Design of Aperture-Multiplexing Metasurfaces via Back-Propagation Neural Network: Independent Control of Orthogonally-Polarized Waves

Representation Learning-Driven Fully Automated Framework for the Inverse Design of Frequency-Selective Surfaces

“…1) compared with [12], [13], [18], [19], we establish a mapping from EM constraints to the topology candidates to enable auto-selection of the suitable topology, which can be accomplished without experienced designers; 2) one advantage over [20]- [22] is that we establish a cross-evolution system to customize the topology and evolve it into an optimal design instead of building a dictionary mapping in advance through tedious training, hence we take only hundreds of simulation cycles while they took tens of thousands of simulation cycles to collect data; 3) the other advantage is, unlike [20]- [22] that obtained the optimal design by optimizing several candidate geometries, the auto-evolution system can evolve the auto-selected topology into new geometries that satisfy various constraints, thanks to the proposed fan-based representation defined by only tens of parameters instead of the pixelated metal layer defined by hundreds of parameters [20]. We validate the proposed framework in four cases where different EM constraints were provided.…”

“…Fanbased representation reaches a balance between increasing possibilities and reducing complexity. Unlike the traditional method [18] that swept the parameters of a specific geometry, fan-based representation creates new geometries and provides more possibilities and functionalities. Compared with [19] that pixelated the metal layer and involved hundreds of parameters, fan-based representation creates sufficient structures by utilizing only tens of parameters, hence training complexity and the amount of training data are reduced significantly.…”

“…For validation, we pre-defined the constraints as a band-pass behavior from 6 GHz to 7.5 GHz, an insert loss less than 2 dB, as expressed in (18):…”

“…Xiao et al [17] proposed an inverse learning system to predict geometry variables of selected metasurfaces for a desired EM behavior without further optimization, which consists of a data classification technique and a cascade of two inverse learning machines for enhanced learning performance. Zhu et al [18] artificially built up a modified Jerusalem Cross structure at first; afterward, they utilized a three-layer back-propagation neural network to project the reflection phase to the length of the metal arm. Likewise, Wei et al [19] targeted the appropriate topology in advance through analysis of filter equivalent circuit before applying the genetic algorithm (GA) to reach an optimal design.…”

“…The continuity of mapping determines the degree of automation. The works in [12], [13], [18], [19] achieved only semi-automation because their mappings did not start from the constraints, which is the initial state of inverse design. The compactness of the mapping impacts the difficulty to establish and the overall efficiency.…”