Hundreds of Possible Futures for New Zealand's Extreme Rainfall

An Earth Sciences New Zealand-led team used generative AI to produce more than 15,000 simulated years of New Zealand rainfall and temperature in six hours. Those simulations cover hundreds of possible futures through to 2100, rather than a forecast stretching thousands of years ahead. In the simulations, the range of plausible rainfall extremes widened as the climate warmed, alongside increases in rainfall intensity.

The work, published in Climate Dynamics and described by Earth Sciences New Zealand, addresses a difficulty in preparing for rare weather events: the historical record contains relatively few examples. Detailed regional climate models can explore how those events might change, but their computational cost limits how many possible futures researchers can examine.

Study lead author Dr Neelesh Rampal, a data scientist at Earth Sciences New Zealand, says the constraint is particularly acute at the fine spatial scales needed to understand risks to communities and critical infrastructure. The team used AI to extend the reach of physics-based modelling, while retaining those simulations as the scientific foundation.

Training on warmer conditions

A model trained only on historical weather can struggle when future conditions differ from anything in its training data. Earth Sciences New Zealand describes the risk as unrealistic predictions or “hallucinations”, particularly for rare events such as tropical cyclones and extreme rainfall.

The researchers trained their AI on a 140-year simulation from a physics-based regional climate model. That training covered historical conditions and a projected warmer future, so the model encountered simulated temperatures and rainfall beyond those in historical records. Additional physics-informed constraints guided it towards realistic weather patterns and extremes.

The team then tested the AI against a separate set of physics-based climate simulations that had not been used in training. According to the institution’s account, it reproduced the expected climate patterns and projected changes in rainfall and temperature. This was a check against other simulations, rather than confirmation against weather that has yet to occur.

Beyond the average downpour

The resulting dataset contains daily rainfall and maximum temperature at a resolution of 12 kilometres across New Zealand. Producing that volume of detailed projections conventionally would have taken years, the institution reports.

Across New Zealand, the study found annual maximum rainfall, meaning the highest daily rainfall in each year, increased by approximately 5.6% per degree of warming. Earth Sciences New Zealand says the rainfall increases were similar to previous findings, but the finer resolution revealed more local detail.

The institution describes the dataset as New Zealand’s largest collection of high-resolution simulations to date. It also reports a finding not previously identified for New Zealand: increased warming led to greater natural volatility in atmospheric moisture, reflected in year-to-year extreme rainfall. The range of plausible rainfall extremes widened considerably. Further work is underway to understand this variability.

The stated public-good pathway is more informed climate adaptation. Exploring hundreds of possible futures can improve estimates of expected change and of rare, high-impact events that may place the greatest pressure on infrastructure, emergency management, water resources and communities. These remain modelled possibilities, with the institution emphasising careful design and testing of AI climate emulators.

The AI model was developed within Earth Sciences New Zealand. The institution reports that it provided finer regional information from 20 global climate models, four greenhouse gas emissions pathways and two large climate-model ensembles. Those inputs and the testing approach are described in the original Earth Sciences New Zealand account, which also links to the study.