When Algorithms Out-Designed Engineers: The Antenna That Changed What Was Possible

You are reading this on your phone right now. Every text, every image, every video frame that has crossed your screen in the last hour likely passed through an antenna. Most people never think about these invisible workhorses of modern life — the thin strips of metal etched onto circuit boards, smaller than a fingernail, doing the unglamorous job of converting electrical signals into the radio waves that connect us all.

But antennas have a dirty secret: they're incredibly finicky. Change the frequency, the material, the shape by a hair, and the whole system can fall apart. For decades, designing an antenna meant relying on the intuition of experienced engineers, running endless physical prototypes, and making painstaking calculations by hand. It was slow, expensive, and limited what was possible.

A team of researchers from Umeå University, Karlstad University, and Proant AB in Sweden has just demonstrated something remarkable. They used an algorithm — a mathematical optimization process run on supercomputers — to design a dual-polarized microstrip antenna that achieves nearly 10% bandwidth. To understand why this matters, consider that most conventional single-layer designs manage only 1.8% to 3.1%. One design, fabricated on cheap fiberglass circuit board material (FR4, the same substrate used in everyday electronics), matched its computer simulation almost perfectly when it was finally built and tested in the real world. The algorithm had out-thought human engineers. And it did so by treating the antenna not as a component to be engineered, but as a material distribution problem to be solved.

The Science

The paper, published by researchers Pan Lu, Eddie Wadbro, Viktor Lundström, Jonas Starck, Martin Berggren, and Emadeldeen Hassan, attacks a fundamental problem in wireless communications: how do you design an antenna that can send and receive data on two different polarizations simultaneously, while also covering a wider range of frequencies?

Let me back up and explain what that means.

Radio waves are electromagnetic oscillations that travel through space. They have a property called polarization — essentially the orientation of their electric field. A wave polarized vertically won't interact as strongly with a horizontal antenna, and vice versa. Modern wireless systems exploit this. By sending separate data streams on different polarizations, you can effectively double the capacity of a communication channel without using more spectrum or power. This is called dual-polarized (DP) operation, and it's essential for everything from cellular base stations to satellite links.

The challenge is that most dual-polarized antennas are designed by imposing symmetry on a single radiating patch — essentially cutting a metal square in a specific way and hoping it works. The approach is intuitive, based on decades of accumulated knowledge, but it hits hard limits. These antennas typically operate over a very narrow band of frequencies, which makes them fragile. In the real world, signals bounce off buildings, shift with temperature, and vary with humidity. An antenna that only works perfectly at exactly 5.7 GHz is of limited use when the real world is messier than a computer model.

The Swedish team proposed something different. Instead of starting with a human-designed shape and tweaking it, they let an optimization algorithm design the antenna from scratch. This is called topology optimization — a computational technique borrowed from mechanical engineering and structural design, where it's used to figure out where to add or remove material from a bridge, an airplane wing, or a heat sink to make it perform better. The algorithm doesn't know what an antenna is supposed to look like. It only knows what it's supposed to do.

The objective was to simultaneously minimize the reflection coefficient (how much of the signal bounces back instead of being radiated), maximize isolation between the two feeding ports (so each polarization channel stays separate and doesn't interfere with the other), and enhance the bandwidth (the range of frequencies over which the antenna works acceptably). The algorithm was also given constraints: the design must fit within a standard four-layer FR4 printed circuit board stack-up, must be manufacturable, and must enforce a diagonal symmetry that ensures the two polarizations are truly orthogonal.

The researchers used a density-based approach. The design domain — essentially where the copper traces would go on two separate layers of the circuit board — was divided into 250,000 cells per layer. Each cell could be either copper (conducting) or empty (insulating). The optimization algorithm would adjust the "density" of each cell, gradually converging on a distribution of copper that minimized the objective function.

