An artificial-intelligence system detected thin air gaps behind ceramic tiles using only 0.5 gigahertz of radar spectrum. That is one seventy-second of the 36-gigahertz sweep used to collect the original signals. The Singapore and Japan team tested the method on 12 manufactured wall specimens, and the stricter specimen-by-specimen result shows that the bandwidth saving does not yet equal field reliability.
An air gap can form when a tile or mortar layer begins to separate from a wall. Radar can inspect the hidden layers without removing the tile because the gap changes the strength and timing of the returning radio wave. A wider sweep separates nearby layers more clearly, but it needs more capable hardware and may be harder to use under local spectrum rules.
The new system, RF-VoID, avoids reconstructing a full depth profile. At each scanned point it reads 28 measurements within the narrow band and examines both small changes between neighbouring frequencies and broader patterns across the band. It then assigns a probability that a gap lies beneath that point and combines the points into a defect map.
The wall sections contained controlled gaps 0.5 to 1 millimetre thick. In the easier test, training and evaluation points could come from the same specimens. In the stricter test, the model trained on 11 entire specimens and was evaluated on the twelfth, repeating the process until every wall had been held out. This second design better measures whether the model can handle a new wall.
The easier test reached 95.84% on macro F1, a score that balances missed defects with false alarms. In the stricter test, the average fell to 62.12% at 0.5 gigahertz. Doubling and quadrupling the bandwidth raised it only to about 69%, although RF-VoID still led the comparison models. The large gap between the two tests indicates that specimen-specific patterns contributed heavily to the headline accuracy.
The experiment used a mechanical scanner in a laboratory and did not include occupied buildings, outdoor weather or the variety of tile, mortar and concrete found in practice. A broader training set and blind tests on real facades are required before narrowband radar can support inspection decisions.
Sources
- Xinyan Chen and colleagues, arXiv, September 11, 2026. Paper summary, authorship, submission date and principal results.
- Full paper. Measurement system, model sequence, test protocols, detailed scores and limitations.
- Lead image: Daniel Ramirez, “Window Cleaners @ Waena”, used without alteration under the CC BY 2.0 licence. The photograph is illustrative and does not show the experiment.
.jpg){kind=link}