New paper published in Geophysical Journal International
In collaboration with G. A. Meles and N. Linde from the University of Lausanne, S. Marelli published a new paper on the combination of dimensionality reduction and surrogate modelling for geophysical tomographic imaging.
Dr. G. A. Meles and Prof. N. Linde from the Institute of Geophysics at the University of Lausanne published a external page new paper together with S. Marelli on the use of prior-knowledge based parametrization of inverse problems, in conjunction with surrogate models.
In the tomographic inversion community, parameterization of the inverse problem is most often derived from the discretization of the forward model solvers, generally resulting in highly underdetermined problems that require strong regolarization constraints to avoid ill-posedness.
In this paper, we remove the need of regularization by representing the available prior information on a suitable low-dimensional manifold, starting from a set of realizations of the generative prior model. This approach allows us to strongly reduce the dimensionality of the problem, from O(10^{4-5}) to O(10^{1-2}) parameters, thus enabling the use of surrogate modelling techniques such as polynomial chaos expansions (PCE).
We demonstrate the effectiveness of this approach by fully characterizing the posterior distribution of a stochastic subsurface GPR travel-time tomography problem, at the cost of O(10^2) finite difference -based forward model evaluations. We demonstrate that this approach compares favourably both in terms of accuracy and computational costs to more widely used approximations such as ray-tracing or straight-line solvers.
For more information, please follow external page this link for the publication and this link for the associated report on our internal archive.