Inverse Design of Broadband Optimal Power Amplifiers Enabled by Deep Learning

An inverse design method for broadband power amplifiers (PAs) is presented in this article that does not rely on the designer’s expertise and can directly synthesize the optimal layout of the output matching network (OMN). A highly randomized patterned microstrip (PM) is employed as the OMN, which can significantly expand the design space. A high-speed electromagnetic (EM) simulator is implemented through deep learning (DL) to accomplish highly accurate prediction of S-parameters. Based on the fast responsiveness of DL, two inverse design methods are proposed. First, inverse synthesis is accomplished using a mutation-driven genetic algorithm (MDGA), requiring only a small amount of target impedance. The OMN synthesis for high electron mobility transistors (HEMTs) across different sizes is achieved, demonstrating the PM’s robust capability in impedance characterization. Second, a novel matching technique incorporating power flatness considerations is proposed, built upon the MDGA. The OMN’s optimization is guided by the constraints formed by load–pull contours, resulting in a synchronous increase in the output power of the PA over broadband, thereby ensuring excellent flatness. As a verification of the method, a 6–18 GHz PA is designed and fabricated in a 0.25- $\mu $ m gallium nitride (GaN) technology. The measurement results show that, within the operating frequency range, the PA achieves an output power of 34.4–36.2 dBm with a flatness of 1.8 dB, and a power added efficiency (PAE) ranging from 20% to 31%, averaging 26.3%.