Automatic differentiation: a tool for variational data assimilation and adjoint sensitivity analysis for flood modeling

William Castaings (1) , Denis Dartus (2) , Marc Honnorat (1) , François-Xavier Le Dimet (1) , Youssef Loukili (1) , Jerome Monnier (1)

1 MOISE - Modelling, Observations, Identification for Environmental Sciences
2 IMFT - Institut de mécanique des fluides de Toulouse

Flood modeling involves catchment scale hydrology and river hydraulics. Analysis and reduction of model uncertainties induce sensitivity analysis, reliable initial and boundary conditions, calibration of empirical parameters. A deterministic approach dealing with the aforementioned estimation and sensitivity analysis problems results in the need of computing the derivatives of a function of model output variables with respect to input variables. Modern automatic differentiation (ad) tools such as Tapenade provide an easier and safe way to fulfill this need. Two applications are presented in this paper: variational data assimilation and adjoint sensitivity analysis.

Domaines: Modélisation et simulation

William Castaings, Denis Dartus, Marc Honnorat, François-Xavier Le Dimet, Youssef Loukili, et al.. Automatic differentiation: a tool for variational data assimilation and adjoint sensitivity analysis for flood modeling. Lecture Notes in Computational Science and Engineering, 2006, 50, pp.249-262. ⟨inria-00259072⟩