The work, reported by Huang, Jiang, Gao and colleagues in Nature Sensors, is based on an artificial retina made from indium nitride, or InNₓ, waveguide-coupled photodetectors. The researchers took inspiration from the human eye, where the retina is not simply a flat sheet of independent light sensors. Instead, retinal cells are arranged in a layered, vertically organized structure that captures, filters and begins processing optical information before signals travel to the brain. The new device uses a comparable vertical one-dimensional pixel architecture, allowing light collection and signal conversion to take place in a compact stack rather than across a large lateral footprint.

At the heart of the system are pixels coupled to optical waveguides. A waveguide confines and directs light, much like a microscopic channel for photons. In this artificial retina, waveguide coupling helps guide incoming light toward the active InNₓ photodetection regions, improving the interaction between the optical field and the semiconductor material. This arrangement can increase the effective absorption of light while allowing individual pixels to occupy less surface area. By moving part of the optical path into the vertical dimension, the architecture addresses a fundamental trade-off in conventional image sensors: smaller pixels often collect fewer photons, while larger pixels limit resolution and integration density.

The reported artificial retina contains 7.84 million pixels per square centimetre, equivalent to approximately 7,000 pixels per inch. That density is about an order of magnitude higher than the density of photoreceptors in the human retina, according to the researchers. High pixel density is important for robotic vision because it can improve the detection of fine edges, small objects and rapid changes in a scene. Yet density alone does not guarantee useful imaging. Tiny pixels can suffer from weak signals, optical crosstalk and increased noise. The waveguide-coupled design is intended to preserve light-collection efficiency even as the pixels shrink, while the flexible substrate allows the retina to conform to curved optical surfaces.

The device also achieved a reported specific detectivity of 2.17 × 10¹⁴ jones, a figure used to describe how effectively a photodetector can distinguish weak optical signals from background noise. In practical terms, high detectivity means that the sensor can respond to very low levels of illumination without being overwhelmed by dark current or electronic fluctuations. This capability could be valuable in applications where artificial eyes must operate across changing conditions, from bright outdoor environments to dim indoor scenes. The reported performance suggests that the InNₓ photodetectors can combine miniaturization with a strong electrical response, although real-world operation will also depend on system-level factors such as calibration, temperature stability and optical alignment.

The researchers integrated the retina into a bionic-eye platform with a tunable depth of field. Depth of field describes the range of distances that appear acceptably sharp in an image. Human eyes alter their focus by changing the shape of the lens, but many artificial vision systems rely on fixed optics or mechanical focusing elements. A depth-tunable bionic eye can instead adapt its imaging configuration to objects at different distances. In the reported demonstrations, this capability supported three-dimensional spatial imaging and motion tracking, allowing the system to gather information not only about where objects appear in a scene, but also about their relative depth and movement.

The optical system was further equipped with coordinated aberration compensation. Aberrations are imperfections introduced when lenses or curved imaging surfaces fail to bring all light rays to the same ideal focus. They can blur details, distort shapes and reduce image quality, particularly near the edges of a wide field of view. The researchers report that their approach produced a 110-degree field of view while reducing field curvature by 45.7%. Field curvature causes a flat scene to project onto a curved focal surface, meaning that an image cannot remain uniformly sharp from centre to edge. Correcting this effect is essential for a bionic eye intended to operate in dynamic environments, where important information may appear anywhere within a broad visual field.

The system’s most ambitious feature is that some visual processing takes place inside the sensor rather than being sent entirely to an external processor. This approach, known as in-sensor processing, can reduce the amount of raw data that must be transferred and analysed by a separate computer. Conventional cameras may generate enormous streams of pixel data, creating delays and consuming significant energy when every frame must be transmitted for interpretation. By performing functions such as signal conditioning, denoising or feature extraction at the point of image capture, a neuromorphic sensor can respond more efficiently to events. The researchers demonstrated denoising-based recognition, suggesting that the device can suppress unwanted fluctuations while retaining visual features needed for identifying objects or patterns.

In motion experiments, dynamic focus adaptation helped the bionic eye reconstruct trajectories with a reported accuracy of 96.1%. The result points toward applications in autonomous robots, wearable devices, intelligent cameras and machines that must react to moving objects in real time. A robot equipped with such a vision system could potentially track a fast-moving target while adjusting focus and compensating for optical distortion, rather than relying on multiple cameras or computationally intensive corrections after image capture. The system’s combination of dense sensing, depth awareness, wide-angle imaging and local processing is particularly relevant to embodied intelligence, in which machines must interpret their surroundings while physically interacting with them.

The researchers describe the artificial retina as a step toward vision systems that more closely resemble the integrated operation of biological eyes and brains. Its vertical pixel architecture addresses the physical challenge of packing large numbers of sensitive detectors into a small, flexible area, while waveguide coupling helps preserve optical efficiency at the pixel scale. Tunable depth of field and aberration compensation extend the system beyond basic image capture, giving it the ability to adapt to different distances and maintain image quality across a broad view. Further development will be needed to assess durability, manufacturing scalability, power consumption and performance outside controlled demonstrations. Even so, the work offers a striking glimpse of how future machines might see: not through a camera that merely records images, but through a responsive artificial retina that focuses, filters, interprets and acts on visual information as it arrives.

Subject of Research: Flexible neuromorphic artificial retina and bionic-eye vision system using ultradense InNₓ waveguide-coupled photodetectors for tunable-depth 3D imaging, motion tracking and in-sensor processing.

