Operational TEWS play a key role in reducing tsunami impact on populated coastal areas around the world in the event of an earthquake-generated tsunami. Traditionally, these systems in the NEAM region have relied on the implementation of decision matrices. The very short arrival times of the tsunami waves from generation to impact in this region have made it not possible to use real-time on-the-fly simulations to produce more accurate alert levels. In these cases, when time restriction is so demanding, an alternative to the use of decision matrices is the use of datasets of precomputed tsunami scenarios. In this paper we propose the use of neural networks to predict the tsunami maximum height and arrival time in the context of TEWS. Different neural networks were trained to solve these problems. Additionally, ensemble techniques were used to obtain better results.

This study focuses on the adaptation and enhancement of the Non-hydrostatic Multilayer version of the Tsunami-HySEA code to better model tsunamis triggered by granular flows. The aim is to optimize the code for operational hazard assessment, particularly for the Stromboli Island, and to meet the integration requirements for early warning systems for events occurring along the Sciara del Fuoco scarp. Significant improvements have been achieved, including a 50% reduction in computational time compared to the previous version. A prototype simulation has been developed to model tsunami propagation in the Tyrrhenian Sea, initiated by a landslide at Sciara del Fuoco on Stromboli Island. This simulation employs a coupled approach: first, it models the initial minutes of the event near Stromboli using the Multilayer-HySEA code in UTM coordinates. Then, the resulting data are transferred as input to other Tsunami-HySEA codes (hydrostatic or non-hydrostatic) operating in lat-lon coordinates. This two-step process enables efficient and accurate modeling of tsunami wave propagation across the entire Tyrrhenian Sea. Acknowledgments: This contribution was supported by the Center of Excellence for exascale in Solid Earth (ChEESE-2P) funded by the European High Performance Computing Joint Undertaking (JU) under grant agreement No 101093038 and by the grants PID2022-137637NB-C21 and PID2022-137637NB-C22 funded by MCIN/AEI/10.13039/501100011033 and by “ERDF A way of making Europe”.

This recipe aims at proposing a modelling tool to infer tsunami arrival times and maximum wave heights at selected points of interest (PoI) for an operational tsunami early warning system. The operational tool is based on the use of machine learning techniques, more specifically, on neural networks models. The methodoloy is applied prototypically for the coasts of Huelva and Cádiz, on the Atlantic coast of Southern Spain. The methodology proposed here can be applied to any other prone-to-tsunami coastal area.

Tsunami Early Warning Systems (TEWS) in the NEAM region (North-East Atlantic, the Mediterranean, and connected seas) operate under strict time constraints, particularly for near-field events where coastal impact may occur within a few minutes. In the NEAM region, operational chains typically use decision matrices and precomputed scenario databases. In Spain, the TEWS is operated by the Instituto Geográfico Nacional (IGN), and this work is carried out jointly with IGN to support operational decision-making. These established tools can be reinforced with rapid products that provide early indicators of coastal impact within some minutes or even seconds of the first source estimate. One example is the use of Faster-Than-Real-Time (FTRT) simulations, already implemented in the current system.Here we present a workflow in which neural-network surrogates are trained on large sets of physics-based tsunami scenarios, enabling fast inference of coastal impact metrics. The Tsunami-HySEA code is used to generate large-scale simulation sets, providing the data required by models designed for near-instant inference on standard CPUs. The surrogates models learn to map solid Earth earthquake source descriptors (capturing some uncertainty in fault parameters) to warning-relevant coastal metrics, focusing on maximum wave height and first-arrival time at multiple sites. Once trained, the models deliver predictions within seconds, facilitating rapid updates as source estimates evolve. Model interpretability is assessed using SHAP values, confirming how each input influences the predictions. The results confirm that the patterns follow the physical principles of tsunami generation and propagation. In an operational workflow, model results are fed into an automated reporting layer that produces tables, maps and graphics for Civil Protection within seconds, enabling rapid situational updates as source estimates evolve.We first report initial results for Atlantic sources affecting SW Spain. Approximately 250,000 HySEA simulations covering multiple Atlantic fault segments, focal mechanisms and magnitudes were used to train models. The results for forecast points along the Huelva–Cádiz coast show good agreement with observed patterns of maximum wave height and meet operational speed requirements, with errors remaining within the acceptable range for TEWS procedures. We then describe the extension of the methodology to the Western Mediterranean, covering the Spanish Mediterranean coast and the Balearic Islands. This extension involves defining and parameterising multiple tsunamigenic fault systems, assembling and controlling the quality of high-resolution topo-bathymetric datasets, and designing robust training and validation strategies.A practical limitation is that, despite comprehensive coverage of the targeted fault systems, rare source realisations or parameter combinations may fall outside the effective support of the training distribution, which can reduce reliability of point predictions. To handle such cases in operations, we complement deterministic estimates with threshold exceedance probabilities, enabling risk-aware decisions while preserving consistency with established TEWS procedures.

