The performance of large language models, image recognition algorithms, or any other machine learning (ML) paradigms that underpin modern artificial intelligence (AI) systems fundamentally relies on efficient computation over massive training datasets. All-optical computing, with its inherent advantages of low power consumption, fast processing speed, high throughput, and parallel computation capability, has emerged as a promising approach. Integrating high-resolution imaging with on-demand, arbitrary-order differentiation on a robust, low-size, weight, and power platform is critical for the advancement of all-optical computing systems, yet remains an outstanding challenge that warrants further investigation.
In a paper recently published in Light: Science & Applications, a team led by Professors Xinliang Zhang and Cheng Zhang from Huazhong University of Science and Technology, along with collaborators from the University of Cambridge and Kyung Hee University, reports a new class of metasurface optical differentiators that overcomes these limitations. Unlike previous embodiments that either require Fourier transform lens pair or supplementary imaging optics, this new device integrates multiple optical functionalities within a single-layer architecture: it performs spin-multiplexed differentiation directly on target objects while simultaneously enabling high-resolution imaging.
The team implemented two types of metasurface differentiators, each capable of performing 0th/1st-order and 2nd/3rd-order spin-multiplexed differentiation. The devices operated over a broad wavelength range spanning from yellow to near-infrared, and achieve fine spatial resolution up to 228.0 lp/mm (corresponding to a line width of 2.19 μm).
To demonstrate practical utility, the researchers imaged both amplitude-type object (custom-made binary metallic pattern on a coverslip) and phase-type object (transparent diatom cell). The 0th-order differentiation (similar to conventional bright-field imaging) preserves the original scene information, while the higher order differentiation emphasizes high-frequency information, highlighting regions with rapid change in the light field.
In contrast to conventional image processing architectures where the intensity information of a target scene is first captured by an imaging system and the resulting images are subsequently processed by digital methods, the proposed all-optical differentiator platform provides enhanced details by performing differentiation directly over the light field, enabling instantaneous multi-dimensional light field data acquisition and processing. The team also demonstrated the device's robustness under high-intensity illumination and further validated its real-time imaging capability by observing live Euglena cells.
"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."
Potential applications span real‑time biological imaging, material inspection, machine vision, and other fields that require fast, parallel processing of optical information. The team also anticipates that the same PSF‑engineering strategy could be extended to implement a broader range of optical computing functions, such as denoising or feature extraction, further enriching the toolkit of meta‑optics for intelligent imaging and sensing.
