Second order adjoint sensitivity analysis in variational data assimilation for tsunami models

TL;DR

This paper develops a second order adjoint sensitivity analysis framework for variational data assimilation in tsunami models, enabling precise assessment of how observation perturbations affect initial condition and bathymetry reconstructions.

Contribution

It introduces a novel second order adjoint method to quantify sensitivities in variational tsunami data assimilation, extending previous first order approaches.

Findings

Derives the Hessian of the cost function for tsunami data assimilation.

Provides a mathematical framework for sensitivity analysis of observations.

Supports optimization of observation networks for improved tsunami prediction.

Abstract

We mathematically derive the sensitivity of data assimilation results for tsunami modelling, to perturbations in the observation operator. We consider results of variational data assimilation schemes on the one dimensional shallow water equations for (i) initial condition reconstruction, and (ii) bathymetry detection as presented in Kevlahan et al. (2019, 2020). We use variational methods to derive the Hessian of a cost function J representing the error between forecast solutions and observations. Using this Hessian representation and methods outlined by Shutyaev et al. (2017, 2018), we mathematically derive the sensitivity of arbitrary response functions to perturbations in observations for case (i) and (ii) respectively. Such analyses potentially substantiate results from earlier work, on sufficient conditions for convergence, and sensitivity of the propagating surface wave to errors in bathymetry reconstruction. Such sensitivity analyses would illustrate whether particular elements of the observation network are more critical than others, and help minimise extraneous costs for observation collection, and efficiency of predictive models.