1-D multi-channel CNN with transfer functions for inverse electromagnetic behaviors modeling and design optimization of high-dimensional filters.

As an essential passive component in modern wireless communication systems, the design of high-frequency filters has become increasingly crucial. To achieve the target behavior specifications, traditional design methods are constrained by designers' expertise or reliant on repetitive frequency sweeps using commercial software. Such processes suffer from low efficiency, limited applicability, and high computational costs. Artificial neural network-based modeling has become an important tool for designing devices. To realize accurate and fast electromagnetic modeling and design of passive components, this work proposes an inverse model integrating transfer functions and one-dimensional multi-channel convolutional neural networks (TF-1DMC-CNN). This model introduces transfer functions to ensure precise representation of electromagnetic responses while addressing the challenge of input dimensionality in wideband modeling. Input dimensions are reduced from 161 to 20 and 221 to 20 for two examples. The 1DMC-CNN processes distinct TF coefficients in each channel and extracts features in parallel. The geometrical parameters can be directly predicted in a single feedforward pass through the trained inverse model without needing iterative optimization. Compared to other inverse neural networks, the proposed model achieves the smallest testing errors. It obtains better model accuracy with fewer training samples, reducing data generation time. Compared to the traditional EM optimization method, this approach reduces CPU time for optimizations, enabling predictions of geometric structures that meet different design indexes. For multi-objective optimization, the proposed model predicts the structure within 0.16 seconds. [ABSTRACT FROM AUTHOR]