Integrated GRM-based efficient multi-performance prediction method for reconfigurable Fabry–Perot antennas

Smart antennas for a crowded airwaves era

As our phones, cars, and homes compete for space on the wireless spectrum, antennas must do far more than simply send and receive signals. They need to adapt on the fly to changing conditions, pointing their beams toward desired users while rejecting interference, all without growing bulky or power-hungry. This paper introduces a fast, data-driven way to design a special kind of “shape-shifting” antenna called a reconfigurable Fabry–Perot antenna, promising smarter wireless hardware for future communication systems.

Antennas that can change their mind

Traditional antennas are usually built for one main job: a fixed frequency band and a mostly fixed beam pattern. In modern networks, that rigidity becomes a liability. Reconfigurable antennas address this by letting key parts of the structure switch between different states, changing how radio waves are reflected or radiated. In Fabry–Perot antennas, a feed antenna sits under a partially reflective surface made of many small “unit cells.” By altering the state of each cell—using technologies such as liquid-filled channels, switches, or smart materials—the antenna can steer its beam or adjust its performance without any mechanical motion.

Why brute-force design hits a wall

Designing these reconfigurable surfaces is extremely challenging. Each tiny cell can be on or off, or tuned in several ways, and a practical surface can contain dozens or hundreds of cells. The number of possible combinations explodes, and testing each configuration with full electromagnetic simulations is painfully slow. Engineers are left with a problem: they need an efficient way to predict how any given pattern of cell states will affect key performance measures—such as how well the antenna is matched to the radio circuit (return loss), how strong the signal is (gain), and where the beam points (radiation pattern)—without simulating every possibility from scratch.

A learning shortcut between switches and signals

The authors propose a generalized regression model (GRM) that acts as a smart shortcut between switch settings on the surface and the resulting antenna behavior. They encode each design as a simple binary string, where a “0” or “1” marks whether a cell is inactive or activated. This string is fed into three parallel neural networks, each specialized in predicting one aspect of performance: return loss, gain across frequency, and the shape of the radiation beam. These networks belong to a family called generalized regression neural networks, chosen because they work well even when only a modest number of carefully simulated examples are available.

Tuning the model and reversing the problem

To make the predictions both accurate and reliable, the authors automatically adjust internal settings of each network using a particle-swarm optimization method that searches for the best balance between fitting known data and generalizing to new cases. They then refine the model through a recursive “correction” process, which iteratively learns the gaps between predictions and high-fidelity simulations until the error falls below a set threshold. With this forward model in place, they tackle the inverse problem: instead of asking “what will this pattern do?”, they ask “which patterns will meet my performance goals?”. By rapidly scanning and pruning the space of possible configurations—discarding any that violate simple return-loss or beam-direction constraints—the framework homes in on promising designs without relying on heavy-duty simulations at every step.

Liquid-filled hardware put to the test

To prove the approach in real hardware, the team uses a Fabry–Perot antenna whose reflective surface contains tiny channels filled with deionized water. Changing which channels are filled alters how the surface reflects radio waves. With only 100 simulated examples covering a range of activation patterns, the GRM learns to predict the antenna’s return loss, gain, and beam shape far more accurately than several competing surrogate models. It also does so at a fraction of the computational cost of full electromagnetic solvers. When the inverse design stage is applied, the model quickly finds liquid patterns that satisfy both a strict matching requirement and a desired beam steering angle, and measurements of a fabricated prototype closely track the predictions.

Faster paths to agile wireless hardware

In plain terms, this work shows how a tailored machine-learning model can stand in for thousands of expensive physics simulations, giving antenna designers a powerful “calculator” that maps switch patterns on a smart surface directly to real-world performance. Just as weather models guide forecasts without measuring every cloud, the GRM guides antenna configuration without simulating every wave. The result is a practical route to rapidly designing antennas that steer and shape their beams on demand, an ability that will be increasingly crucial for future wireless, radar, and sensing applications.

Citation: Huang, Y., Liu, Z., Wang, D. et al. Integrated GRM-based efficient multi-performance prediction method for reconfigurable Fabry–Perot antennas. Sci Rep 16, 12492 (2026). https://doi.org/10.1038/s41598-026-42164-0

Keywords: reconfigurable antennas, Fabry–Perot cavity, surrogate modeling, neural network design, beam steering