To evaluate each candidate design, the team ran full-wave simulations of Maxwell's equations — the fundamental equations governing all electromagnetic phenomena — using a finite-difference time-domain (FDTD) solver. This is computationally intensive work. Each optimization iteration required four separate simulations: one forward and one adjoint simulation when a port was excited, and one forward and one adjoint simulation with a plane wave source to evaluate receiving performance. The entire optimization process for one design took 210 iterations, running on NVIDIA H100 GPUs at the High Performance Computing Center North.

The adjoint method is crucial here. Calculating how each of the 250,000 design variables affects the final objective would be computationally prohibitive — you'd have to perturb each cell individually and re-simulate. The adjoint method lets you compute all sensitivities (how sensitive the objective is to each design variable) with just two additional simulations. It's a mathematical trick that transforms an intractable problem into a feasible one.

To ensure the final designs were manufacturable, the team used filtering techniques that prevented the algorithm from generating features too small to fabricate. They also thresholded the final result: anything below 50% density became air (insulator), anything above became copper (conductor). This created sharp, clean boundaries between metal and dielectric.

The FR4 stack-up is worth noting. FR4 is the cheap fiberglass-epoxy substrate found in virtually all consumer electronics. It's lossy (it absorbs some radio energy as heat), its electrical properties vary with frequency, and it's generally considered unsuitable for high-performance RF applications. Yet the researchers chose it deliberately. If their optimization approach could make FR4 work, it could make any substrate work.

What They Found

The optimization produced two distinct designs, each the result of 210 iterations of gradient-based optimization using the globally convergent method of moving asymptotes (GCMMA).

Design I connected the feed probe to the middle layer of the four-layer stack-up, leaving the outer layer disconnected. Over 210 iterations, the algorithm settled on a copper distribution that achieved resonance at 5.59 GHz with a reflection coefficient of approximately -14 dB. The -10 dB impedance bandwidth — the frequency range over which at least 90% of the input power is not reflected back — spanned 5.43 GHz to 5.75 GHz, a fractional bandwidth of about 5.8%. More importantly, the reflection coefficient remained below -5 dB over a much wider range, indicating useful performance beyond the strict -10 dB threshold. The isolation between ports stayed below -20 dB across most of the operating band, confirming that the symmetry constraints successfully enforced orthogonal polarization modes.

When the team fabricated Design I and measured it in a lab, the results matched the simulation with striking fidelity. The measured -10 dB bandwidth shifted slightly toward lower frequencies due to fabrication tolerances — a normal consequence of the etching and drilling processes not being perfectly precise — but the bandwidth actually increased slightly to 7.5%. The resonant frequency moved from 5.59 GHz to somewhere in the 5.30-5.71 GHz range.

Design II was more aggressive. Here, the feed probe connected to the outer layer, with the middle layer serving as a parasitic element — not directly fed, but influencing the antenna's behavior through electromagnetic coupling. The optimization algorithm exploited this arrangement to create two distinct resonant modes that overlapped in frequency, effectively doubling the usable bandwidth.

In simulation, Design II showed resonances at 5.67 GHz and 6.00 GHz. The -10 dB bandwidth stretched from 5.59 GHz to 6.09 GHz — substantially wider than Design I. When fabricated and measured, the antenna performed even better than predicted. The two resonances appeared at 5.56 GHz and 5.92 GHz, with a minimum reflection coefficient of approximately -26 dB. The measured bandwidth ran from 5.46 GHz to 6.03 GHz, corresponding to a fractional bandwidth of 9.9% — almost exactly one-tenth of the center frequency.

This 9.9% bandwidth is the paper's most striking number. The comparison table in the paper tells the story clearly. Previous dual-polarized microstrip antenna designs, whether based on traditional theoretical analysis or other optimization algorithms, achieved bandwidths ranging from 1.8% to 3.1% (Chen 2023; Zhu et al. 2022; He and Li 2020; Huang et al. 2025). The new design, using topology optimization on a two-layer stack-up, achieved 9.9% — more than three times the bandwidth of the best previous result.

