Integrating Transfer Learning and GPU Acceleration in MMIC Design: A Neural Network Approach for a 6–13 GHz LNA

Abstract

The design of monolithic microwave integrated circuits (MMICs) is a laborious process that involves exploring a vast design space, requiring multiple iterations to identify the optimal circuit design. In this research, we propose a design approach that combines GPU-based high-performance computing and transfer learning techniques. To improve modularity and reusability, we decompose the MMIC into multiple substructures and then combine these substructures to restore the overall circuit structure and performance. To achieve this, we adopted schematic simulation, which is more time-efficient, to construct a data set and pre-train the circuit substructure models. We then fine-tune the pre-trained models using a limited amount of electromagnetic (EM) simulation data, aiming to obtain layout-level subcircuit models. Leveraging the parallel processing capabilities of neural network models, we employ GPU to conduct extensive exploration and design within the circuit design space, utilizing cascade connection theory to optimize the performance of the complete circuit. We apply this methodology to a low-noise amplifier (LNA) circuit operating in the 6–13 GHz frequency range, achieving favorable outcomes.