A Princeton University research project suggests that future RF components may no longer be designed by refining existing geometries, but by defining the desired performance first and allowing AI-assisted algorithms to discover entirely new solutions.
For decades, progress in RF and microwave engineering has been driven by better semiconductor technologies, more advanced manufacturing processes, increasingly accurate electromagnetic simulations and ever-growing computing power. Yet despite these advances, the fundamental design methodology has remained largely unchanged.
Engineers typically begin with a known circuit topology or physical geometry, simulate its behaviour, analyse the results and gradually refine the design through multiple iterations until it meets the required specifications.
A research programme led by Professor Kaushik Sengupta at Princeton University, recently highlighted by IEEE Spectrum, challenges this long-established workflow. Rather than starting with the physical structure of a component, the researchers propose reversing the process: first define the desired electrical performance, then allow computational algorithms to search for the geometry capable of achieving it.
Known as Inverse Design, the approach represents more than another optimisation technique. It questions one of the basic assumptions behind RF engineering: that the physical structure must always be the starting point of the design process.
As RF systems continue to grow in complexity, this seemingly subtle change could have profound implications.
From Geometry to Performance
Traditional RF development follows a familiar path.
Engineers select an architecture, define an initial layout, perform electromagnetic simulations, evaluate the results and repeat the process—sometimes dozens or even hundreds of times.
The methodology has produced generations of successful amplifiers, filters, antennas and RF integrated circuits. However, it also assumes that designers already have a reasonable idea of what the final structure should look like.
Inverse Design takes a different perspective.
Instead of asking how to improve an existing design, engineers specify the desired operating frequency, bandwidth, efficiency, output power, manufacturing constraints, size limitations and any additional performance requirements. The algorithm then searches for geometries capable of satisfying those constraints, without being limited by conventional design intuition.
In other words, the computer is not refining an existing concept—it is searching for entirely new ones.
Why Is This Becoming Practical Now?
The idea of Inverse Design is not new.
Researchers have explored it for years, but its practical application was constrained by one unavoidable obstacle: computational cost.
Every candidate RF geometry requires a full electromagnetic simulation based on Maxwell’s equations. For complex RFICs or mmWave structures, a single simulation may take minutes or even hours. Searching through hundreds of thousands of possible geometries quickly becomes impractical.
Recent advances in machine learning have changed that equation.
Rather than performing a full electromagnetic simulation for every candidate, researchers first train surrogate models using large datasets of highly accurate simulations. Once trained, these models can estimate the behaviour of new geometries in fractions of a second.
Only the most promising candidates are then passed to full-wave electromagnetic simulation for final verification.
The result is not a replacement for traditional simulation tools, but a dramatic reduction in the search space—making Inverse Design practical for increasingly complex RF problems.
When the Best Design Doesn’t Look Like Engineering
Perhaps the most fascinating aspect of the Princeton research is not the computational efficiency, but the type of solutions it discovers.
Conventional optimisation algorithms usually improve existing layouts. They adjust dimensions, tune parameters and refine structures that already resemble familiar RF circuits.
Inverse Design removes those assumptions.
Any geometry capable of meeting the required specifications becomes a valid candidate—even if it bears little resemblance to anything an engineer would naturally draw.
One of the most striking demonstrations presented by Professor Sengupta involves a silicon power amplifier operating between 30 and 100 GHz.
Rather than optimising an existing amplifier, the researchers simply defined the required electrical performance and allowed the algorithm to search for a suitable layout.
The resulting structure looked remarkably unconventional.
Instead of the clean, symmetrical geometries common in RF design, the final layout appeared as a dense, irregular pattern that IEEE Spectrum compared to a QR code.
More importantly, the design worked.
After full electromagnetic verification, fabrication and laboratory testing, the amplifier demonstrated an impressive combination of bandwidth, output power and efficiency, providing one of the first practical demonstrations that AI-assisted inverse design can generate manufacturable RF hardware—not merely interesting theoretical concepts.
Beyond Artificial Intelligence
Although artificial intelligence receives much of the attention, the Princeton work combines several computational techniques rather than relying on a single AI model.
Surrogate models reduce the number of expensive electromagnetic simulations.
Reinforcement learning continuously improves the search strategy.
The research team is also exploring diffusion models capable of generating entirely new RF geometries directly from electrical specifications such as S-parameters.
Yet regardless of how the candidate structures are created, every promising design ultimately returns to the same destination: rigorous full-wave electromagnetic simulation.
Physics remains the final authority.
Only the path towards discovering new solutions is changing.
What Could This Mean for RF Engineers?
It is far too early to suggest that Inverse Design will replace conventional RF design workflows.
Many challenges remain, including manufacturing constraints, integration into existing EDA environments and validation across a much broader range of applications.
However, the research already points towards an important shift in engineering thinking.
For decades, the central question has been:
“How can we improve the geometry we already have?”
Inverse Design asks something different:
“Are we even starting from the right geometry?”
If approaches like this continue to mature, computational models may increasingly take responsibility for exploring vast design spaces, while engineers focus on defining system requirements, engineering constraints and evaluating the most promising solutions.
The engineer’s expertise does not become less important.
Instead, its focus moves from manually searching for solutions to defining better problems to solve.
Looking Ahead
Whether Inverse Design becomes a standard engineering tool remains to be seen.
Like every new design methodology, it will need to prove itself across a broad range of technologies before it becomes part of mainstream RF development.
Nevertheless, Professor Kaushik Sengupta’s research has already achieved something significant.
It has reopened a question that RF engineers have rarely asked over the past several decades:
Does every RF design really have to begin with geometry?
The answer may still be “yes” in many applications.
But if, in some cases, the answer turns out to be “no”, Inverse Design could represent one of the most significant methodological shifts RF engineering has seen in a generation.
