ISFNN: an enhanced neural network for parametric modeling of passive devices with input skip-connections

High-performance passive devices play a critical role in the radio frequency front-end of wireless systems. Accurately characterizing the electromagnetic (EM) responses of these devices poses a formidable challenge, particularly in high-frequency design. Commercial numerical methods often demand substantial computational resources and necessitate complete recalculation for any structural modifications. This paper proposes an input skip-connections feedforward neural network (ISFNN) for the parametric modeling of passive devices. The ISFNN architecture incorporates multiple skip connections within the layer blocks, which periodically integrate the input design variables into the intermediate hidden layers. This design facilitates feature combination and enhances the extraction capability of design variables. The intermediate layers contain both the original input features and the learned feature information from preceding layers, enabling the model to effectively and robustly capture the nonlinear relationships between the design variables and EM responses. Additionally, a systematic algorithm is proposed to develop and train the ISFNN model. The ISFNN offers a unified solution for both single-physics EM analysis and EM-centric multi-physics (MP) analysis. Compared to other ANN-based models, the ISFNN achieves smaller testing errors with fewer training samples for single-physics EM modeling. Furthermore, in certain applications, conducting MP analysis is more aligned with the actual operating conditions of high-performance microwave components. The nongeometrical parameters are incorporated into the input design features. The ISFNN accurately predicts EM responses using only MP training data, without requiring additional single-physics EM data, transfer functions, or multiple mapping modules. Validation examples demonstrate the better modeling accuracy and efficiency of the ISFNN. Within the design space, the trained ISFNN can generate new MP data in just 0.02 seconds per calculation, significantly reducing time costs while maintaining high modeling precision. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.