Assessing the interpretability and robustness of MLP in tsunami prediction
Please login to view abstract download link
In this work, we employ multilayer perceptron (MLP) to predict tsunami wave heights and arrival times at various points along the coast of Spain, specifically focusing on the regions of Huelva and Cádiz. Using numerical simulations generated by the Tsunami-HySEA code, we create a comprehensive dataset with thousands of scenarios that model different tsunami conditions. These simulations serve as the training data for the MLP, allowing us to develop models capable of making rapid predictions based on seismic inputs. This approach is aimed at improving the efficiency and accuracy of the Spanish tsunami early warning system. While the MLP is widely recognized for its predictive capacity, its interpretability and robustness require careful attention. To address these issues, we conduct an advanced analysis using SHAP (Shapley Additive Explanations) values, which provide insights into the contribution of individual features (e.g., seismic parameters, geographic location) to the model's predictions. This helps increase transparency in decision-making, particularly critical in life-threatening scenarios like tsunamis. Additionally, we evaluate the model's resilience through adversarial attack simulations, analyzing how small perturbations in input data can lead to significant changes in output, potentially compromising the model's reliability. We also study the impact of outliers and data points outside the expected range, given the real-world variability of oceanographic conditions. By combining these techniques, we ensure that the model is not only accurate but also interpretable and robust, enhancing its applicability within Spanish tsunami early warning system for emergency response along the Spanish coast.
