Knowledge-Guided Active-Base-Element Modeling in Machine-Learning-Assisted Antenna-Array Design

Inverse design of reflective metasurface antennas using deep learning from small‐scale statistically random pico‐cells

“…By establishing a mapping between the geometry and scattering coefficients of known unit cells, the trained network can be used for predicting the characteristics of unknown geometries automatically. [19][20][21][22][23][24][25][26][27][28] The success or accuracy of mapping highly depends on the quality of the training data, for which a large number of training data is usually needed, known as the big-data-based method.…”

“…For this purpose, prior knowledge is introduced for selecting the training data. [23][24][25][26][27][28] For example, a pretrained transfer learning (TL) network that includes uncorrelated or domain-unspecific experience, the size of training samples is reduced to 20 000, 23 where the actual candidate pool or the number of simulations performed is expected to be larger for the 20 000 training samples to be randomly selected and form an identical distribution as required. In another example, correlated or domain-specific prior knowledge is introduced in combination with the variational auto-encoder (VAE) network for the design of transmittarrays.…”

“…The key challenge is how to select a small‐scale but representative or diverse enough data for being used in a following training process. For this purpose, prior knowledge is introduced for selecting the training data 23–28 . For example, a pretrained transfer learning (TL) network that includes uncorrelated or domain‐unspecific experience, the size of training samples is reduced to 20 000, 23 where the actual candidate pool or the number of simulations performed is expected to be larger for the 20 000 training samples to be randomly selected and form an identical distribution as required.…”

“…Recently, the method of deep learning (DL) has been introduced for accelerated inverse design of MAs, where a DL network is trained by a large set of unit cells including their geometric information and electromagnetic characteristics (or data samples). By establishing a mapping between the geometry and scattering coefficients of known unit cells, the trained network can be used for predicting the characteristics of unknown geometries automatically 19–28 . The success or accuracy of mapping highly depends on the quality of the training data, for which a large number of training data is usually needed, known as the big‐data‐based method.…”

Sparse linear array synthesis using artificial neural network and nonconvex optimization

“…To apply ANNs to sparse array and thinned array synthesis, existing approaches often combine ANN with other optimization algorithms. 19,21 ANNs are employed for forward prediction of the pattern, while optimization algorithms handle the distribution and excitation of the elements. For example, ANNs establish the relationship between subarray element positions and the resulting pattern, accounting for mutual coupling effects.…”