A Deep Learning Dataset for Pre‐Drill Geohazard Assessment in Taranaki Basin New Zealand

ABSTRACT Scientific ocean drilling is crucial for understanding subsurface processes but remains vulnerable to geohazards such as blowouts and wellbore instability, posing catastrophic risks to operations, human safety, and ecosystems. Progress in pre‐drill risk assessment remains limited due to the shortage of high‐quality, integrated datasets for deep learning applications. Here, we release an interpretated Taranaki Basin dataset that combines geophysical, drilling data, and overpressure data to advance pre‐drill geohazard assessment. The dataset contains 17 seismic attributes extracted along well trajectories and paired rock‐physical interpretations. This enables characterisation of four drilling geohazard categories: normal (safe) formations, wellbore collapse, overpressure, and combined overpressure‐collapse conditions. We propose three deep learning benchmarks: an unsupervised clustering model using solely seismic attributes, an informed model incorporating geological prior as constraints, and an enhanced variant of the informed model using increased trainable parameters. These benchmarks evaluate the ability of seismic data, and its integration with complementary data, to distinguish drilling geohazard factors. Validation against traditional methods highlights the dataset's utility for advancing predictive geohazard frameworks. This work promotes risk mitigation, fosters collaboration, and enables reproducible research.