A new Bayesian framework for Imaging the Subsurface and associated HydrOdynamic Processes (BISHOP)
A new Bayesian framework for Imaging the Subsurface and associated HydrOdynamic Processes (BISHOP)
Abstract
In geosciences, inverse approaches are commonly used to estimate model parameters based on observed data. Deterministic methods suffer from critical difficulties: an implicit prior and/or conceptual model assumptions need to be rightfully chosen and geophysical imaging is inherently non-unique. Moreover, for applications such as hydrogeophysics where hydrodynamic processes are occurring, realistic patterns of heterogeneity are necessary. A way to overcome those issues is to use stochastic inversion methods, based on Bayesian formalism that enables to build an ensemble of models that fit the data. Whereas interesting on its foundations, this method is computationally expensive and is therefore not widely used in real field. Our project proposes a new Bayesian framework called Bayesian evidential learning to overcome those impediments, for instance in complex systems. The proposed framework consists in four main steps: prior definition from broad geological information and site-specific details using Monte Carlo sampling, experimental design using a distance based global sensitivity analysis, prior falsification and prediction focused approach in a reduced dimensional space. This new framework addresses important difficulties of deterministic or stochastic inverse modeling methodologies: a need for a well-defined prior distribution, a focus on prediction of key decision variables and a definition of relevant data, informative to the key decision variables within a computationally-friendly framework. The project focuses on cases where predictions are geological images or key subsurface parameters, on experimental design and on prior falsification with applications on surface nuclear magnetic resonance and surface waves dispersion interpretation and experimental design. Nonetheless, the developed framework could be applied to any prediction problem.
Publications
BEL1D: 1D imaging using geophysical data in the framework of Bayesian Evidential Learning (2020)
1D geological imaging of the subsurface from geophysical data with Bayesian Evidential Learning (2020)
Improving BEL1D accuracy for geophysical imaging of the subsurface (2020)
1D geological modeling of the subsurface from geophysical data with Bayesian Evidential Learning (2019)
Improving the accuracy of 1D SNMR surveys using the multi-central-loop configuration (2018)
