Experimental Evaluation of a Deep Learning Approach to Transmissive Metasurface Design for Mask-Based Power Pattern Synthesis

An end-to-end deep-learning-based approach is considered for the design of lossless and passive transmissive metasurfaces that transform known incident electromagnetic waves into new ones whose far-field power patterns reside within user-defined lower and upper masks. The far-field pattern obtained from this design approach is evaluated using two simulation strategies relying on zero-thickness and finite-thickness metasurface models, as well as experimental measurements performed in a planar near-field antenna range. It is shown that the adherence of the achievable power pattern to the desired far-field masks deteriorates as we transition from the simulated zero-thickness model to the simulated finite-thickness model and eventually to the fabricated metasurface that is evaluated experimentally. Finally, the reflectivity of the fabricated metasurface is evaluated using the time gating feature of a vector network analyzer, confirming the low level of reflectivity under normal plane wave illumination for the fabricated transmissive metasurface.