Research

Overview

My group develops probabilistic and machine learning models for natural hazard forecasting. We integrate time series analysis, Bayesian inference, and deep learning with real-time observational data to build tools that generalise across hazard systems and operate in under-resourced monitoring settings. Key hazards include volcanic eruptions, wildfires, snow avalanches, and floods.

A unifying thread across our work is transferability: extracting seismic, meteorological, or geophysical signal patterns at well-monitored sites and applying them, via transfer learning and ergodic statistical frameworks, to sites with sparse historical records. Our publications appear in Nature Communications, Journal of Geophysical Research, Geophysical Research Letters, and the International Journal of Wildland Fire.

🌋 Volcanic Eruption Forecasting

Volcanic eruptions pose major risks to communities and infrastructure, yet forecasting them remains difficult, particularly at volcanoes with little monitoring history. We use machine learning trained on seismic data to identify transferable eruption precursors.

Seismic Precursors & Transfer Learning

Our ergodic framework pools seismic features across 24 volcanoes to build generalisable eruption forecast models. Transfer learning then adapts these to data-scarce targets, achieving short-term forecast skill at volcanoes never seen during training.

Key paper: Nature Communications (2025) • doi:10.1038/s41467-025-56689-x

Whakaari / White Island (2019 Eruption)

Using the displacement seismic amplitude ratio (DSAR) as a proxy for shallow hydrothermal sealing, we showed that precursory signals before the deadly 2019 Whakaari eruption generalise to eruptions at other volcanoes worldwide.

Key paper: Nature Communications (2022) • doi:10.1038/s41467-022-29681-y • 60 citations

Template Matching for Hidden Fluid Release

Multi-timescale template matching applied to Ruapehu, Copahue, and Kawah Ijen detects hidden fluid-release episodes beneath crater lakes, a key precursory process rarely captured by standard monitoring.

Key paper: JGR Solid Earth (2023) • doi:10.1029/2023JB026729

Steamboat Geyser Eruption Forecasting

We investigate time scales, differentiability, and detectability of seismic precursors at Steamboat Geyser (Yellowstone) using data-driven methods, bridging volcano and geyser forecasting science.

Key paper: JGR: Machine Learning and Computation (2025) • doi:10.1029/2024JH000569

🔥 Wildfire Danger Forecasting

As climate change intensifies wildfire risks globally, decision-makers urgently need high-frequency, real-time fire danger information. We build ML models that deliver sub-hourly fire potential estimates from standard weather station data.

Sub-hourly Fire Potential Forecasting

Using time series of surface weather variables (temperature, humidity, wind, moisture), our ML pipeline produces fire-potential updates every 30 minutes, far finer resolution than existing daily danger indices.

Key paper: IJWF (2025) • doi:10.1071/WF24113

Cross-Region Transfer Learning

Models trained on regions with rich fire-weather records are adapted via transfer learning to Australian regions with limited historical data, demonstrating economic value of sub-hourly forecasts across diverse climatic settings.

Key paper: IJWF (2026) • doi:10.1071/wf25221

From Forecast Skill to Economic Value

We link forecast skill metrics to decision-theoretic and socio-economic value frameworks, providing practitioners with guidance on when and where ML fire forecasts justify operational deployment.

♨️ Geothermal & Geyser Systems

Geothermal systems and geysers offer natural laboratories for understanding hydrothermal fluid dynamics. We apply Bayesian geophysical inversion and multi-channel data modelling to image subsurface structure and quantify heat transfer.

Bayesian Magnetotelluric Inversion

We developed a Bayesian MT inversion method using methylene blue (MeB) structural priors to image shallow clay caps in geothermal fields, providing uncertain, probabilistic estimates of subsurface conductors at Wairakei-Tauhara.

Key paper: Geophysics (2021) • doi:10.1190/geo2020-0226.1 • 25 citations

Geothermal Heat Flux Quantification

Multi-channel data modelling of temperature, MT, and geochemical data constrains heat transfer through the Wairakei-Tauhara system, with implications for reservoir management and sustainability.

Key paper: GRL (2021) • doi:10.1029/2020GL092056 • 21 citations

Geyser Plumbing Geometry

Ground deformation data at El Tatio and other geyser fields are used to infer plumbing geometry (reservoir shape, conduit dimensions), improving our understanding of geyser eruption dynamics and seismic coupling.

Key paper: JGR Solid Earth (2019) • doi:10.1029/2018JB016454 • 18 citations

🌊 Floods, Hydrology & Avalanches

Real-Time Flood Forecasting

Physics-informed LSTM models for real-time river flow prediction in New Zealand catchments (e.g. Buller River), combining process understanding with the pattern-recognition strengths of deep learning.

ML in Hydrology Education

We demonstrate how modern ML approaches can be integrated into hydrology curricula, using river flow modelling as a case study to bridge data science and environmental engineering pedagogy.

Key paper: Journal of Hydrology NZ (2024) • Ardid, Pahlow, & Dempsey

Avalanche Detection via Infrasound

Using infrasound sensors in the Hooker Valley (Aoraki/Mt Cook area), we detect and characterise snow avalanche activity, developing automated ML-based detection pipelines for remote mountain settings.

Key paper: NZ Journal of Geology and Geophysics (2026) • Watson, Miller, Anderson, Toney, & Ardid