Fingerprint authentication has long depended on a digital chain of custody: a sensor captures a fingerprint, software converts the image into a mathematical template, and an algorithm compares that template with stored biometric data. The arrangement is efficient, but it creates a valuable target. Databases containing biometric templates can be hacked, copied or misused, and unlike a password, a fingerprint cannot simply be changed after exposure. A study published in Nature Sensors now describes a radically different approach in which the comparison is performed directly by waves moving through a physical device. Instead of digitizing the fingerprint and matching it against a stored record, the proposed system uses ultrasonic propagation and a learned metasurface to make the authentication decision in a single physical computation.
The platform, developed by S. Lee, B. Oh, W. Kim and colleagues, is based on an all-wave computational framework built around a loop-diffractive neural network. In conventional neural networks, information is transformed through layers of numerical operations carried out by processors. In a diffractive neural network, the equivalent computation occurs as a wavefront travels through carefully designed structures that alter its phase and amplitude. The new system places a single learned metasurface within an ultrasonic loop, allowing acoustic waves to propagate, diffract and be redirected through the device. The result is a form of hardware-level inference: the physical behavior of the waves performs the calculation instead of software reconstructing the process after measurement.
Ultrasound is central to the design because it can carry structured information through space while interacting with engineered surfaces in a controllable manner. When an acoustic wave encounters a metasurface, the surface can modify how the wave advances, including the timing of different portions of the wavefront and the distribution of energy across space. These changes are governed by the geometry and physical properties of the surface. In the reported framework, the metasurface is trained so that the wave generated by a registered fingerprint is guided toward a designated detector region. A nonmatching fingerprint produces a different propagation pattern, causing the strongest energy to accumulate elsewhere. Authentication is then determined by reading the relative intensity at the predefined regions.
A significant feature of the device is its three-parameter modulation strategy. Rather than controlling only the phase of an acoustic wave, the metasurface can manipulate phase, amplitude and switch state. Phase determines how different parts of the wavefront interfere as they propagate; amplitude controls the strength of the contributions entering the computation; and the switch state provides an additional discrete degree of control over whether a local element participates in the interaction. Combining these variables expands the number of physical transformations the surface can perform. For a diffractive network, that larger design space can make it easier to shape complex wave patterns and separate the response associated with a genuine registered input from responses generated by other fingerprints.
The loop architecture adds another dimension to the computation. In a simple one-pass diffractive system, an input wave travels through the device once before reaching a detector. A loop-diffractive neural network instead allows the acoustic information to circulate through the learned structure, creating repeated interactions with the metasurface. Each pass can redistribute the wave’s energy and alter the information carried by the field. Physically, this resembles giving a compact optical or acoustic processor multiple opportunities to refine a decision without requiring a long sequence of separate digital layers. The reported design uses a single metasurface as the learned computational element, suggesting that repeated propagation can provide processing depth while keeping the physical system comparatively compact.
During authentication, the fingerprint is converted into an acoustic input rather than being treated solely as a digital image for algorithmic matching. The wave entering the system carries the spatial pattern associated with the user’s fingerprint. As it propagates through the loop and interacts with the metasurface, diffraction transforms that pattern according to the learned response of the device. If the input corresponds to the registered user, acoustic energy is directed toward the system’s designated “match” detector region. If it does not, the energy is concentrated in an alternative region. The authentication outcome therefore comes from a physical intensity comparison, not from a conventional software routine that searches a database of stored templates.
This distinction is important because the system’s security model is based on what the device does rather than on what it stores as a digital record. In a traditional biometric architecture, a template must be retained somewhere so that a future fingerprint can be compared with it. In the proposed all-wave design, the learned response is embedded in the physical metasurface and its operating conditions. The device is trained to produce a particular output behavior, but the authentication process does not require retrieving a digital representation of the user’s biometric pattern. The researchers present this as a way to reduce exposure to template theft and data breaches. It does not eliminate every possible security concern: the physical device itself would still need protection against tampering, duplication or unauthorized characterization.
The approach also illustrates why physical computing has become an increasingly prominent theme in advanced sensing. Waves naturally propagate and interact in parallel, meaning that many parts of a spatial input can be processed at the same time rather than being scanned sequentially by a central processor. Diffraction performs transformations through interference, while the metasurface supplies a carefully engineered response. Because the calculation is carried out by the motion of the wave, the system can avoid some of the data movement and repeated numerical operations associated with conventional digital processing. In an authentication setting, where the desired answer may simply be “match” or “no match,” directing energy to one detector region or another could offer a particularly direct route from sensing to decision-making.
The work also carries a strong practical and conceptual message: biometric security does not necessarily have to be built around a database. A fingerprint can be treated as a physical signal, and the authentication process can be designed as a transformation of that signal in space. The metasurface-driven loop network turns recognition into an event that unfolds in matter, with the answer encoded in the destination of acoustic energy. That idea could attract attention beyond fingerprint identification, especially as researchers investigate optical, microwave and acoustic systems capable of carrying out machine-learning operations without conventional digital postprocessing. The reported framework remains a research platform rather than a universal replacement for deployed authentication systems, and real-world adoption would require testing under variations in finger placement, pressure, noise, ageing and deliberate spoofing. Even so, the study points toward a future in which some of the most sensitive computing tasks are performed not by storing more personal data, but by designing physical systems that never need to keep the data in the first place.
Subject of Research: All-wave computational ultrasonic fingerprint identification using a metasurface-driven loop-diffractive neural network.
Article Title: All-wave computational ultrasonic fingerprint identification with metasurface-driven loop-diffractive neural network.
Article References: Lee, S., Oh, B., Kim, W. et al. “All-wave computational ultrasonic fingerprint identification with metasurface-driven loop-diffractive neural network.” Nature Sensors 1, 728–739 (2026). https://doi.org/10.1038/s44460-026-00102-7
Image Credits: AI Generated
DOI: 10.1038/s44460-026-00102-7
Keywords: Ultrasonic computing, fingerprint identification, biometric security, metasurfaces, diffractive neural networks, physical computing, wave-based authentication, acoustic signal processing.
