Intelligent Optimization of RDL-TSV Interconnect Structures Using Physics-Guided CNN and Multi-Objective GA

High-frequency transmission loss in Redistribution Layer-Through Silicon Via (RDL-TSV) interconnect structures is a critical factor influencing the performance of three-dimensional integrated circuits. This study aims to enhance the prediction accuracy of high-frequency losses by balancing the training accuracy and computational efficiency of traditional full-wave simulation and equivalent circuit models. A Physical Information Convolutional Neural Network (PI-CNN) prediction model was developed based on convolutional neural networks, incorporating the skin effect as physical guidance. A multi-criteria decision-making framework was then proposed by integrating the PI-CNN model with a genetic algorithm. Results show that the PI-CNN model achieves stable single-prediction times under 3 s, with prediction loss errors below 0.1 dB and an R2 value of 0.987, significantly improving the accuracy of high-frequency loss prediction. Through multi-criteria decision optimization, the randomness inherent in genetic algorithms enables systematic exploration of favorable design options within the design space. This approach ensures that the final design maintains consistent performance and robustness under anticipated manufacturing variations. The study provides a data-driven, physics-guided approach for evaluating and optimizing high-frequency performance in advanced packaging.