Radiation Source Reconstruction for TSV-Based Chips and Multilayer Circuits via Optimized Neural Networks

In this article, a radiation source reconstruction method (SRM) is proposed for through silicon via (TSV)-based chips and multilayer packaged circuits. By integrating various neural networks (NNs) with advanced optimization algorithms, the proposed method achieves a measurable accuracy improvement in electromagnetic field prediction, as validated by experimental data. Specifically, the method employs enhanced NNs to accurately fit the dipole moments generated by the chips or circuits. These fitting results are then utilized to construct an equivalent dipole model, which serves as a reliable substitute for the actual radiation sources during their operation, particularly in TSV-based structures. Numerical simulations and experimental measurements demonstrate that the prediction error of the proposed three-dimensional (3D) SRM is as low as 4.74%, representing an absolute error reduction of 2.5–3.8 percentage points and a relative reduction of 34% –44% compared to the traditional regularization method. Moreover, the proposed SRM exhibits effective performance in characterizing the electromagnetic compatibility issues of complex 3D TSV chips, as evidenced by stable low-error results across different dipole configurations.