Numerical methods for multiscale inverse problems

We consider the inverse problem of determining the highly oscillatory coefficient $a^{\epsilon}$ in partial differential equations of the form $-\nabla \cdot (a^{\epsilon} \nabla^{\epsilon}) + bu^{\epsilon} = f$ from given measurements of the solutions. Here, $\epsilon$ indicates the smallest characteristic wavelength in the problem ($0 \lt \epsilon \ll 1)$. In addition to the general difficulty of finding an inverse is the challenge of multiscale modeling, which is hard even for forward computations. The inverse problem in its full generality is typically ill-posed, and one common approach is to reduce the dimension by seeking effective parameters. We will here include microscale features directly in the inverse problem and avoid ill-posedness by assuming that the microscale can be accurately represented by a low-dimensional parametrization. The basis for our inversion will be a coupling of the parametrization to analytic homogenization or a coupling to efficient multiscale numerical methods when analytic homogenization is not available. We will analyze the reduced problem, $b=0$, by proving uniqueness of the inverse in certain problem classes and by numerical examples and also include numerical model examples for medical imaging, $b \gt 0$, and exploration seismology, $b \lt 0$.

Keywords

inverse problems, stability, heterogeneous multiscale method, periodic homogenization

2010 Mathematics Subject Classification

35B27, 35R25, 65N21, 65N30

Full Text (PDF format)

Published 21 February 2017