Bandwidth Comparison with Prior Work

Label	Value
This work (Design II)	9.9
Zhu et al. 2022	3.1
Huang et al. 2025	3.1
He and Li 2020	3
Chen 2023	1.8

Design II achieves 3× the bandwidth of the best previous dual-polarized microstrip antenna design. All designs operate in the 5 GHz range (C-band for 5G/Wi-Fi).

The radiation patterns confirmed that both designs maintained good pattern symmetry and orthogonality. At 5.7 GHz, Design I achieved a maximum realized gain of 6.61 dBi, with the co-polarization (the intended polarization) far stronger than the cross-polarization (the unwanted orthogonal component). Design II showed similar behavior, with the two ports exciting symmetric and orthogonal radiation modes as intended.

The current distributions tell an interesting story. When port 1 was excited, the surface currents on the copper layers were nearly symmetric around the diagonal axis — exactly as the symmetry constraints required. The influence of the feeding probes was minimal, confirming that the structural symmetry, not the feed location, determined the polarization behavior. This is a crucial validation of the design methodology: the algorithm respected the constraints, and the result was predictable, consistent behavior.

Simulated vs. Measured Bandwidth

Label	Value
Design I (Simulated)	5.8
Design I (Measured)	7.5
Design II (Simulated)	9.6
Design II (Measured)	9.9

Simulated and measured results show excellent agreement, validating the optimization-to-fabrication pipeline. Design II achieves nearly 10% fractional bandwidth.

Why This Changes Things

To understand why this matters, you need to understand the bandwidth problem in wireless communications.

Modern wireless systems demand more from antennas than ever before. 5G networks, vehicle-to-vehicle communication, Internet of Things devices, and next-generation radar systems all need antennas that can operate across wider frequency ranges. Wider bandwidth means more data can be transmitted, the system is more resilient to interference and frequency shifts, and it can support multiple communication standards simultaneously without needing separate antennas for each.

Traditional dual-polarized microstrip antennas have been limited by a fundamental trade-off: achieving good dual-polarization performance required imposing symmetry on the design, but symmetry constrained the possible shapes, and those constrained shapes were inherently narrowband. You could have dual-polarization or you could have wideband operation, and getting both required increasingly complex structures that were difficult to design, manufacture, and tune.

Topology optimization breaks this constraint. By allowing the algorithm to explore the full space of possible copper distributions — subject only to the necessary symmetry and manufacturing constraints — the researchers discovered configurations that human designers hadn't considered. The two-layer approach is key: by having one layer actively fed and the other serving as a parasitic element, the system develops two resonance modes that can be overlapped to expand the bandwidth. This is a known technique in antenna design, but the topology optimization approach found a better implementation of it than had been previously achieved.

The use of FR4 is perhaps even more significant than the bandwidth numbers. FR4 is cheap, ubiquitous, and easy to manufacture with standard PCB processes. It's the substrate of consumer electronics, which means any design that works on FR4 can be mass-produced at low cost. High-performance RF substrates like Roger's Rogers or PTFE-based materials are better electrically but cost ten to a hundred times more and require more specialized manufacturing. If topology optimization can make FR4 competitive for dual-polarized wideband applications, the implications for cost-sensitive markets are substantial.

The agreement between simulation and measurement deserves emphasis. In computational electromagnetics, the gap between "it works in simulation" and "it works on the bench" is often large. Fabricated antennas behave differently from their models due to manufacturing tolerances, material property variations, connector effects, and the messiness of the real world. The excellent agreement in this paper — Design II's measured bandwidth of 9.9% versus simulated 9.9%, with resonant frequencies within 2% of predictions — suggests that the optimization process is robust and that the FDTD modeling is accurate enough to trust.

Maximum Realized Gain Comparison

Label	Value
Design I	6.61 dB
Design II	6.5 dB

Design I achieved 6.61 dBi gain at 5.7 GHz; Design II showed similar performance with two resonant modes supporting wider bandwidth.

