Numerical results on the use of the L-SVD approach for the solution of the inverse source problem from amplitude-only data

We consider a Learned Singular Value Decomposition (L-SVD) approach to face a canonical 2D scalar inverse source problem from amplitude-only, far-field data. We compare the reconstruction performance of L-SVD from amplitude-only data against the Truncated SVD (TSVD) regularized inversion using amplitude and phase (complex) information. The numerical tests show that phaseless L-SVD provides, with a proper training on a well-organized dataset accommodating significant a priori information on the set of relevant unknown current distributions, superior performance as compared to TSVD.