AI-Enabled Radio-Frequency Chip Design

Minutes

On October 23, 2026, Members of the Society and guests joined the speaker for a reception and dinner at 5:45 PM in the Members’ Dining Room at the Cosmos Club. Thereafter they joined other attendees in the Powell Auditorium for the lecture proceedings. In the Powell Auditorium of the Cosmos Club in Washington, D.C., President Larry Millstein called the lecture portion of the 2,543rd meeting of the Society to order at 8:02 p.m. ET. He began by welcoming attendees, thanking sponsors for their support, announcing new members, and inviting guests to join the society. Scott Mathews then read the minutes of the previous meeting which included the lecture by Emanuele Berti, titled “Black Hole Spectroscopy: Listening to a Black Hole’s Gravitational-Wave Ringing”. The minutes were approved as read.

President Millstein then introduced the speaker for the evening, Kaushik Sengupta, of Princeton University. His lecture was titled "AI Enabled Radio Frequency Chip Design."

The speaker began by tracing the deep historical roots linking the Philosophical Society of Washington and Princeton University to the origins of telecommunications. He mentioned Alexander Graham Bell’s demonstration of the working telephone before the Society in 1877, noting the presence that evening of Joseph Henry. He said that Henry made seminal breakthroughs in electromagnetism decades before Maxwell and Hertz formulated the underlying theory. Sengupta noted that chip-scale iterations of Henry’s solenoids remain ubiquitous, embedded in billions of modern radio-frequency integrated circuits (RFICs).

Sengupta discussed the historical beginnings of wireless communication; including contributions from Joseph Henry, Heinrich Hertz, J.C. Bose, Guglielmo Marconi, Edwin Armstrong, and Jack Kilby- leading up to Nokia’s 1992 release of the world’s first mass produced digital mobile phone. He said that today’s commercial cellular networks transmit approximately 210 exabytes of data per month, and that modern mobile receivers detect signals as small as 100 femtowatts. Sengupta explained that this revolution rests upon the dual foundations of Maxwell’s equations and Moore’s law of integrated circuit scaling. He said that transistor dimensions have contracted from 10’s of microns in the early 1970’s to 10’s of nanometers in modern IC’s. Sengupta claimed that gate costs plummeted from $1,000 per gate in 1959 to less than a dollar for 300 million transistors on a single silicon die. He said that this evolution fulfilled Gordon Moore’s 1965 prediction that integrated electronics would revolutionize phased array radar.

The speaker detailed the physical mechanisms of high-frequency RFICs and phased-array sensing. He showed how performance breakthroughs resulted from advanced multi-sensor integration, particularly 77 GHz radar transceivers. He said that these RFIC’s are operating at frequencies 15 to 20 times higher than computer CPU’s. Sengupta explained that phased arrays emulate biological binaural localization. He described multiple integrated antenna elements radiating coherently, with electronically controlled delay lines steering the electromagnetic beams without moving parts.

Sengupta argued that despite decades of progress, conventional RFIC design remains fundamentally constrained by bottom-up, human-intuitive methodologies. Like civil engineers assembling trusses from discrete chords and joints or optical engineers arranging standard lenses, RF designers rely on finite libraries of well-characterized discrete elements, including: inductors, capacitors, transistors, transmission lines, and patch antennas. He argued that these fundamental building blocks do not necessarily constitute the most general or most efficient basis for exploring the complete RF design space.

To illustrate an alternative paradigm, the speaker drew an analogy to the structural design of the Sagrada Família in Barcelona. Gaudí, the architect, built a hanging-string model suspending weights, which was subsequently inverted into freestanding stone arches. Sengupta proposed developing an analogous "hanging chain model" for RF circuits: an AI-driven inverse design framework that bypasses human-designed templates, directly synthesizing optimal structures from desired electromagnetic performance.

The speaker explained the computational challenges of inverse electromagnetic synthesis. Discretizing a passive metallic surface into a 25-by-25 binary grid (a “QR-code”) yields approximately 10188 possible configurations, exceeding the number of atoms in the universe, and making brute-force optimization computationally intractable. To resolve this, Sengupta’s laboratory developed a two-tiered framework. First, they constructed AI-driven “forward models” that predict full-wave electromagnetic responses in microseconds; as compared to computing numeric solutions to Maxwell’s equations, which would take minutes or even 10’s of minutes. Second, they implemented generative diffusion models which “evolve complex EM structures” from stochastic noise. The generative diffusion models optimize the “QR-code” of metallization such that the desired EM properties are achieved. Because multiple distinct geometries can generate identical EM responses, the researchers integrated a control mechanism allowing designers to push the diffusion process along a spectrum ranging from classical topologies to highly non-intuitive pixelated geometries.

Sengupta presented experimental validations of this methodology. He showed several examples of RFIC’s designed using this method, including: ultra-wide bandwidth LNA’s (Low Noise Amplifiers), dual-band LNA’s, phased array feed structures, dual-band filtering networks, and millimeter wave power amplifiers. The speaker claimed that the automated inverse synthesis produced multi-band satellite amplifier designs in under five minutes, compared to 21 days for conventional optimization.

The speaker addressed prevailing skepticism regarding AI-driven circuit design, rejecting the idea that analog and high-frequency electromagnetics are too reliant on human insight to be automated. Sengupta concluded by reiterating the point that “ground-up” design limits the accessible region of the complete design space, and that there are opportunities outside this limited region.

The lecture was followed by a question and answer session.

A member asked about “spectrum sharing”, asking whether Sengupta’s approach could address some of the issues associated with spectrum shortages. Sengupta responded that because spectrum sharing problems inherently involve hardware filtering, he thinks that this approach is clearly relevant. He noted, however, that hardware solutions “must scale-up across other levels of abstraction”.

A guest asked about the asymmetric design of a particular RFIC that Sengupta presented, asking “Can you describe how the asymmetry makes the design more efficient”. Sengupta responded that the asymmetric design relates to power combiners, where a signal is split, amplified, and recombined. He said that the asymmetric splitting allows a more favorable loss/bandwidth trade-off than the symmetric configuration.

A member on the livestream asked if human designers had thought of splitting signals into separate frequency bands and processing through separate pathways on the chip. Sengupta replied that this approach had been used, in what was called a classical frequency diplexing structure, but that the challenge of impedance matching the passive networks over a wide range of frequencies made it “very hard to do”.

After the question and answer period, President Millstein thanked the speaker and presented her with a PSW Rosette, a signed copy of the announcement of the lecture, and a signed copy of Volume 17 of the PSW Bulletin. He then announced upcoming lectures and adjourned the 2,543rd meeting of the Society at 9:56 p.m. Eastern Time.

Temperature in Washington, D.C.: 11.7 °C
Weather: Passing Clouds

Dinner Attendance: 40
Lecture Attendance (In-Person): 66
Livestream Attendance: 36
For a Total of 102 viewers
Views of the video in the first two weeks: 863

Respectfully submitted, Scott Mathews, Recording Secretary