EMI Prediction and Optimization for Pinmap Design Using Deep Transfer Learning and an Enhanced Genetic Algorithm

With the rapid increase in the operating frequency and integration density of ball grid array packages, pin assignment (pinmap) significantly impacts electromagnetic interference (EMI). However, the previous deep reinforcement learning (DRL) approaches required time-consuming evaluation and training procedures. In this article, we propose a novel design methodology for predicting and optimizing the EMI of pinmaps. First, we present a deep learning-based predictor that can accurately and quickly evaluate the EMI levels of pinmaps, thereby supporting fast pinmap design. Furthermore, transfer learning achieves excellent predictor performance with less training data, resulting in effective data savings. Based on the presented predictors, an enhanced genetic algorithm is developed to optimize the EMI and can quickly find better solutions compared with the DRL approaches. As a result, this article proposes a detailed guideline for predicting and optimizing the EMI of pinmaps, and the proposed methodology can be further developed for promising intelligent system-level packaging design.