Operational Tsunami Early Warning Systems (TEWS) are crucial for mitigation and highly reducing the impact of tsunamis on coastal communities worldwide. In the North-East Atlantic, the Mediterranean, and connected Seas (NEAM) region, these systems have historically utilized Decision Matrices for this purpose. The very short time between tsunami generation and landfall in this region makes it extremely challenging to use real-time simulations to produce more precise alert levels and the only way to include a computational component in the alert was to use precomputed databases. Nevertheless, in recent years, computing times for a single scenario have been progressively reduced to a few minutes or even seconds depending on the computational resources available. In particular, the EDANYA group at the University of Málaga, Spain, has focused on this topic and developed the GPU code Tsunami-HySEA for Faster Than Real Time (FTRT) tsunami simulations. This code has been implemented and tested in TEWS of several countries (such as Spain, Italy, and Chile) and has undergone extensive testing, verification and validation.In this study, we propose the use of neural networks (NN) to predict the maximum height and arrival time of tsunamis in the context of TEWS. The advantage of this approach is that the inference time required is negligible (less than one second) and that this can be done in a simple laptop. This allows to consider uncertain input information in the data and still providing the results in some seconds. As tsunamis are rare events, numerical simulations using the Tsunami-HySEA are used to train the NN model. This part of the workflow requires producing a large amount of simulations and large HPC computational resources must be used.Machine learning (ML) techniques have gained widespread adoption and are being applied in all areas of research, including tsunami modeling. In this work, we employ Multi-Layer Perceptron (MLP) neural networks to forecast the maximum height and arrival time of tsunamis at specific locations along the Chipiona-Cádiz coast in Southwestern Spain. In the present work, initially several individual models are trained and we show that they provide accurate results. Then ensemble techniques, which combine multiple single models in order to reduce variance, are explored. The ensemble models often produce improved predictions.The proposed methodology is tested for tsunamis generated by earthquakes on the Horseshoe fault. The goal is to develop a neural network (NN) model for predicting the maximum height and arrival time of such tsunamis at multiple coastal locations simultaneously. The results of our analysis show that deep learning is a promising approach for this task. The proposed NN models produce errors of less than 6 cm for the maximum wave height and less then 212 s for the arrival time for tsunamis generated on the Horseshoe fault in the Northeastern Atlantic.

Operational Tsunami Early Warning Systems (TEWS) are crucial for mitigation and highly reducing the impact of tsunamis on coastal communities worldwide. In the North-East Atlantic, the Mediterranean, and connected Seas (NEAM) region, these systems have historically utilized Decision Matrices for this purpose. The very short time between tsunami generation and landfall in this region makes it extremely challenging to use real-time simulations to produce more precise alert levels and the only way to include a computational component in the alert was to use precomputed databases. Nevertheless, in recent years, computing times for a single scenario have been progressively reduced to a few minutes or even seconds depending on the computational resources available. In particular, the EDANYA group at the University of Málaga, Spain, has focused on this topic and developed the GPU code Tsunami-HySEA for Faster Than Real Time (FTRT) tsunami simulations. This code has been implemented and tested in TEWS of several countries (such as Spain, Italy, and Chile) and has undergone extensive testing, verification and validation.In this study, we propose the use of neural networks (NN) to predict the maximum height and arrival time of tsunamis in the context of TEWS. The advantage of this approach is that the inference time required is negligible (less than one second) and that this can be done in a simple laptop. This allows to consider uncertain input information in the data and still providing the results in some seconds. As tsunamis are rare events, numerical simulations using the Tsunami-HySEA are used to train the NN model. This part of the workflow requires producing a large amount of simulations and large HPC computational resources must be used. We utilized the Tsunami-HySEA code and the Spanish Network for Supercomputing (RES), to apply neural networks (NN) and obtain the numerical results.Machine learning (ML) techniques have gained widespread adoption and are being applied in all areas of research, including tsunami modeling. In this work, we employ Multi-Layer Perceptron (MLP) neural networks to forecast the maximum height and arrival time of tsunamis at specific locations along the Chipiona-Cádiz coast in Southwestern Spain. In the present work, initially several individual models are trained and we show that they provide accurate results. Then ensemble techniques, which combine multiple single models in order to reduce variance, are explored. The ensemble models often produce improved predictions.The proposed methodology is tested for tsunamis generated by earthquakes on the Horseshoe fault. The goal is to develop a neural network (NN) model for predicting the maximum height and arrival time of such tsunamis at multiple coastal locations simultaneously. The results of our analysis show that deep learning is a promising approach for this task. The proposed NN models produce errors of less than 6 cm for the maximum wave height and less then 212 s for the arrival time for tsunamis generated on the Horseshoe fault in the Northeastern Atlantic.