The framework, described in a paper published in Frontiers in Antennas and Propagation by researchers from MathWorks and Belden’s Ruckus Networks, tackles two of the hardest problems in contemporary radio-frequency design. The first is miniaturization: shrinking microstrip patch antennas while keeping their bandwidth and gain within usable limits. The second is phase stability: ensuring that reconfigurable intelligent surfaces, or RIS, maintain a consistent reflection phase across wide frequency bands and a broad range of scan angles. Both problems share a common enemy — an enormous, nonlinear design space that would exhaust any human designer and overwhelm conventional parametric tuning.

The core idea behind PARISO is deceptively simple. Instead of tweaking a handful of traditional parameters, the software slices a conducting patch or unit cell into a grid of discrete squares, each representing a pixel that can either hold metal or not. Every pixel becomes a binary gene in a chromosome, and optimization algorithms — either a Genetic Algorithm or a Surrogate-Assisted Differential Evolution approach known as SADEA — mutate and recombine these genes across generations of candidate designs. Because the technique is frequency-independent, the same framework can target everything from tri-band Wi-Fi access points operating at 2.4, 5, and 6 GHz to millimeter-wave and future sub-terahertz 6G systems.

What makes this pixel-based representation so powerful is its topology freedom. Traditional antenna design constrains engineers to recognizable shapes — rectangles, circles, slots in predictable positions — because a human has to sketch them. Pixelization removes that constraint entirely. Toggling a single metallization cell can re-route surface currents, open or close slots, or alter the effective inductive-capacitive loading of the structure, producing fractal-like geometries that no designer would draw by hand but that full-wave electromagnetic simulation reveals to be highly effective. This is precisely the kind of non-intuitive exploration at which evolutionary algorithms excel.

The physics of why this works is worth appreciating. Gradient-based optimizers assume a smooth objective landscape, but full-wave electromagnetic responses are anything but smooth. Resonances, mode switching, and discrete geometry changes produce jagged, multi-modal cost surfaces where a single-pixel change can cause abrupt jumps in the S11 reflection coefficient, gain, or bandwidth. Genetic Algorithms, which rely on stochastic selection, crossover, and mutation rather than derivatives, navigate such landscapes naturally. Each candidate design is evaluated using a hybrid Method of Moments solver tailored for planar and layered printed-circuit-board structures, and the resulting scores — resonance frequency, minus-10 dB bandwidth, and peak realized gain — feed a weighted cost function that balances the competing goals of miniaturization, bandwidth preservation, and radiation robustness.

The demonstration results are striking. Starting from a plain square patch antenna resonating at 2.65 GHz, a genetic search on a 10-by-10 pixel grid drove the resonance down to 2.35 GHz — a genuine size reduction — while realized gain slipped only from 8.88 dBi to 8.55 dBi and bandwidth from 120 MHz to 110 MHz. In an infinite-array configuration using Floquet-type periodic boundary conditions, the same approach shifted a unit cell’s resonance from 2.66 GHz to 2.26 GHz while actually expanding the bandwidth more than tenfold, from 20 MHz to 232 MHz, with element gain essentially unchanged. Crucially, the framework enforces fabrication-aware constraints throughout: metallized pixels overlap by ten percent to guarantee continuous current flow, edge cells are protected, the feed region is masked from modification, and symmetry is enforced in both directions so that far-field radiation patterns remain well-behaved.

For reconfigurable intelligent surfaces, the optimization goal flips. Rather than shifting resonance, the framework minimizes the variation of the reflection-coefficient phase across a specified band and over scan angles from zero to sixty degrees, a property essential for RIS deployments in Wi-Fi environments where multipath propagation and channel variation are severe. In a representative 20-by-20 unit-cell example, both GA and SADEA substantially flattened the phase response compared with the baseline square patch, with phase variation at 2.2 GHz dropping from 6.08 degrees to roughly 2.5 to 3.1 degrees depending on the optimizer.

The comparison between the two algorithms reveals a compelling computational story. SADEA augments differential evolution with a learned surrogate model — typically built on radial basis functions or Gaussian processes — that approximates the expensive electromagnetic objective and proposes new candidates by balancing expected improvement against exploration. In the RIS study, where each candidate required thirty-five separate field computations spanning five frequencies and seven scan angles, GA consumed roughly 5,460 total field computations while SADEA achieved comparable design quality with about 525 — a roughly tenfold reduction in simulation effort. That kind of efficiency gain matters enormously when every evaluation involves a full-wave solve, and it positions surrogate-assisted methods as the practical route to larger pixel grids and more complex multi-objective problems.

Accessibility is as central to the project as the algorithms themselves. The entire workflow — geometry generation through Boolean operations on multilayer PCB stack-ups, electromagnetic evaluation, fitness formulation, and optimization — is implemented in MATLAB and wrapped in a graphical user interface built with App Designer, featuring panels for input parameters, patch geometry, optimization settings, and execution control. The complete code, examples, and interface are freely available on GitHub, an explicit bid to encourage reproducibility and lower the barrier for research groups that lack the resources to build such pipelines from scratch. The authors are candid that the relative performance of GA versus SADEA is problem-dependent, framing the contribution as a flexible framework rather than a verdict on any single method.

The roadmap ahead is ambitious. The team plans to extend the cost function toward direct phase targeting, allowing individual RIS unit cells to realize prescribed reflection phases such as 0, 90, 180, or 270 degrees across wide bandwidths; to generalize the binary pixel encoding into multi-state, two-bit metallizations that yield four discrete phase states per cell; and to co-optimize feed location alongside the pixel mask for miniaturized patches. Hybrid GA-SAdeA pipelines, hierarchical coarse-to-fine pixel grids, and large-scale parallel execution are also on the agenda, along with the eventual fabrication and measurement of optimized designs. As 5G densifies, Wi-Fi 7 spreads across its new 6 GHz allocation, and 6G research pushes toward sub-terahertz frequencies, tools like PARISO suggest a future in which the antennas and smart surfaces surrounding us are not drawn by hand but grown — pixel by pixel — inside an evolutionary search that discovers shapes human intuition would never reach.

Subject of Research: Pixel-based optimization of planar antennas and reconfigurable intelligent surfaces using genetic algorithms and surrogate-assisted evolution

Article Title: PARISO: pixelated antennas and reconfigurable intelligent surface optimizer

Article References: Sekharan, S., Fazel-Rezai, R., & Noghanian, S. (2026). PARISO: pixelated antennas and reconfigurable intelligent surface optimizer. Frontiers in Antennas and Propagation, 4, Article 1830853. https://doi.org/10.3389/fanpr.2026.1830853

Image Credits: AI Generated

DOI: 10.3389/fanpr.2026.1830853

Keywords: antenna optimization, reconfigurable intelligent surfaces, genetic algorithms, surrogate-assisted optimization, pixelated antennas, miniaturization, RIS phase stability, 6G, Wi-Fi 7, massive MIMO, full-wave simulation, open-source software