Learning-Based Subsurface Quantitative Imaging via Near-Field Scanning Microwave Microscopy

Abstract

Accurate characterization of sample dielectric property is essential for scientists and engineers to analyze and evaluate the performance of the invested sample. This article proposes a learning-based (LB) method for quantitative subsurface imaging via coaxial-resonator-based near-field scanning microwave microscopy (CR-NFSMM) in a nondestructive way. To efficiently generate the database, a well-designed forward solver is critical since many samples will be included in the database, each sample needs a scanning procedure, and each scanning point involves solving the forward problem once. A fast forward problem solver is used to evaluate the tip–sample interaction in CR-NFSMM that avoids repeated meshing during the scanning process and it can apply to an arbitrary tip shape. Once the neural network is trained, it generates the reconstructed image within 1 s. Numerical and experimental results show that the proposed LB method could improve the resolution of the image and recover the dielectric properties of the subsurface perturbation pixel-by-pixel. Moreover, the proposed method outperforms the traditional objective function approach in terms of image resolution and time cost. The proposed method is promising to realize a nondestructive and real-time local dielectric evaluation of the subsurface object.