A Deep-Learning Approach for Wideband Design of 3D TSV-Based Inductors

Abstract

A high-efficient wideband through-silicon vias (TSVs) modeling method based on deep learning is proposed, and a compact three-dimensional (3D) spiral inductor is designed using the proposed method. By comparing different activation functions and loss functions, an adaptive deep neural network (DNN) based on Gaussian Error Linear Unit (GELU) and Huber functions for constructing parameterized TSV models is proposed. The model has much higher accuracy and better robustness than commonly used circuit equivalent models over a wide range of bandwidths. Moreover, a compact 3D spiral inductor using ground TSV is designed based on the DNN model. This 3D inductor greatly reduces the inductor area compared to planar inductors and has weak crosstalk between TSV pairs. The designed inductor is simulated by direct electromagnetic calculation to verify the proposed method and design.

Keywords

• Deep learning
• equivalent-circuit model
• through-silicon vias (TSVs)
• 3D spiral inductor