Recent advances in machine learning have enabled neural weather models (NWMs) to achieve forecast skill comparable to traditional numerical weather prediction (NWP) models.

This presentation describes our ongoing work to develop an NWM trained as an emulator of the New Zealand Reanalysis (NZRA), a 1.5 km resolution product. Initial experiments focused on the Auckland region, where a U-Net and alternative architectures, including Vision Transformers and FourCastNet, were trained for next-step prediction. Recent work has incorporated generative modelling to produce ensemble forecasts and better resolve fine-scale detail. To compare against high-resolution NWP, we generated a one-year hindcast of 48-hour forecasts.

The current system demonstrates that NWMs can generate plausible, skillful high-resolution forecasts that are competitive with NWP. We discuss the role of reanalysis-based initialisation and large-scale atmospheric forcing in forecast quality, and the implications for transitioning towards an operational framework based on NWP guidance.

Finally, we present ongoing work towards a nationwide NWM, including the architectural and computational considerations required to scale from a regional prototype to a national forecasting system.

Data Assimilation (DA) is fundamental to numerical weather prediction. Currently, the state-of-the-art DA is done by a technique called 4D-Var, which optimizes an initial state to generate a sequence that is consistent with a prior forecast, and observations over the assimilation window. 4D-Var performs iterative updates to the initial state, which requires multiple numerical weather prediction model runs at a high computational cost.

Recent advances in generative AI have enabled sampling from learned distributions of realistic weather states. Using a technique known as posterior-free guidance, these models can be steered to generate samples that are consistent with available observations. This conceptual similarity to variational DA has motivated growing interest in AI-based approaches to data assimilation.

While current AI data assimilation systems do not yet generate initial conditions suitable for initializing full numerical weather prediction models, they can generate useful initial conditions with minimal computational cost. These analyses have the potential to support short-range nowcasting applications (up to 24 hours), where rapid updates are critical for sectors such as renewable energy and emergency management. Furthermore, some AI data assimilation systems do not require observations while training, meaning that new sources of guidance can be added without retraining.

Surface melt over ice sheets is a critical component of climate change. Therefore, there is an urgent need for skillful and efficient large-ensemble surface melt forecasts for both Greenland and Antarctica. Traditionally, researchers use weather forecast models to drive high-resolution regional atmospheric models to predict surface melt. However, running the physical regional models is computationally expensive, which limits large-ensemble forecasts to short lead times (fewer than 7 days for a few ensemble members).

In this talk, I will first present an emulator that successfully reproduces Greenland surface melt simulated by a regional climate model, using variables from the ECMWF ERA5 reanalysis as predictors. The emulator uses a convolutional neural network (CNN) architecture, which performs slightly better than a linear Ridge Regression model and significantly outperforms climatology baselines. Second, I will show case studies using large-ensemble (>50 members) ECMWF weather and seasonal forecasts to drive the emulator, demonstrating the success of this approach. Finally, I will discuss plans to provide real-time Greenland surface melt forecasts starting next summer.

This talk introduces Artificial Indigenous Relational Intelligence (AIRI), an emerging framework that draws on mātauranga Māori to rethink how AI can understand and support complex environmental systems. Grounded in concepts such as whakapapa (relational genealogy), mauri (life force), and reciprocal responsibility, AIRI focuses on the networks of relationships that generate ecological health and resilience. The presentation explores how Indigenous perspectives can contribute to conversations about AI, environment, and sustainability by highlighting the interconnectedness of human and more-than-human systems. Through this lens, environmental challenges are understood not simply as technical problems, but as questions of relationship, responsibility, and collective wellbeing. The talk illustrates how Indigenous knowledge can provide both cultural context for AI and alternative ways of thinking about the role of intelligent systems in supporting thriving environmental futures.

Earthquakes in the crust are the result of elastic strain accumulated between earthquakes being suddenly released in rapid fault slip. Our understanding of earthquake processes is inhibited by the long inter-seismic timescales of large earthquakes, however smaller earthquakes have proportionally shorter inter-event times. As such, detailed analysis of small earthquakes can provide insight into the processes governing large earthquakes. To test physics-based models of earthquake occurrence requires detailed observations of these small earthquakes over multiple seismic cycles. AI-based methods are now enabling the interrogation of long-duration (multi-decade) large (tens to hundreds of TB) datasets to build robust catalogues of earthquakes. However, the use of such methods requires significant testing and verification of observations. In this talk we discuss the application and testing of such methods, and the observations that they enable across Aotearoa New Zealand.

Advancements in AI-based detection methods enable us to interpret seismicity at scales not previously possible. My research focuses on utilising these tools to develop and apply a novel automated workflow to generate research-grade earthquake catalogues for New Zealand that balance the best of large- and small-scale earthquake catalogues.

