AI-Based Transformer Matching Network Synthesis Using a Data-Efficient Two-Step Method

This work presents an artificial intelligence-driven approach for automated on-chip impedance-matching (IM) transformer design, achieving layout generation within seconds. Unlike direct prediction approaches, where machine learning models directly estimate the optimal transformer geometry based on input and output impedances, this method follows a two-step process. This new approach significantly reduces the required training data for the neural network while maintaining high accuracy and efficiency. As a proof of concept, the synthesis of a D-band transformer in IHP's SG13G2 technology is presented and optimized using the algorithm proposed in this paper. The approach is versatile and applicable across technologies, frequency ranges, and matching network structures, making it a scalable solution for automated IM design.