Machine Learning to Predict EMI Radiation and Optimize the Structure of Pinmap Packages

Abstract

With the rapid increase in integrated circuit chip frequencies, the complexity of electromagnetic interference issues and the time-consuming nature of traditional pin mapping (pinmap) design methods have become increasingly evident. This article leverages advanced machine learning techniques to accurately and efficiently predict the maximum 3-m radiated electric field of pinmap packages. Among the various models tested, the convolutional neural network (CNN) demonstrated the best performance. When combined with the Adam optimizer, the CNN achieved an average relative error of less than 1.5% across the 6 to 20 GHz frequency range. The trained CNN model’s prediction speed is several orders of magnitude faster than full-wave simulation methods. Furthermore, this article proposes a structural optimization method to minimize radiation in pinmap packages. By integrating ground ball position optimization with the trained CNN model, the method achieves optimal radiation suppression. Validation across three different cases demonstrated a reduction in radiation by 4.85 to 10.37 dB (mV/m), confirming the effectiveness and applicability of the proposed approach.