Can AI fix New Zealand’s weather forecasting gap?

When severe weather hits New Zealand, the damage is often intensely localised.  

When Wellington was hit by heavy rain in April, some suburbs were impacted far more than neighbouring areas. In Canterbury, repeated hailstorms in January devastated some farms while nearby properties saw little damage. 

For communities, farmers and emergency responders, this kind of variability makes planning harder. Forecasts don’t always capture what happens at the local level, and the cost of getting it wrong can be significant. 

At Te Herenga Waka Victoria University of Wellington, AI researcher Dr. Emily O’Riordan is exploring whether artificial intelligence and high-performance computing (HPC) could help change that. 

Using REANNZ’s HPC platform, O’Riordan is developing new AI models designed to generate highly localised weather forecasts for New Zealand at a fraction of the computational cost required by traditional approaches. If successful, her research could eventually support more precise operational forecasting, helping meteorologists better identify severe weather risks before they hit. 

Forecasting the unpredictable 

Modern weather forecasting relies on Numerical Weather Prediction (NWP) models that solve the physical equations governing the atmosphere. These models divide the Earth and atmosphere into three-dimensional grid cells, calculating how temperature, pressure, wind, and moisture evolve over time. 

“Simply put, the higher resolution we can make our grid, the more precise we can get our weather forecasts to be,” says O’Riordan. 

Global forecasting models currently operate at resolutions of roughly 10 square kilometres. In New Zealand, regional forecasting systems refine that further, typically down to around four kilometres. But many of the most damaging weather events such as intense rain, hail and convective storms occur on much smaller scales. 

“Damaging weather events that we’re seeing often happen on a very fine scale,” says O’Riordan. 

Increasing model resolution, however, comes at a significant computational cost. Every reduction in grid-cell size dramatically increases the number of calculations required to model atmospheric behaviour. This means forecasting weather at hyperlocal scales using conventional physics-based methods quickly becomes prohibitively expensive. 

It’s this challenge that sits at the centre of O’Riordan’s research. 

Teaching AI to see hyperlocal weather 

Emily O’Riordan’s research trains AI models to add fine-scale local detail to broader weather forecasts, enabling more precise predictions at lower computational cost.

O’Riordan’s project explores whether generative deep learning models can learn the fine-scale patterns missing from lower-resolution weather forecasts. 

Rather than replacing existing forecasting systems, the research is designed to work alongside them. A neural network is trained using large volumes of historical meterological data, learning how to generate high-resolution local detail from broader NWP outputs. 

“We want to learn how to merge these together,” says O’Riordan, describing her hybrid approach combining neural networks with conventional physics-based forecasting systems. 

The goal is to create a hybrid neural-network and NWP approach capable of producing forecasts at scales measured down to the hundreds of metres.  

To train the model, O’Riordan needed to create high-resolution “target” datasets. By comparing lower-resolution forecast inputs with known high-resolution outputs, the model learns how to reconstruct fine-scale local weather patterns.  

"One of the key parts is that you need lots of examples of what you want the model to produce so it can learn how to map from lower-resolution inputs to higher-resolution outputs,” O’Riordan says.  

The research also explores ‘uncertainty awareness’ or how confident the model is in its own predictions, alongside methods for incorporating real-world observational data into forecasts. 

This awareness is critical. Weather systems are inherently chaotic, and forecasting severe localised events remains extraordinarily difficult. The challenge is finding a balance between not warning people enough, and sending them too many notifications that may turn out to be unnecessary.  

“If you give out a red warning every week, no one’s going to listen,” says O’Riordan. 

Scaling the impossible 

Training AI weather models at the scale needed for this project requires enormous computational capability. 

O’Riordan’s research involves terabytes of meteorological datasets, including hourly observations collected across New Zealand over more than a decade. The models process multiple atmospheric variables including rainfall, pressure, wind, temperature and humidity, all across densely gridded spatial datasets. 

“You wouldn’t be able to work through this without access to GPUs,” says O’Riordan. “You’ve got to use HPC for this stuff.” 

Before beginning the project, O’Riordan had little experience using HPC infrastructure and had previously avoided it because it felt overly complex. The research was originally started during her time at Bodeker Scientific, where she continues to collaborate, and later expanded using REANNZ’s services.  

Using REANNZ’s HPC platform, O’Riordan was able to train larger and more complex deep learning models while experimenting with different architectures and workflows. REANNZ consultancy support also helped optimise and debug her code. 

“The support team reply to your emails instantly. It’s incredible,” she says. 

From research to resilience 

While the project remains a work in progress, early prototype models are already showing promising results. 

If successful, the technology could eventually be incorporated into operational forecasting systems by the likes of Metservice, helping meteorologists generate more localised and computationally efficient weather forecasts for New Zealand. 

The implications extend well beyond more accurate forecasting alone. 

More precise severe weather prediction could help emergency management agencies better prepare for floods and storms, support growers facing increasingly volatile weather conditions, and improve risk assessment for industries exposed to climate-related disruption. 

Communities across New Zealand are already adapting to increasingly severe weather events. In regions such as Tairāwhiti and Hawke’s Bay, repeated flooding and cyclones have reshaped expectations around what constitutes “normal” weather. 

“They’re adjusting to a new kind of semi-normal,” says O’Riordan. 

As extreme weather becomes more frequent and more localised, the ability to forecast what happens on one street, one farm or one valley may become increasingly critical. That’s the scale where forecasting needs to work, and where access to serious computing power makes it possible.