Summary
Conventional imaging separates concerns: optics form a sharp image on a focal plane, the sensor records it, software cleans it up. The cluster of work collected here abandons that separation. In a compact interferometric imager, the large aperture of a telescope is replaced by a dense array of small apertures combined by a photonic integrated circuit, which records complex visibilities rather than an image; the components that traditionally form the image are simply absent, and the object is reconstructed numerically. In end-to-end meta-imagers, a metasurface and its computational back-end are optimised jointly, and recent work reduces that optimisation to closed-form, data-free objectives built on Shannon capacity and Fisher information, showing that the optimal transfer matrix is a permutation matrix, each source mapped to one distinct detector, a condition the authors call generalised focusing.
The algorithm layer is being pulled in the same direction, but by a different pressure: cost. On AR glasses, continuous high-resolution egocentric capture is the bottleneck, and a co-optimised system that uses gaze, pose and inertial cues to discard uninformative input cuts memory 27.5-fold and energy 24.3-fold while preserving task accuracy. In X-ray imaging of laser powder bed fusion keyholes, a purpose-built segmentation algorithm processes a cropped frame in 19.9 ms on CPU against 54 ms and 5284 ms for the machine learning baselines tested.
The parameters that decide this field are: whether the reconstruction penalty is acceptable (the interferometric imager work is explicitly an SNR comparison against a monolithic telescope at equal photon count, and it imposes a hard constraint that spectral resolution be no less than the field-to-resolution ratio); whether reconfigurable optics survive the environment (multimode fibre imaging is crippled by bending, and the finding that higher-order modes are far more bend-robust, with the perturbation dominated by changes in effective propagation constants rather than inter-modal power coupling, is what could make it usable); and whether design tools scale, which is what the data-free objectives and the differentiable neural surrogate for scintillator cascades address.
Evidence base
- 15 June 2026: a concept study replaces a telescope’s large aperture with a PIC-combined array of small apertures, deriving that spectral resolution must be no less than the field-to-resolution ratio and comparing SNR against a monolithic telescope in Fourier space at equal photon count.
- 16 June 2026: closed-form, data-free objectives based on Shannon capacity and Fisher information yield metasurface designs matching end-to-end optimisation, with the optimum proved to be a permutation matrix regardless of source or detector geometry.
- 16 June 2026: EPIC cuts memory 27.5x and energy 24.3x versus a full-video baseline on AR glasses while preserving egocentric video understanding accuracy.
- 16 June 2026: MOSAIC segments L-PBF keyhole X-ray frames at 19.9 ms per image on CPU against 54 ms and 5284 ms baselines, with average F1 0.894 and precision 0.953 across 12 samples and no training required.
- 16 June 2026: DarkFlow keeps packet loss ultra-low at billion-photon event rates where conventional digital SiPM readout can exceed 80% data loss, with 2.14x better effective refresh rate, evaluated in GlobalFoundries 22nm.
- 8 July 2026: a 20x20 microfluidic acoustic lens driven by 80 micropumps gives 400 addressable phase pixels, roughly one second reconfiguration, and 3D focusing resolution comparable to a 400-element transducer array.
Open questions
- Does the compact interferometric imager’s noise propagation beat or lose to a monolithic telescope at equal photon count, and by what factor? The excerpt of states the comparison is made but is truncated before the result.
- Do the closed-form Shannon-capacity and Fisher-information objectives of reproduce end-to-end optimisation quality when replicated outside the originating group and on fabricated, not simulated, metasurfaces?
- How much bending does the higher-order-mode robustness in actually tolerate before the input-independence assumption fails, given that clinical endoscopy involves more than moderate bending?
- Can the Sb2S3 varifocal metalens of be cycled between states enough times, and fabricated at enough yield, to survive a product lifetime? No source addresses switching endurance or manufacturing yield.
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.