The learned range test method for the inverse inclusion problem (Q6939614)

To reconstruct a perfectly insulated inclusion \(B\) inside a homogeneous medium \(\Omega\subset\mathbb{R}^d\) from measurements of Cauchy data at the boundary of \(\Omega\) is a variant of Calderón's inverse problem. The range test (RT) is a method designed to decide whether given test domains \(G\subset\Omega\) contain \(B\). Under certain assumptions, RT recovers from a single Cauchy datum convex polygons \(B\) with diameter strictly less than the distance to the boundary of \(\Omega\). A numerical algorithm realizing RT was proposed and studied in [\textit{S. Sun} et al., J. Comput. Phys. 485, Article ID 112099, 18 p. (2023; Zbl 1560.65786)].\N\NIn the paper under review, the authors propose a learned range test (LRT) to overcome weaknesses (selection of parameters, computational cost) of the numerical RT mentioned above. The LRT consists of the RT algorithm read as a 5-layer neural network equipped with learnable weights. The norm of Cauchy data depends strongly on the distance of \(B\) to the boundary of \(\Omega\). Therefore, LRT contains a second neural network, called classification step, which learns probabilities for these distances. Numerical experiments presented in the paper show that the LRT gives more precise results than the (deterministic) RT with much less computational cost. The experiments also show that reconstruction by purely data driven neural networks is much worse than LRT. The authors regard this as an indication that the architecture of the networks used in the LRT, which takes into account the physical model and the RT method, is a key aspect for a successful recovery of an inclusion.