Deep learning approach to wireless chip design wins best paper at International Microwave Symposium

Written by
Scott Lyon
Sept. 14, 2022

Graduate students from Princeton ECE have won the Best Advanced Practice Paper Award at the 2022 IEEE International Microwave Symposium, the flagship conference on radio-frequency and wireless technologies.

The students, Zheng Liu and Emir Ali Karahan, developed a method that uses deep learning to massively speed up the design of RF chips for next-generation wireless systems. The traditional approach relies heavily on intuition, making it labor-intensive and expensive. The new approach exploits data-driven methods to virtually eliminate the need for highly specialized expertise and slashing the most time-consuming aspects of the process. The paper was selected as the top submission among hundreds, including all student and industry submissions related to practical applications.

Sengupta poses in his lab

Kaushik Sengupta in his lab. Photo by Sameer A. Khan/Fotobuddy

Advised by Kaushik Sengupta, associate professor of electrical and computer engineering, Liu, a fifth-year Ph.D. student, and Karahan, a fourth-year Ph.D. student, inverted the conventional design approach. They used the chip’s desired performance as their starting point, and the complex electromagnetic structures emerged through a fast optimization algorithm. Prior attempts to automate this design process were limited by time and resource-intensive simulations.

In the winning paper, the authors showed that those simulations can be eliminated by a deep convolutional neural network (CNN) model that predicts the performance of each complex structure from a single image of that structure. This opens a new, searchable set of designs with an astronomical number of total configurations.

Liu said that while they can’t guarantee the globally optimal solution, the new approach rapidly uncovers many near-optimal solutions that defy intuition and outperform state of the art designs. Sengupta added that this new method can also be generalized to include designs for antennas, filters and end-to-end RF chips.

Their paper, “Deep Learning-Enabled Inverse Design of 30–94 GHz Psat,3 dB SiGe PA Supporting Concurrent Multiband Operation at Multi-Gb/s,” was supported in part by the Defense Advanced Research Program Agency and the Army Research Office.