Reinforcement-Learning-Based Optimization of Bonding Wires for EMI Mitigation

Abstract

Wire bonding as a metallic interconnection is widely used to transmit high-speed signals and supply power within the integrated circuit (IC) packages. However, bonding wires also effectively radiate power noise and the harmonics of the output signals, causing electromagnetic interference and radio frequency interference issues. In this study, a current-loop model using a transfer admittance matrix for estimating the equivalent radiation sources of an IC/package featuring bonding wires is proposed. Based on the proposed modeling method, a novel reinforcement learning algorithm is applied to optimize the configurations of signal, power, and ground bonding wires, mitigating the radiation from the IC/package. The proposed modeling method is validated experimentally by a self-designed IC with an inverter-type buffer based on a complementary metal–oxide–semiconductor 0.18-μm process, and a radio frequency victim antenna built on the same printed circuit board. From 720 to 900 MHz, the maximum difference between the proposed modeling method and the measurement results is only 2.3 dB. In addition, full-wave simulation is performed to evaluate the optimization results of the reinforcement learning algorithm, showing radiation mitigation of over 7 dB compared to the randomly selected bonding-wire configurations.