The devastating 2011 Tohoku-oki tsunami highlighted the need for improved real-time tsunami inundation prediction, particularly in the near field. Current early warning systems lack robust inundation forecasting capabilities. Traditional physics-based models, while accurate, are computationally expensive, requiring high-performance computing for real-time application. The delay in processing, coupled with uncertainties in tsunami source estimations, limits their effectiveness. Japan's S-net, a large-scale cabled seafloor observatory, provides high-frequency real-time data from 150 offshore stations. This study leverages S-net's capabilities to develop a machine learning-based model for rapid and accurate near-field tsunami inundation prediction, aiming to significantly reduce computational time and improve forecast lead time by directly utilizing offshore observations, thereby bypassing the need for computationally expensive source estimations that introduce uncertainty.

Previous studies have explored machine learning for tsunami inundation prediction. Fauzi and Mizutani (2020) and Mulia et al. (2020) proposed machine learning models, but these still required tsunami source estimations and linearized shallow water equation solutions. More recent works (Makinoshima et al., 2021; Liu et al., 2021) utilized offshore data to predict water level fluctuations at limited sites, providing detailed temporal information. However, these methods lacked the spatial detail and broad coastal coverage necessary for comprehensive near-field inundation mapping. This study addresses these limitations by developing a model that predicts the spatial distribution of tsunami inundation flow depths across a broad coastal region with high resolution, directly using offshore observations without relying on intermediate source estimations.

The study focuses on seven cities along Japan's southern Sanriku coast, a region highly vulnerable to tsunamis. A stochastic slip model was used to generate 3093 hypothetical tsunami scenarios from megathrust and outer-rise earthquakes, varying in magnitude. These scenarios served as the training data for a fully connected artificial neural network. The input layer consisted of 150 neurons representing S-net stations, with the output layer representing flow depths at over 200,000 grid points across the study area. Two hidden layers with 150 neurons each were used, with a ReLU activation function and a 20% dropout rate. The Adam optimization algorithm was used for training. The model was tested against 480 unseen hypothetical scenarios and three historical tsunami events (2011 Tohoku-oki, 1896 Meiji Sanriku, and 1933 Showa Sanriku). Model performance was evaluated using Aida's numbers (geometric mean ratio K and standard deviation x), accuracy percentage, goodness-of-fit (G), and misfit analysis. The impact of malfunctioning S-net stations on prediction accuracy was also investigated.

The machine learning model achieved high accuracy (98% for a 20-min prediction window using maximum tsunami amplitudes as input) comparable to physics-based models but with significantly reduced computational time (0.05 seconds vs. 30 minutes). Aida's numbers for the test set were K = 1.02 and x = 1.47, indicating good agreement with the target inundation maps. Analysis of misfits showed a normal distribution centered around zero, suggesting unbiased predictions. Goodness-of-fit values were higher (worse fit) for smaller magnitude earthquakes, as expected due to their higher variability. The model performed better in predicting larger tsunamis. Spatial variability analysis revealed higher misfits in areas with complex coastal geometry. Experiments simulating malfunctioning S-net stations showed that the model maintained reasonable accuracy (84.8%) even with 140 malfunctioning stations. Application to real events showed comparable accuracy to physics-based models, although underestimation of smaller flow depths (0.2-1m) was observed. Prediction uncertainty due to observational errors was quantified using ensemble predictions with perturbed inputs.

The machine learning model offers a significant improvement over traditional physics-based models for real-time near-field tsunami inundation prediction, achieving comparable accuracy with drastically reduced computational time. The model's ability to directly utilize offshore observations eliminates uncertainties associated with source estimation and allows for faster predictions. The results highlight the model's effectiveness, particularly for larger tsunamis, which are most critical for early warning. While the model generally performs well, limitations exist, particularly in areas with complex coastal geometries and when many S-net stations malfunction. The study's findings demonstrate the potential of machine learning to enhance tsunami early warning systems.

This study demonstrates a highly efficient and accurate machine learning-based method for real-time near-field tsunami inundation prediction. The model leverages the extensive S-net data to bypass computationally expensive source estimations, achieving near-real-time predictions with accuracy comparable to physics-based methods. Future research should focus on improving robustness against malfunctioning sensors, expanding the training data to encompass a wider range of events, and developing methods for a more comprehensive quantification of prediction uncertainty. The approach holds significant potential for enhancing global tsunami mitigation capabilities.

The model's performance is affected by complex coastal geometries and malfunctioning sensors. While the study used a wide range of tsunami scenarios, unprecedented events outside this range could lead to inaccurate predictions. The quantification of prediction uncertainty, while addressed in the study, could be further improved by considering model uncertainty and integrating with probabilistic tsunami hazard assessments (PTHA).