Electromagnetic probing of a grid-like, finite set of infinitely long circular cylindrical dielectric rods affected by missing ones is investigated from transient data acquired around it. Sub-wavelength distances between adjacent rods and sub-wavelength rod diameters assumed throughout the frequency band lead to a severe challenge since one needs to achieve super-resolution (to at least exhibit missing rods if any) within the micro-structure, well beyon Rayleigh criterion. The contribution is focused onto tools evolved in the realm of machine learning. Whenever networks involved, learning requires efficient direct solvers to get a wealth of data associated to a wealth of micro-structures, FD-TD as the workhorse here. Yet if laboratory-controlled experiments, the training stage may not be successful if synthetic data; the prototype is not as ideal as simulated, antennas are physical ones (like UWB ridged ones that one uses), which requires careful calibration, and they do not yield scattered fields as commonly assumed in the synthetic situations in the literature (and herein as well), while training data are costly to get as succession of experiments needed. Both cases (synthetic and experiments data) will illustrate a CNN solution and a RNN solution with aim to yield some sort of superresolution, with comparison to a standard time reversal approach.