Using GPUs to generate realtime forecasts of tsunami hazards

Local tsunamis can reach shores in minutes. REANNZ compute resources support modelling research for better forecasts and future tsunami warnings.

Research background

New Zealand faces significant tsunami risk due to its location on the Pacific Ring of Fire, bordered by the Hikurangi Subduction Zone where the Pacific Plate dives beneath the North Island. Distant tsunamis from sources like Chile, Japan, or Alaska provide hours of warning time, allowing effective evacuation. However, local and regional tsunamis generated close to shore — particularly along the Hikurangi margin — can reach coastlines within minutes, leaving virtually no time for official warnings to be issued and acted upon.

The core challenge for scientists is that seismic monitoring, tsunami modeling, and public alert systems all take time. Detecting an earthquake, running ground deformation models, propagating these through tsunami simulations and disseminating results to the public within a useful timescale involves making assumptions about the earthquake size, depth and type (e.g. seismic, landslide, meteo). The problem for decision makers is how to account for these assumptions when generating warnings.

Energy map for a HySEA simulation in the North-Eastern Atlantic. From: https://www.uma.es/edanya/info/151578/hysea-platform/

 

Project challenges

Tsunami-HySEA (Hyperbolic Systems and Efficient Algorithms) is a high-performance GPU-accelerated numerical model developed by the EDANYA group at the University of Málaga, designed to simulate tsunami propagation and inundation by solving the shallow water equations on structured grids. Its GPU architecture allows simulations that would take hours on traditional CPU clusters to complete in minutes, making it practical for operational forecasting and large ensemble runs.

The Monte Carlo configuration of Tsunami-HySEA, T-HySEA MC, exploits this speed by running thousands of simulations across a randomised ensemble of earthquake source parameters — varying fault location, strike, dip, rake, slip, and depth within physically plausible ranges derived from seismic hazard models. Each realisation produces a distinct seafloor deformation and resulting tsunami. Aggregating these ensembles allows scientists to construct probabilistic hazard maps showing, for example, the 10% probability of exceedance wave height at any given coastal location. This probabilistic approach captures source uncertainty far more rigorously than single deterministic scenarios, producing forecasts that better inform risk assessment, evacuation zone planning, and insurance modelling for tsunami-exposed communities.

Maximum water height for a HySEA simulation in the Western Mediterranean. From https://cheese-coe.eu/flagshipcode/t-hysea-tsunami-hysea

 

What was done

Together with Mika Liao from the Natural Hazards and Resilience Platform, REANNZ Research Software Engineers deployed HySEA on different GPU resources to determine performance characteristics of the model. By systematically varying the number of GPUs against the number of parallel simulations, Mika explored how workload distribution affects throughput and efficiency across different hardware configurations. To maximise flexibility across REANNZ's heterogeneous GPU infrastructure, HySEA was compiled as a fat binary targeting NVIDIA L4, A100, and H100 architectures, allowing a single binary to run seamlessly across all available GPUs. This work provided valuable insights into optimal GPU resource allocation for probabilistic tsunami hazard assessment workflows, as it highlighted the tradeoff between probabilistic accuracy and resource utilisation based on earthquake location and approximate arrival time of the tsunami.

Maximum water height for a HySEA simulation in Wellington for a magnitude 7.5 earthquake in the middle of the Harbour

 

Predicted tide gauge response for a HySEA simulation in Wellington for a magnitude 7.5 earthquake in the middle of the Harbour

 

Researcher feedback

Without REANNZ there was no way we could experiment with different GPUs and model configurations to tune the model to New Zealand conditions. This will provide guidance as to how to operationalise HySEA based tsunami warnings..

- Mika Liao, Natural Hazards and Resilience Platform

 


 

This case study shares some of the technical details and outcomes provided through our Consultancy Service. This service supports projects across a range of domains, with an aim to lift researchers’ productivity, efficiency, and skills in research computing. Get in touch to discuss how our Research Software Engineers and specialist support could help advance your project.

Find anything about our products, services, and more. Enter a query in the search input above.