This validation matters because it opens the door to more ambitious designs. If the simulation-to-hardware pipeline is reliable, designers can use topology optimization for problems where the solution space is larger, the objectives are more complex, or the trade-offs are more severe. The current work solves a well-defined problem with clear objectives. The next step might be multi-objective optimization that explicitly trades bandwidth against gain, or size, or cross-polarization levels. The fact that the method works at all is a proof of concept; the fact that it works accurately is an invitation to push further.

What's Next

The paper leaves several questions unanswered, and several directions open for future work.

The 9.9% bandwidth is impressive, but it's not the theoretical limit. The authors note that their approach allows both copper layers to be optimized simultaneously, yet in the configurations they explored, only one layer was actively connected to the feed probe. The other layer acted as a parasitic element, influencing the antenna through coupling. Future work might explore different feed configurations, different layer arrangements, or multi-layer designs with more than two active layers. The optimization framework is general; the designs presented are initial results, not final answers.

The trade-off between isolation and bandwidth is mentioned but not fully explored. Design I achieved higher isolation (>20 dB across most of the band) with lower bandwidth (5.8-7.5%). Design II achieved wider bandwidth (9.9%) but with isolation that dropped to around 15 dB in some frequency ranges. The abstract mentions achieving "more than 40 dB" isolation in some configurations, but the detailed results for those configurations are not presented in this paper. Understanding and controlling the isolation-bandwidth trade-off systematically would be valuable for different applications. High-isolation designs might be preferred for MIMO systems where cross-talk between channels is harmful; lower-isolation designs might be acceptable for applications where bandwidth is more critical.

The paper focuses on a single frequency range around 5.7 GHz. This is in the C-band, used for 5G communications in many countries, and for Wi-Fi in some regions. The methodology is general and could be applied to other frequency ranges — millimeter-wave 5G, satellite communications, automotive radar — but this hasn't been demonstrated. Different frequencies would require different design domains, different substrates, and possibly different optimization parameters.

Manufacturing at scale remains an open question. The paper demonstrates that a single prototype can be fabricated and validated, but mass production introduces variability. Etching tolerances, drilling accuracy, material lot-to-lot variation, and environmental factors (temperature, humidity) all affect real-world performance. A robust design methodology for volume manufacturing would need to account for these variations, perhaps through sensitivity analysis or robust optimization techniques.

The computational cost is significant. The optimization ran on high-performance GPU nodes, and while the exact run time isn't stated, the 210 iterations with four simulations each implies substantial compute. For this to become a routine design tool, the process needs to be faster, more accessible, or both. Parallelization improvements, surrogate models, or hierarchical optimization (where a coarse design is refined progressively) could address this.

The intellectual implications extend beyond antennas. The success of density-based topology optimization for this electromagnetic problem suggests the approach could be applied to other wave-based devices — acoustic metamaterials, photonic crystals, thermal conductors. The adjoint method for sensitivity analysis works for any linear system governed by partial differential equations; the topology optimization framework is agnostic about what physics it models. The antenna design was a challenging test case, but it's not the only case.

What does this mean for the world?

Every time you make a video call, stream a movie, or download a file over cellular data, your signal passed through an antenna designed by someone who had to guess, calculate, prototype, iterate, and guess again. The process is slow, expensive, and limited by human imagination. The shapes that engineers could conceive and verify were shapes they could draw and analyze; the solution space was bounded by intuition.

Topology optimization removes those bounds. The algorithm doesn't know what an antenna is supposed to look like. It doesn't have preconceptions about what shapes are "reasonable" or "elegant." It only knows the physics and the objectives. When given a dual-polarized antenna design problem on an FR4 substrate, it found a shape that human engineers hadn't imagined — one that achieves more than three times the bandwidth of previous designs. When that shape was built and tested, it worked exactly as predicted.

This is a glimpse of what engineering might look like when computers do the inventing. Not artificial intelligence replacing human engineers, but AI augmenting human engineers — exploring solution spaces too vast for intuition, optimizing for objectives too complex for manual calculation, and generating designs that push against the boundaries of what we thought was possible. The antenna in your next smartphone might not look like anything a human would design by hand. And it might work better for it.