Projects

DeepWeather

AI-enhanced Numerical Weather Prediction for Aotearoa New Zealand

DeepWeather is a research programme developing hybrid AI–physics weather prediction systems tailored to the unique meteorology of Aotearoa New Zealand.

Traditional Numerical Weather Prediction (NWP) models are computationally expensive and often struggle to resolve fine-scale variability, particularly in regions with complex terrain and coastlines. DeepWeather addresses this by combining dynamical models with machine learning, enabling higher-resolution, more efficient, and more adaptable forecasts.

Current work focuses on:

• Coupling AI models with dynamical systems (e.g. WRF) to enhance spatial detail and forecast skill
• Learning corrections to model biases and unresolved processes
• Evaluating performance on extreme weather events and regional variability

A key component of the project is ensuring that models are validated against real-world, high-resolution observations, particularly in regions most vulnerable to extreme weather.

Spectral Machine Learning for Physical Systems

Beyond Fourier: learning in alternative spectral representations

Many modern AI weather models rely on Fourier transforms, which impose global smoothness and can limit the representation of sharp gradients and localised extremes.

This project investigates whether alternative spectral representations — including wavelets and other localised basis functions — can improve the fidelity of AI-based forecasting systems.

Core research questions:

• Can localised spectral methods better capture extreme events and fine-scale structure?
• How should neural networks operate directly in spectral space?
• What loss functions (e.g. power spectrum–based) best preserve physically meaningful structure?

This work includes:

• Designing architectures that operate in transform space (FFT / wavelet / hybrid)
• Developing physically informed loss functions (e.g. PSD alignment)
• Systematic comparison across climatology, extremes, and regional domains

The goal is to build spectral AI models that are both mathematically principled and operationally useful.

Mātaki Marangai

Community-led rainfall monitoring and data sovereignty

Mātaki Marangai is a community-driven rainfall monitoring initiative based in Tairāwhiti, developed in collaboration with schools, scientists, and local partners.

The project deploys:

• ~100 rain gauges across households
• Automatic weather stations at local kura
• Data collection tools co-designed with students

Its goals are to:

• Improve understanding of highly variable rainfall patterns in a vulnerable region
• Build local capability in environmental data collection and analysis
• Support decision-making through locally owned data

Tairāwhiti experiences strong spatial variability in rainfall due to its terrain and exposure, making forecasting particularly challenging. By generating dense, local observations, Mātaki Marangai creates a high-resolution dataset that complements traditional meteorological networks.

Importantly, the project embeds:

• Kaupapa Māori approaches, including integration with maramataka
• Community ownership of data and tools
• STEM engagement for rangatahi

The data collected also plays a role in validating AI-based forecasts developed in DeepWeather, linking community science directly to cutting-edge modelling.

Data Assimilation with Neural Processes

Learning from sparse, irregular environmental observations

Environmental observations are inherently sparse, noisy, and unevenly distributed. This project explores the use of neural processes and related probabilistic models for data assimilation.

Focus areas:

• Learning continuous representations from irregular observations
• Combining observational data with model outputs
• Quantifying uncertainty in predictions

This work is particularly relevant for:

• Weather forecasting in data-sparse regions
• Integrating citizen science datasets
• Real-time environmental monitoring