Intelligent Prediction Method for Radiation Emission of Printed Circuit Board by Using Deep Learning

Abstract

This article introduces a novel deep learning approach, utilizing a seq2seq model, to predict far-field radiation emissions (REs) in high-density, multilayer, and complex printed circuit boards (PCBs). To address the challenges of weak long-term dependencies and information ambiguity arising from excessively long network sequences in RE prediction for complex PCBs, our method begins by tokenizing and modeling each network combination. These combinations are then fed into the seq2seq model for training, supplemented with a multihead attention mechanism to enhance information connectivity. To validate the model's accuracy and efficiency, we compared our prediction results with those obtained from traditional full-wave electromagnetic simulations. Our approach significantly reduces the RE prediction time for the entire board from several hours to nearly one second. Moreover, the model exhibits high performance in RE prediction, with a mean absolute error of approximately 0.001 V/m. To verify the model's reliability, we applied the trained model to predict RE for brand-new entire boards and key networks, demonstrating its ability to accurately predict RE in previously unseen PCBs. Therefore, this method can be implemented for intelligent, full-automatic RE prediction in practical applications.