PhD defense of Pierre ANDRAUD

Pierre ANDRAUD

Rapid estimation of tsunami floods and currents from Deep Learning algorithms

Abstract

In a context where the time constraints imposed on the French National Tsunami Warning Center make high-fidelity numerical simulations unusable, this thesis proposes an alternative approach based on artificial intelligence to rapidly predict the coastal effects of a tsunami in the Mediterranean.

A first model, inspired by the literature and based on a dense neural network, was designed to directly estimate coastal wave heights, maximum velocities, and run-up from deep-ocean simulation outputs. Despite yielding promising results, its complexity limits its usability in operational settings.

A second, lighter approach leverages the latent structure of the data, obtained through Principal Component Analysis (PCA), to significantly reduce model complexity and training time. Two new models were developed: a dense neural network applied to PCA coefficients and a Gradient Boosting model. These approaches enable the generation of high-resolution flood maps within seconds, achieving performance levels comparable to classical simulations.

However, underestimation is observed in flooded areas, which are underrepresented in the training dataset. To improve the robustness of the predictions, a dedicated part of this thesis focused on the implementation of Bayesian methods for the quantification of epistemic uncertainties. Applied in particular to the latent models, these techniques yielded well-calibrated confidence intervals without increasing computational costs, paving the way for a fast, reliable, and operationally compatible tsunami warning tool.

Key words

Tsunami, Hybrid modeling, Machine Learning, Uncertainties

Supervision

• Nicolas VAYATIS
• Frédéric DIAS
• Audrey Gailler

Jury

• Stéphane ABADIE, Professeur des universités, Université de Pau et des Pays de l'Adour (UPPA), Rapporteur
• Ioan NISTOR, Professeur des universités, Université d'Ottawa, Rapporteur
• Maria Ana BAPTISTA, Professeure des universités, Université de Lisbonne, Examinateur
• Laure QUIVY, Maîtresse de conférences, ENS Paris-Saclay, Examinateur
• Gaël POËTTE, Ingénieur de recherche, CEA, Examinateur