Sea ice plays a critical role in regulating Earth’s climate and ocean–atmosphere interactions, yet its behaviour around Antarctica has proven highly unpredictable. While Arctic sea ice has steadily declined since satellite observations began, Antarctic sea ice confounded expectations by continuing to expand until it collapsed with unprecedented speed in recent years. The causes and predictability of these changes remain poorly understood, as both the prolonged growth and its sudden reversal defied climate model predictions. In this work, we apply modern graph neural network architectures (DeepGAT, GraphSAGE, and DCRNN) designed for spatio-temporal data to forecast Antarctic sea ice concentration and its anomalies up to eight weeks ahead. Using a novel graph representation of the sea ice, ocean and atmospheric dynamics derived from reanalysis and satellite observations, we show that graph-based models substantially outperform statistical baselines and dynamical systems. The results demonstrate that graph neural networks can capture complex spatial and temporal dependencies, offering a promising new pathway for subseasonal sea ice prediction and polar climate research.

Antarctica plays a critical role in regulating the global climate, yet its remote location and harsh conditions make accurate weather monitoring and forecasting particularly difficult. A sparse and unevenly distributed network of automatic weather stations, combined with the known limitations of existing numerical weather prediction models such as AMPS in capturing surface-level events, underscores the need for new predictive approaches. With heatwaves increasing in frequency and severity globally, and no prior machine learning-based study addressing extreme temperature prediction for Antarctica, this work investigates that gap. We employ a graph neural network (GNN) to predict average daily temperatures and heatwave events across the Antarctic continent, trained on a newly constructed dataset combining automatic weather station records with satellite-derived temperature products. Our findings contribute a novel data-driven framework for Antarctic temperature forecasting, with implications for climate science, glaciology, and operational safety in the region.

Foehn winds are accelerated, warm and dry winds that can have large socioeconomic and environmental impacts as they descend into the lee of a mountain range. They can be found in many alpine areas around the world, including the New Zealand Alps and the McMurdo Dry Valleys (MDV) in Antarctica. In the latter, foehn winds can cause ice and glacial melt and by extension destabilise ice shelves. Hence, detecting past and future foehn events from in-situ sensors or climate models is of significant interest. However, to the best of our knowledge, no unsupervised machine-learning approaches have been applied to this domain to provide a data-driven analysis of foehn wind patterns. In this study, we show that unsupervised methods such as k-means clustering and autoencoder-based anomaly detection (AD) can produce comparable results to current rule-based and supervised approaches.

The potential of leveraging network science in the area of law has long been advocated and highlighted through case and legislation networks. Yet this particular subdomain of data management and information retrieval is still heavily underutilised in both practice and research. One of the contributing factors to this problem is the lack of openly available legal data. This work describes the development of a legal citation network for New Zealand. In contrast to traditional case citation networks, this data repository also includes legislation and court data. Our network provides the data and references from over 300,000 decisions, 10,000 legislations and 115 courts from all levels of jurisdiction. Additionally, we present exemplifying network analysis results to reveal previously hidden information about the New Zealand legal system and to motivate future research in this domain.