Rapid Layout Optimization of Active RF Circuit with Deep Learning: from Rectifiers to Power Amplifiers

This paper proposes a deep learning-powered method with transfer learning capability to accelerate RF active circuit design. A schematic-level surrogate model for broadband matching is constructed through a fully connected neural network at extremely low computational cost, which is then transferred to its layout-level counterpart with minimal electromagnetic simulation data. Innovatively, the surrogate model is integrated into the SPICE simulation environment for comprehensive evaluation of active circuit performance. To validate the methodology, a 1-4 GHz broadband rectifier and a 2-4 GHz broadband power amplifier are implemented. Measurement results show good agreement with predictions and simulations, demonstrating average efficiencies of ≥ 65% and 60% within their respective operational bandwidths. The proposed approach significantly reduces labor costs while shortening the design cycle from several days to minutes.