Zhejiang University researchers built a deep-learning compiler that converts a desired radar picture into a sequence of time-coded commands — and a detector reconstructed patterns with no object behind them.
A metasurface is a flat array of sub-wavelength elements that controls how electromagnetic waves scatter off it. If each element can be switched electronically, the surface’s response becomes programmable — and if it can be switched quickly, the response can be changed within the time it takes a radar pulse to arrive and leave.
A study in National Science Review from a Zhejiang University team exploits that second property in a way the paper’s title describes as sequencing: treating time-domain modulation patterns as reusable units to be assembled, and letting a neural network decide the order.
The expert contact for the work is Hongsheng Chen of Zhejiang University.
Modulation as a vocabulary
The framework treats each time-domain modulation waveform applied during one radar pulse as a reusable modulation unit, drawn from a predefined library and arranged pulse by pulse.
The genetic metaphor the release reaches for is apt in a specific way. A fixed alphabet of units, combined in different orders to produce different outcomes, is a compositional system — and the practical advantage is that the hard engineering work goes into building the library once, after which new behaviours are new sequences rather than new hardware.
What converts an intention into a sequence is a deep-learning compiler. “The compiler takes a target illusion pattern as its input,” the release explains, “extracts its spatial features, and predicts the corresponding pulse-by-pulse modulation sequence.”
Calling it a compiler rather than a model is a reasonable description of what it does. The problem it solves is an inverse-design problem: a forward calculation from a modulation sequence to the resulting radar image is tractable, while working backwards from a desired image to the sequence that produces it has no closed-form solution. Training a network to approximate that inverse is the standard modern approach, and it is what makes the system usable by someone specifying a picture rather than a waveform.
What was demonstrated
In a synthetic aperture radar imaging experiment, a detector reconstructed customised two-dimensional illusion patterns with no matching physical target.
The choice of synthetic aperture radar is what makes the demonstration meaningful rather than trivial. SAR builds an image by combining returns collected as the radar moves relative to the scene, over many pulses — so the image a SAR produces depends on the coherent relationship between those returns across time. A surface that changes its response between pulses is intervening in exactly the quantity SAR relies on, which is why modulating pulse by pulse can write a shape into the final image rather than merely altering how bright a point appears.
The applications the release names follow from that: adaptive camouflage, dynamic signature manipulation, reconfigurable coherent imaging, and task-oriented radar-scattering control.
The first two are the obvious ones and are worth stating plainly rather than euphemistically — a surface that can make a radar see a shape that is not there, or fail to see one that is, is a counter-detection technology, and most funding in this area reflects that. The third runs the other way: the same control over coherent returns could be used to improve imaging, by having a cooperative surface structure the illumination of a scene.
What the release does not report
The absences here are unusually comprehensive, and they limit what can be said about the result.
There are no quantitative results at all. No figure for how closely the reconstructed illusion matched the intended pattern, no frequency band, no modulation switching rate, no metasurface element count or aperture size, and no indication of the range or geometry at which the demonstration was performed. For a system whose feasibility depends entirely on whether the surface can switch fast enough relative to the radar’s pulse repetition frequency, the absence of any timing figure is the central gap.
The release also does not list the paper’s authors. Chen is given as expert contact without his role on the paper being stated, and no author roster appears, so the work cannot be attributed to individuals from the release alone.
Nor are funding sources or conflicts of interest stated.
The constraint the concept implies
One limitation follows from the physics rather than from the reporting, and the release does not address it.
A system of this kind has to know what it is being illuminated by. The modulation sequence that produces a given illusion in a given radar’s image depends on that radar’s waveform, pulse repetition frequency, bandwidth and the geometry of its pass — which means the compiler is solving the inverse problem for an assumed observer. Against a radar whose parameters differ from the assumption, or which varies them deliberately, the carefully composed sequence produces something other than the intended picture.
That is the standard dynamic in this field: countermeasure and counter-countermeasure, where agility on one side invites agility on the other. It does not diminish the demonstration, and it is the reason a laboratory result against a known radar is some distance from a capability against an uncooperative one.
What the work establishes, within those limits, is an architecture: a reusable library of temporal modulation units, a learned inverse map from pictures to sequences, and a programmable surface to execute them. The paper was published on 10 August 2026.
Notes
- “Deep Learning-Enabled Temporal Sequencing of Metasurface for Rewritable and Customizable Electromagnetic Illusions.” National Science Review, 2026. DOI: 10.1093/nsr/nwag263. The release does not name the paper’s authors.
- Science China Press. “Sequencing electromagnetic waves like genes: a metasurface creates customizable electromagnetic illusions.” EurekAlert!, 10 August 2026. https://www.eurekalert.org/news-releases/1139489.
Photo by Prasanjeet Shyam on Pexels.com