How to Train Your Tsunami Emulator: From Open Science and Research Infrastructure to Stakeholders' Needs

Modern tsunami hazard assessment requires moving beyond slow high-fidelity simulations toward scalable hybrid frameworks that integrate physics-based numerical modelling with machine learning (ML) emulation. To ensure these "tsunami emulators" are trusted by stakeholders for tasks like hazard assessment, evacuation planning and real time forecasting, they must be developed through transparent, reproducible but tailored workflows. We present our attempt at building and testing tsunami inundation emulators designed for rapid probabilistic inundation assessment.

This work utilises of large simulation dataset derived from European research projects and computing infrastructure for training our emulator, that will be made available on the CINECA-hosted Simulation Data Lake (SDL) linked to the Geo-INQUIRE and EPOS project along with codes on open repository to allow other researchers to reproduce results, test, and also benchmark against new ML models.

We demonstrate through rigorous testing and benchmarking for an application at inundation sites in Sicily the emulator performance against full ensembles of numerical simulations and importance sampling Monte Carlo methods. Our emulation framework enables for uncertainty quantification of the emulator essential for trust and reliability in operational setting. The resulting products include probabilistic hazard maps, evacuation maps, inundation forecasts which are directly actionable for stakeholders. This example showcases a scalable path for integrating AI into solid Earth science using upcoming research infrastructures, helping bridge the gap between open science and real-world disaster resilience.