A mirage in the desert makes the eye see water where none exists. In radar imaging, could a device make a detector see a user-designed electromagnetic illusion pattern that has no physical object behind it?
Metasurfaces are artificially engineered electromagnetic interfaces composed of arrays of subwavelength elements. By tailoring the scattering responses of these elements, metasurfaces can regulate key properties of electromagnetic waves, including amplitude, phase, frequency, polarization, and temporal behavior. This capability has made them an important platform for wavefront engineering, with applications in cloaking, holography, beam control and imaging.
For radar imaging, this capability raises an intriguing question: could a metasurface make a detector see a user-designed electromagnetic illusion pattern where no physical target exists? Achieving this on demand remains challenging. Static metasurfaces can create complex scattering responses, but their functions are largely fixed once fabricated. Traditional time-modulated metasurfaces add dynamic control, yet they often rely on globally repeated modulation waveforms, producing regular and predictable illusion patterns. Generating complex two-dimensional illusions that can be rewritten from task to task therefore requires a more flexible temporal programming strategy.
In the study, a research team led by Zhejiang University proposed a deep-learning-enabled temporal sequencing framework for programmable electromagnetic illusions. Inspired by the combinatorial logic of gene regulation in living organisms, the framework regards each time-domain modulation waveform applied during one radar pulse as a reusable “modulation unit.” These units are selected from a predefined library and arranged pulse by pulse to form an executable temporal control sequence. A deep-learning compiler is then used to map a desired illusion pattern to the corresponding sequence, which is implemented by a programmable metasurface to experimentally generate rewritable and customizable electromagnetic illusions. The work, titled “Deep Learning-Enabled Temporal Sequencing of Metasurface for Rewritable and Customizable Electromagnetic Illusions,” has been accepted by National Science Review.
A gene-inspired modulation strategy
The first advance lies in how the metasurface is programmed over time. In many conventional time-modulated metasurface schemes, the modulation waveform or coding strategy remains globally unchanged, or is periodically repeated, throughout the radar observation process. Such stationary temporal control is effective for producing regular responses, but it offers limited freedom for creating complex illusion patterns that need to vary from task to task.
To move beyond this limitation, the team drew inspiration from the combinatorial logic of gene regulation in living organisms. In biology, different selections and arrangements of basic genetic elements can give rise to diverse observable traits. The researchers adopted a similar idea for electromagnetic control: each time-domain modulation waveform applied during one radar pulse was treated as a reusable modulation unit.
These modulation units are collected in a predefined library and assigned to different pulses across the observation process. As a result, different time segments no longer perform the same repeated operation; instead, each segment can impose its own electromagnetic control. The final illusion pattern is therefore shaped by the collective effect of many ordered modulation units. Both the choice of the units and their temporal arrangement influence how the scattered electromagnetic energy is redistributed, and ultimately determine what two-dimensional illusion pattern is reconstructed by the detector.
An intelligent compiler for on-demand illusions
The second advance is the construction of an intelligent compiler that connects a user’s desired illusion pattern with the physical control commands required by the metasurface. Designing such a sequence by hand is difficult because many different modulation units must be selected and ordered across the full radar imaging process. As the desired pattern becomes more complex, manual tuning or brute-force searching becomes increasingly impractical.
To overcome this design bottleneck, the researchers built a deep-learning-based vision-to-sequence compiler. The compiler takes a target illusion pattern as its input, extracts its spatial features, and predicts the corresponding pulse-by-pulse modulation sequence. This sequence is then converted into time-varying voltage commands and executed by a programmable metasurface.
The complete framework was demonstrated in a synthetic aperture radar imaging experiment. After receiving a user-defined illusion pattern, the compiler generated the required temporal sequence, and the metasurface physically implemented it during radar measurement. The detector then reconstructed customized two-dimensional electromagnetic illusion patterns that did not correspond to real targets in space.
The work points toward future intelligent electromagnetic systems that can adaptively shape how they are perceived by imaging sensors. Potential directions include adaptive camouflage, dynamic signature manipulation, reconfigurable coherent imaging, and task-oriented radar-scattering control. The researchers note that richer coding dimensions, more complex temporal waveforms, and faster hardware could further expand the diversity and resolution of achievable electromagnetic illusions.
National Science Review
Experimental study