Parallel Joint Inversion

This ISCRA class C project has two main aims: (1) complete both the development and the test of an innovative approach to the problem of reconstructing observed objects by jointly inverting noninvasive, differently acquired observational data sets; (2) apply the proposed new approach, as well as recent state-of-the-art methods, to large-scale 3D geophysical data and possibly to electronic microscopy data. Inverse problems is an extremely active research field, which got in recent years increased attention due to both the huge amount of data provided by new technologies and the impressive development of effective methods. We are interested in an innovative approach to complex, large-scale volumetric imaging applications in Geophysics, but it can be very useful also in a number of other fields such as Medicine, Astronomy and Microscopy. The idea is that by jointly using more than one data set in the inversion process, one can get better results than those obtainable by separately processing each single data set and then "merging" the outcomes. This situation is very common in many different fields, where multiple sets of observations of the same physical system are acquired by different techniques. In the usual approach, each data set is used separately to recover an estimate of the system at hand and only afterwards these estimates are assembled together, most often on the basis of geometrical/physical considerations of the experts in the field. The current theoretical knowledge and the algorithmic advances on the numerical optimization side allow to face the harder difficulties associated with the solution of the joint inversion problem. Following the Tikhonov's approach, we consider to minimize a functional given by a misfit and a regularization term for each object's property to be reconstructed and, differently from the classical approach, a "joining" term which depends on all the investigated properties and embeds the additional requirement we want to impose to the solution. A large amount of data can thus be treated at the same time, coming from models obtained by different discretization techniques. We developed the current version of the code based on PETSc and TAO parallel libraries, using C and MPI. To face real-world applications, the power of modern multiprocessor architectures is strongly needed. In addition to the described benefits, a further outcome of this project will be to allow facing the difficult task of studying and testing techniques for the optimal choice of the various parameters involved, which most often requires to solve the same class of problems many times. The project is connected with Ambra Giovannini's Ph.D. work. PROJECT'S BUDGET: 15000 core hours. (CINECA's national ISCRA projects are issued as part of the European PRACE Tier0 access initiative)