Sensing and imaging is the engineering layer that turns physical, chemical and biological quantities into readable signals, and the current wave of work replaces bulk optics and benchtop instruments with chip-scale structured photonics, quantum-limited readout and learned signal decoding.
Summary
Sensing and imaging as a taxonomy area covers any device that transduces a physical quantity (a photon flux, a magnetic field, a rotation rate, a molecular fingerprint, a conformational change) into a measurable electrical or optical signal, plus the computation that turns that signal into an estimate. The recent literature clusters around one structural idea: move the function that used to be performed by large, precisely aligned optics or a benchtop spectrometer into a nanostructured surface, a photonic integrated circuit or a quantum-coherent solid-state defect, and recover the lost information in software.
Three distinct engineering routes appear in the sources. The first is resonant nanophotonics: dielectric metasurfaces that concentrate light into a molecular fingerprint band so absorption can be read without a spectrometer. One group fabricated Ge pillar arrays on CaF2 by resist-free stencil lithography, supporting polarisation-insensitive Mie resonances tuned across the 950-1700 cm^-1 fingerprint region, and used germanium’s thermo-optic response to reconfigure the resonance after fabrication. Related work engineers the photonic environment itself, using air-defined silicon “Mie voids” to tune excitation enhancement and quantum yield independently within a subwavelength unit, and flat-band nonlocal membrane metasurfaces to obtain angle-robust chiral resonances.
The second route is quantum-limited readout: pushing detectors down to the fundamental noise floor rather than the classical one. Examples include direct measurement of quantum projection noise in nitrogen-vacancy ensembles of up to 43 spins at room temperature, enabled by optically-detected magnetic resonance contrast above 20 per cent; a continuously monitored superconducting-island detector for single 10 GHz microwave photons with 10 per cent efficiency, sub-50 ns time resolution and roughly 1 microsecond dead time; and a 64 m2 passive resonant gyroscope with a shot-noise limit of 5.7(1)x10^-13 (rad/s)/sqrt(Hz) at 1 mW incident power, against roughly 10^-11 (rad/s)/sqrt(Hz) for today’s large ring laser gyroscopes.
The third route is computational: the sensor is deliberately under-specified and a model does the inversion. A twelve-chirality carbon nanotube array read as time-resolved near-infrared excitation-emission matrices, decoded by a dual-objective convolutional autoencoder, classified ALS serum at 84.6 per cent cross-validated accuracy (AUC 0.87) in 39 subjects. On the imaging side, a dense array of small apertures combined by a photonic integrated circuit is analysed as a replacement for a monolithic telescope aperture, with the finding that spectral resolution must be at least the field-to-resolution ratio and with an explicit signal-to-noise comparison at fixed photon count. The parameters that decide all of this are the same three everywhere: how much signal the structure concentrates, what noise floor the readout reaches, and how much of the discarded physical information the algorithm can recover.
Evidence base
- A 64 m2 passive resonant gyroscope reported a shot-noise limit of 5.7(1)x10^-13 (rad/s)/sqrt(Hz) at 1 mW incident optical power, against roughly 10^-11 (rad/s)/sqrt(Hz) resolution for large ring laser gyroscopes, while noting demonstrated passive-gyroscope resolutions remain about two orders lower.
- Sub-bandgap AlN Schottky photodetectors showed non-saturating linear response to blue light above 40 W/cm2 and undistorted linearity to at least 300 degrees C, attributed to deep-level-mediated photoresponse with a narrow space charge region.
- Ge pillar metasurfaces on CaF2, made by resist-free stencil lithography, gave polarisation-insensitive Mie resonances spanning the 950-1700 cm^-1 molecular fingerprint region with thermo-optic post-fabrication tuning.
- Room-temperature projection-noise measurement and spin counting of up to 43 nitrogen-vacancy spins was achieved using ODMR contrast above 20 per cent from polarisation-selective excitation plus spin-to-charge conversion, without cryogenics or high bias fields.
- A superconducting-island detector monitored by microwave reflectometry detected single 10 GHz photons at 10 per cent efficiency with sub-50 ns time resolution and about 1 microsecond dead time.
- A 12-chirality DNA-functionalised SWCNT array read as excitation-emission matrices at 0, 6 and 24 hours, decoded by a dual-objective convolutional autoencoder, classified ALS versus matched control serum at 84.6 per cent cross-validated accuracy and AUC 0.87 in 39 subjects.
Open questions
- Does resist-free stencil lithography for Ge-on-CaF2 metasurfaces hold resonance uniformity across a full wafer, and at what yield and cost per die?
- Can the 12-chirality SWCNT corona signature reproduce at AUC near 0.87 across independent sites and serum handling protocols, or does the autoencoder latent space encode batch effects?
- Does the passive resonant gyroscope’s shot-noise limit of 5.7x10^-13 (rad/s)/sqrt(Hz) translate into a demonstrated rotational resolution that closes the two-order gap to ring laser gyroscopes, or do technical noise sources dominate?
- What efficiency can the photo-assisted quasiparticle tunnelling detector reach while keeping sub-50 ns resolution and roughly 1 microsecond dead time, and does it stay competitive against existing microwave photon counting schemes?
- In which regimes of field, resolution and photon budget does a PIC aperture array actually beat a monolithic telescope on signal-to-noise, given the spectral resolution constraint set by the field-to-resolution ratio?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.