Automated Millimeter-wave RFIC Inductor Design using Neural Networks

Abstract

Since 5G communication systems have been commercialized recently, millimeter-wave circuit technology has been expected to play an important role in the information and communication technology (ICT) industry. In millimeter-wave bands, semiconductor design parameters and models are more inaccurate than in micro-wave bands, so the design of the millimeter-wave radio-frequency (RF) circuit requires a lot of experience and time to overcome this inaccuracy. Recent millimeter-wave RF circuits primarily employ inductor-based elements that have an outstanding effect on size and performance. However, unformatted layouts and complex parasitic components of the elements require considerable time and effort to design the elements throughout the whole design process. We propose automated millimeter-wave RF circuit design using neural networks that learned by supervised learning. The proposed system receives electrical characteristics such as inductance, quality factor and impedance, and outputs layout parameters such as width and length in very short time. When the proposed system was tested 10,000 times in the range of operating frequency 1 to 100 GHz and inductance 50 to 500 pH, the root mean square of inductance error rate was 2.81 % and the average design time was 0.039 seconds. This study proposes a technical basis that can ultimately contribute to the design of RF circuits by providing users with the automated system as a part of PDK provided by the semiconductor foundry.