Single Metasurface Performs Optical Differentiation and High-Resolution Imaging at Once

Artificial intelligence has placed unprecedented demands on the hardware that moves and processes information. Every training cycle for a large language model or an image recognition algorithm depends on the rapid manipulation of enormous datasets, and the energy cost of doing that in digital electronics has become a defining constraint of the field. Against this backdrop, all-optical computing has re-emerged as one of the most compelling alternatives in modern photonics. By performing mathematical operations directly on light as it propagates, optical systems can in principle deliver processing speeds limited only by the propagation of photons themselves, while consuming far less power than conventional processors and carrying out many operations in parallel. A team of researchers in China, the United Kingdom and South Korea has now reported a device that brings this vision a significant step closer to practical reality, combining two functions that have long been pursued separately: high-resolution imaging and arbitrary-order optical differentiation, integrated into a single flat optic.

The new work, published in Light: Science & Applications, comes from a collaboration led by Professors Xinliang Zhang and Cheng Zhang of Huazhong University of Science and Technology, working alongside collaborators at the University of Cambridge and Kyung Hee University. Their device belongs to a class of components known as metasurfaces, which are engineered arrays of subwavelength structures that shape the wavefront of light in ways that bulk optics cannot easily achieve. In this case, the team fabricated a dielectric metasurface that acts simultaneously as an imaging lens and as an optical differentiator, a component that computes the spatial derivative of an optical field in real time. The achievement lies in doing both at once, on the same layer of material, without the cascades of lenses and processing stages that earlier designs required.

To understand why this is technically significant, it helps to consider what optical differentiation actually means. In Fourier optics, taking the derivative of an image is equivalent to multiplying its spatial frequency spectrum by a factor proportional to frequency. Conventional implementations of this operation require a pair of lenses to move between real space and Fourier space, with a filter placed at the intermediate focal plane. That architecture works, but it is bulky, alignment-sensitive and difficult to scale. Previous metasurface differentiators inherited some of these burdens, typically needing Fourier transform lens pairs or supplementary imaging optics to deliver a usable result. The new device sidesteps that complexity by engineering the point spread function of the metasurface itself, the characteristic pattern into which a point source of light is spread by the optical system, so that the desired differentiation operation is embedded directly into the imaging process.

The key innovation is spin multiplexing. Light can carry spin angular momentum in two forms, corresponding to left-handed and right-handed circular polarization. By designing the metasurface so that it responds differently to each spin state, the researchers gave a single flat optic two independent operational channels. One polarization channel performs zeroth-order differentiation, which is functionally equivalent to conventional bright-field imaging and preserves the original scene information, while the other channels execute higher-order differentiation that emphasizes high-frequency content, highlighting the regions of a scene where the light field changes rapidly, such as edges, boundaries and fine textures. The team demonstrated two versions of the device, one handling zeroth and first orders and the other handling second and third orders, all within the same single-layer architecture.

The performance figures reported by the team underline the practical maturity of the approach. The metasurface differentiators operated across a broad wavelength range extending from yellow light into the near-infrared, a span that matters for real-world applications where illumination sources vary and biological samples respond differently at different wavelengths. At the same time, the devices maintained fine spatial resolution of up to 228.0 line pairs per millimetre, corresponding to a line width of 2.19 micrometres. That combination of spectral breadth and resolving power in a single flat component is what distinguishes the work from earlier proof-of-concept demonstrations, which often achieved the computing function at the cost of imaging quality or vice versa.

To show that the platform works on real targets rather than only on synthetic test patterns, the researchers imaged two very different kinds of objects. The first was an amplitude-type object, a custom-made binary metallic pattern fabricated on a coverslip, which modulates the intensity of transmitted light. The second was a phase-type object, a transparent diatom cell, which leaves the intensity of light essentially unchanged and instead imprints information in the phase of the transmitted wavefront. Phase objects are notoriously difficult for conventional cameras to capture, because standard detectors register only intensity. The ability of the differentiator platform to extract useful contrast from such samples points toward applications in biological microscopy, where much of the structural information in unstained cells lives in phase rather than amplitude.

The contrast with conventional image processing pipelines is central to the significance of the work. In a standard architecture, an imaging system first captures the intensity of a scene, and the resulting digital images are then processed electronically, with edge detection, sharpening or feature extraction performed by algorithms running on processors. Each step in that chain costs time and energy, and the phase information of the light field is lost at the moment of detection. The all-optical differentiator instead performs the mathematical operation directly on the light field itself, before any detection takes place. The result is instantaneous multi-dimensional acquisition and processing, in which enhanced details emerge in the optical signal without any intermediate digital computation. The researchers also demonstrated that the device remains robust under high-intensity illumination, an important consideration for deployment in demanding imaging environments.

Real-time capability was validated in a living system. The team used the platform to observe live Euglena cells, showing that the metasurface can keep pace with dynamic biological samples rather than being restricted to static targets. This matters because many of the most valuable applications of computational imaging, from tracking motile microorganisms to monitoring cellular responses to stimuli, require the imaging and the computation to happen faster than the scene itself changes. An optical element that computes as it images has no inherent processing latency, so the temporal resolution of the system is set by the optics and the detector rather than by a downstream processor.

“This work provides a practical pathway toward ultracompact all-optical computing devices,” says Professor Cheng Zhang. “By embedding computation directly into the imaging process, we bypass the speed and energy bottlenecks of digital electronics while simultaneously capturing phase information that is inaccessible to conventional cameras.” The statement captures the dual promise of the approach: it addresses the throughput and power constraints that limit electronic processing, and at the same time it recovers information, namely the phase of the optical field, that a standard intensity camera simply discards.

The potential applications extend across several domains that share a common need for fast, parallel processing of optical information. Real-time biological imaging could benefit from a compact optic that highlights cellular boundaries and internal structure without staining or digital post-processing. Material inspection and machine vision systems, which depend heavily on edge and texture detection, could integrate the differentiator directly into their front-end optics, reducing the computational load on downstream electronics. Looking further ahead, the team anticipates that the same point spread function engineering strategy could be extended to implement a broader repertoire of optical computing functions, such as denoising or feature extraction, expanding the toolkit of meta-optics for intelligent imaging and sensing. As artificial intelligence continues to push against the limits of digital hardware, flat optics that compute while they image offer a glimpse of how the front end of machine vision may itself become a computational element.

Subject of Research: Spin-multiplexed point spread function engineering in dielectric metasurfaces for simultaneous optical differentiation and high-resolution imaging

Article Title: Spin-multiplexed point spread function engineering via dielectric metasurface for simultaneous optical differentiation and high-resolution imaging

Article References: Spin-multiplexed point spread function engineering via dielectric metasurface for simultaneous optical differentiation and high-resolution imaging. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: metasurface, all-optical computing, optical differentiation, point spread function, spin multiplexing, high-resolution imaging, dielectric metasurface, machine vision, biological imaging, phase imaging, nanophotonics, Light Science & Applications