more than 15,000 simulated years in six hours using four specialised computer chips. Conventional regional models would have taken years.

This does not mean looking 15,000 years into the future. The total includes many reruns of history and of possible futures to 2100, drawn from 20 global climate models.

Each represents New Zealand in grid cells about 12 kilometres wide, allowing us to examine regional differences. Our earlier review found that similar approaches can also work in other parts of the world, and could help regions better understand their own climate risks.

A rugby analogue

To understand how these potential climate futures might play out, it can be helpful to imagine the All Blacks playing Australia at Eden Park.

As we saw earlier this month, the All Blacks are the stronger team and we expected them to win, especially at Eden Park.

But every match unfolds differently, and if we were to replay that same match again, a missed penalty, injury or last-minute red card could change the final score.

If we could replay the same match hundreds of times, the All Blacks would win most of the time but some matches would end in unexpected losses. Six replayed matches would give only a rough picture of rare outcomes such as a loss to the Wallabies.

Hundreds would give us a better chance of understanding these more surprising possibilities, and how they depend on the venue or other factors.

Climate works similarly. Rising greenhouse gases push up temperature and influence extreme rainfall. But natural variability, including everyday weather and longer-term fluctuations such as the El Niño Southern Oscillation, affects when and where extremes occur and how severe they become.

Even with the same emissions, small differences in starting conditions lead to different outcomes. Earlier research shows how simulating many futures (by slightly varying these conditions) allows researchers to examine both the overall direction of change and less likely outcomes.

A clearer picture

Our study shows that this variability grows with warming, widening the range of possible rainfall extremes faster than their average intensity increases.

Returning to rugby, the equivalent would be if the All Blacks started winning by more points on average, but their results also became much more varied, with unexpected losses and bigger blowout wins.

As more groups now use AI in this way to project future climate conditions, work is underway to coordinate methods globally to develop shared standards.

AI can produce poor outputs under unfamiliar conditions. This problem plagued some earlier studies, but training on both historical and warmer climates helped to address this, alongside a training strategy our previous work showed would produce realistic extreme events.

More simulations cannot tell us exactly where or when the next record-breaking downpour will happen. But they give us a clearer picture of what to prepare for, including extremes beyond anything communities have experienced in living memory.