Article Title: A neuromorphic bionic eye with ultradense waveguide-coupled pixels for depth-tunable 3D vision

Article References: Huang, PY., Jiang, B., Gao, JS. et al. “A neuromorphic bionic eye with ultradense waveguide-coupled pixels for depth-tunable 3D vision.” Nature Sensors (2026). https://doi.org/10.1038/s44460-026-00119-y

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s44460-026-00119-y

Keywords: neuromorphic vision, bionic eye, artificial retina, InNₓ photodetectors, waveguide-coupled pixels, 3D imaging, depth of field, motion tracking, in-sensor processing, embodied intelligence

The concept of diffractive neural networks has captured significant attention in recent years, offering a paradigm shift by exploiting the physical propagation of light through engineered surfaces to perform neural computations. Traditional diffractive neural networks, however, suffer from their fixed architectural design, enabling only a single task or requiring extensive redesign for new functionality. Enter the movable-type coding metasurface: a novel metasurface design featuring discrete, modular units whose optical responses can be independently controlled and dynamically rearranged, unlocking unparalleled versatility in neural computation.

At its core, the researchers developed a metasurface composed of multiple individual coding elements whose positions can be physically adjusted, allowing the phase and amplitude modulation of light waves traversing the surface to be dynamically tailored. This movable architecture is inspired by the uniform movable type printing technology, where individual characters can be rearranged to generate new texts. Here, the optical “characters” are nanoscale coding units manipulated to implement complex diffractive functions that can be reprogrammed on demand without reconstructing the entire metasurface.

To achieve precise reconfigurability, the team engineered a microscale mechanical system capable of fine-tuned lateral displacements of coding units. This mechanical control, integrated with high-fidelity metasurface fabrication techniques, ensures that optical properties across the surface can be swiftly modified to perform different neural operations. Such an approach describes a significant leap forward compared to static metasurfaces limited to fixed light modulation patterns.

The practical implications of this advancement are vast. By enabling reconfigurable diffractive neural networks, a single metasurface device can switch between multiple computational modes or functions, tailoring its response to specific real-time inputs or tasks. This adaptive capability opens new avenues for on-chip optical computing, dynamically programmable holography, and multi-tasking photonic AI systems, enabling machine learning directly in the optical domain at speeds and efficiency unattainable by electronic processors.

In terms of optical performance, the metasurface leverages state-of-the-art nanofabrication to achieve strong light-matter interactions with subwavelength precision modulation of optical wavefronts. The coding elements are designed to support amplitude and phase modulations across the visible to near-infrared spectra, ensuring broad applicability across optical communication, sensing, and imaging domains.

The experimental validation demonstrated remarkable classification accuracy across diverse datasets by reprogramming the coding units to implement different sets of diffractive neural network layers. This multi-modal classification exemplifies how the platform inherently addresses the challenge of hardware inflexibility in previous photonic AI systems. Moreover, the reconfiguration speed achieved by the movable components enables near real-time switching between tasks, a critical requirement for adaptive optical processing systems deployed in dynamic environments.

A striking feature of this work is the multiplexed coding scheme that captures multiple functionalities within the same metasurface footprint. By spatially rearranging the coding units, the system can encode multiple network configurations without increasing device size or complexity. This ingenious strategy promises compact, energy-efficient alternatives to bulky, multi-component optical processors.

Furthermore, the mechanical robustness and repeatability of the movable coding units were carefully engineered to withstand millions of reconfiguration cycles without performance degradation. This durability is crucial to ensure the practicality and longevity of metasurfaces designed for continuous operational use in real-world applications, ranging from telecommunications to autonomous vehicles.

The research team also highlighted the adaptability of the movable-type coding metasurface platform to incorporate emerging materials such as phase-change or electro-optic media, which could enable electronic control of optical properties alongside mechanical repositioning. This multimodal tunability would further accelerate the integration of multifunctional metasurfaces into reconfigurable photonic circuits.

From a theoretical perspective, the framework for designing reconfigurable diffractive neural networks presented here bridges optical physics, machine learning algorithms, and mechanical engineering. It establishes new optimization paradigms where physical displacement patterns correspond to network weights, enabling joint opto-mechanical co-design for learning tasks.

Given the exponential growth of AI demands and the bottlenecks in conventional electronic hardware, this innovative approach to reconfigurable neural computation at the speed of light heralds a promising future. It sets the stage for metasurface-enabled photonic intelligence platforms that are not merely passive optical devices but active, multifunctional processors capable of adapting to diverse computational tasks instantly.

In conclusion, the multifunctional movable-type coding metasurface concept redefines the landscape of diffractive neural networks by introducing mechanical reconfigurability to a static optical computing architecture. This paradigm shift empowers scalable, versatile, and adaptive photonic AI implementations for next-generation computing applications, representing a significant milestone in the evolution of light-based neural processors.

Subject of Research: Reconfigurable optical computing and diffractive neural networks using multifunctional movable-type coding metasurfaces.

Article Title: Multifunctional movable-type coding metasurface enabling reconfigurable diffractive neural networks.

Article References:
Yu, Z., Li, X., Gu, Z. et al. Multifunctional movable-type coding metasurface enabling reconfigurable diffractive neural networks. Light Sci Appl 15, 127 (2026). https://doi.org/10.1038/s41377-026-02216-6

Image Credits: AI Generated

DOI: 10.1038/s41377-026-02216-6

Keywords: Diffractive neural networks, metasurface, reconfigurability, optical computing, photonic AI, movable-type coding, programmable metasurface, nanophotonics, adaptive optics, mechanical tuning