CSE Community Seminar
December 6, 2024, 12-1PM
Enforcing realism in hydrogeological and geophysical inverse modeling
Niklas Linde
Full Professor | Institute of Earth Sciences (ISTE),
University of Lausanne, Switzerland
Abstract:
The classical approach to inverse modeling in the geosciences consists of fitting the available data by perturbing a subsurface model by gradient-based optimization such that one unique model is obtained. A geophysicist or groundwater hydrologist pursuing interdisciplinary collaborations may then be placed in one of two equally uncomfortable situations when presenting the model obtained by such an inversion strategy: the collaborators may consider it as reality (it is not) or as useless as it doesn’t contain realistic-looking texture or features. As an alternative, the late Albert Tarantola proposed a movie strategy in which the disciplinary scientist should judge if the images passing by in the “movie” appeared realistic and the geophysicist should produce realizations that not only had this realistic look, but also explained the available data. Such an inversion approach needs to be probabilistic in nature, using for instance a Bayesian formulation and it comes with several conceptual and computational challenges. This talk will explain how such inversions can be achieved using multiple-point statistics methodologies or deep generative modeling. It will then demonstrate how computational speed-ups can be achieved by Bayesian variational inference and/or using surrogate models emulating computationally expensive physics-based numerical simulations. Finally, Bayesian model selection will be introduced to demonstrate how data can be used to test alternative conceptual models. Illustrative examples will be given in the context of environmental geophysics and groundwater hydrology.
December 6, 2024, CSE Community Seminar
Niklas Linde
Full Professor | Institute of Earth Sciences (ISTE)
University of Lausanne, Switzerland