An Encoder-Only Transformer to Generate Power Patterns from Far-Field Performance Criteria

This paper proposes a method for generation of Maxwellian far-field power patterns, to be used as input for a variety of electromagnetic inversion algorithms. Typically inversion source and scattering algorithms require a desired power pattern to reconstruct equivalent currents and/or material properties, but it is more common for the designer to provide only far-field pattern specifications such as beam directions, null directions, or half-power beamwidth. By using a method that represents an aperture antenna with a grid of infinitesimal dipoles, a set of training data is created consisting of current distributions (with varying phase and amplitude) and their corresponding power patterns as well as power pattern specifications. An encoder-only transformer neural-network model is then trained to recreate the full power pattern given only the specifications.