This methodology was able to pick and locate 99.7% of all earthquakes in the 20-year GeoNet catalogue, totalling 407,553 events. This enabled further research into New Zealand's seismo-tectonic setting and a better understanding of present-day seismic hazard. Collaborating with colleagues at Earth Sciences New Zealand, I have extended and operationalised the workflow to allow for a containerised approach to ongoing automated cataloguing of New Zealand seismicity.

In this talk I consider the key benefits and drawbacks of this approach and discuss how this work may lay the basis for further research and development of seismic catalogues in New Zealand and globally.

Low-frequency (LF) volcanic earthquakes track fluid and magma movement and are used worldwide for forecasting volcanic eruptions. However, these weak and emergent signals are routinely missed by conventional threshold-based detectors, meaning our earthquake catalogues underrepresent the signals that matter most for eruption forecasting. We present work towards a deep-learning pipeline that combines neural phase picking, automated phase association, and waveform classification to detect and characterise LF earthquakes at New Zealand's volcanoes.

As a first step, we compare deep-learning pickers trained on tectonic earthquakes with those fine-tuned on volcano-seismic data (e.g. VOISS-Net, VolPick), assessing their sensitivity to LF arrivals. We are applying the workflow to ~20 years of continuous GeoNet data, with Ruapehu as the initial focus and testbed, before extending it to eight volcanoes. This reproducible approach is intended to enable consistent cross-volcano comparison of seismic behaviour through time, and to explore how automated detection tools can be adapted for volcanic eruption forecasting.

Tsunami inundation is ruled by geometry, the geometry of subduction zones, post-slip stress equilibria, ocean basins, harbours, and more. Whereas the physical geometry of tsunami inundation has profoundly shaped the tsunami modelling toolbox, the corresponding geometry of the inundation estimation problem has not been similarly exploited. This talk considers a state space formulation of tsunami evolution where a forward function evolves a snapshot into the spatiotemporal wavefield it generates. In this context we present a geometric interpretation of the state space, as well as its subspaces and manifolds where feasible tsunami states occur. We then discuss how machine learning methods can exploit this geometric structure to render forward evolution estimation problems tractable.

Monitoring New Zealand's land environment and terrestrial ecosystems requires information on land cover and vegetation composition, structure and change across spatial scales, from individual trees to nationwide maps. Remote sensing from drones, aircraft and satellites provides essential datasets in the form of optical imagery, radar backscatter and 3D LiDAR point clouds. We develop efficient, purpose-built computer vision algorithms for semantic and instance segmentation of imagery and 3D point clouds. Applications include native forest composition and health, carbon stocks, rare ecosystems (such as wetlands, sand dunes and peatlands), erosion and landslide risk, invasive weeds, and urban green spaces. We share operational experiences and discuss real-world challenges in training data generation, as well as the practical limitations of AI. We showcase examples from 3D tree wood segmentation and volume modelling, species classification in complex native forests, and high-resolution land cover segmentation.

Surficial seafloor sediments and their composition (e.g. grainsize, geochemistry) are influenced by a host of dynamic processes such as source sediment provenance, current speed and direction, and seafloor geomorphology. Traditional interpolation methods such as univariate kriging result in smoothed maps that are lacking in enough detail for applications such as habitat mapping. Here we compare traditional methods (kriging, Inverse Distance Weighting) and machine learning methods (RandomForests, RandomForests Spatial Interpolation) to evaluate their ability to predict and map seafloor sediment composition and distribution at high spatial resolution.

A brief tour of how ESNZ is using advanced remote sensing techniques and machine learning to understand and forecast landslide processes at different rates/scales. We will focus on current outputs from the Landslide Watch Aotearoa and Sliding Lands Endeavour programmes and offer a glimpse into future developments.

Rainfall-induced landslides (RIL) present a challenging machine learning problem characterised by rare events, spatial dependence, and environmental heterogeneity. We developed a 25 m resolution RIL susceptibility model using LightGBM and a landslide inventory mapped from pre/post Cyclone Gabrielle satellite imagery. Training data were carefully constructed to improve label quality and minimise spatial bias. Predictor variables were selected to represent physically meaningful controls on slope failure, including rainfall, topography, hydrology, land cover, and critical zone architecture. A positive monotonic constraint was imposed on rainfall to ensure physically realistic model behaviour. LightGBM was chosen for its ability to capture the non-linear relationships, variable interactions, and threshold effects characteristic of RIL processes. Model responses were explored using SHAP analysis, while class imbalance, probability calibration, and spatial generalisation were evaluated. The resulting model was then applied nationally to assess RIL susceptibility under future storm and climate scenarios, supporting hazard assessment and adaptation planning.