Shenzhen University Team Develops Optical Computing Hardware for Integrated Computation and Display

A research team from Shenzhen University's College of Physics and Optoelectronic Engineering, led by Professor Han Zhang and Assistant Professor Songrui Wei, in collaboration with the University of Hong Kong, Southern University of Science and Technology, and Peng Cheng Laboratory, has developed a novel optical computing hardware platform. This platform, based on fluorescence matrix-vector multiplication (FMVM), integrates perception, memory, computation, and display functionalities, addressing key challenges in traditional computing systems and opening new avenues for edge AI perception systems, among other applications. Recent developments in artificial intelligence and smart sensing systems have greatly expanded their use in areas like edge computing, medical imaging, and security identification. However, these advances have also brought unprecedented challenges to underlying information processing hardware. Conventional computing systems, based on the von Neumann architecture, suffer from issues such as functional segregation, frequent data transfer, and high energy consumption. Optical neural networks (ONNs) are seen as a promising solution due to their capabilities for high-speed parallel processing and low latency. Yet, existing ONN systems often rely on separate electronic modules for result output, failing to achieve true integration of computation and display. To overcome these limitations, the Shenzhen University team introduced an FMVM-based architecture. By combining the programmability of photochromic materials and the visual output capability of fluorescence, they created a system that performs optical weight programming, neural computation, and result display all within the material itself. The FMVM architecture uses a reversible photochromic system—the spiropyran (SP)/merocyanine (MC) molecule pair—to achieve non-volatile optical weight storage and high-resolution spatial programming. In the FMVM setup, SP molecules are evenly distributed in a methyl methacrylate (MMA) substrate, creating a film that can exhibit controlled fluorescence responses. Digital light processing (DLP) technology is used to program the film by projecting specific UV patterns onto it, which adjusts the MC concentration and thus controls the fluorescence intensity. During computation, input light signals (UV patterns with varying brightness) interact with the programmed SP-MC film, eliciting fluorescence emissions that correspond to matrix-vector multiplication results. These results are then directly observable or recordable by a camera, eliminating the need for electrical signal converters or traditional display devices. The research, titled "Fluorescence Matrix-Vector Multiplication: Realization of In-Memory-Display Computing," was published in the prestigious journal Optica in July 2025, Volume 12, Issue 7. Mechanism and Principles The core of the FMVM architecture lies in the reversible photochromic behavior of the SP/MC system. When exposed to ultraviolet (UV) light, non-fluorescent SP molecules transform into fluorescent MC molecules. Conversely, visible light causes the MC molecules to revert back to the SP state. This transformation provides the basis for optical weight programming, enabling high spatial resolution, precise weight control, and retrainable capabilities. The team achieved a 5-bit (32 levels) fluorescence intensity resolution and a spatial writing precision of approximately 33 μm using microscopic techniques. The fluorescence response time is less than 10 nanoseconds, with a bandwidth exceeding 100 MHz, making it suitable for high-speed computations. Weight patterns remain stable in the dark and show minimal degradation even after multiple write-erase cycles, demonstrating non-volatile storage properties. Material Characterization Detailed characterization of the SP/MC films revealed their robust photochromic behavior and high-resolution spatial programming capabilities. Experiments showed that the fluorescence intensity could be continuously adjusted by varying the UV light intensity and exposure time, achieving a high level of precision. The stability and longevity of the weight patterns were also confirmed, ensuring reliable long-term storage. Additionally, the films demonstrated excellent display flexibility, capable of generating complex patterns such as the "Mona Lisa" and microbial images with uniform light intensity and high display quality. Applications The FMVM system was tested in a practical application involving fingerprint recognition. By performing matrix-vector multiplication operations, the system successfully identified fingerprints from three different individuals and displayed their corresponding names. Both simulations and experiments were conducted to compare the system's performance in weight programming and result output, with further summation and non-linear processing enhancing the clarity of the displayed letters. Summary This research marks the first successful implementation of a neural network using fluorescence-optical technologies, integrating perception, memory, computation, and display into a single platform. FMVM offers several advantages over traditional ONN systems: Zero ADC/DAC: Entirely optically processed, eliminating the need for analog-to-digital conversions. Low Power Consumption: No continuous power is required to maintain weight storage, ideal for edge computing scenarios. Parallel Display Output: Capable of simultaneously displaying multiple types of output results. Non-Volatile Storage: Supports repeated write-erase cycles and long-term data retention. Material Flexibility: Extendable to various fluorescent materials like quantum dots, metal-organic frameworks (MOFs), and carbon dots for multi-wavelength fluorescence responses. These advancements pave the way for future innovations in wearable technology, edge AI perception systems, natural light signal processing, and optical security recognition. The project was supported by grants from the National Natural Science Foundation of China, the Shenzhen Peacock Plan, and the Shenzhen Science and Technology Innovation Commission. Paper Information: Songrui Wei, Shangcheng Yang, Dingchen Wang, Xiao Tang, Kunbin Huang, Yanqi Ge, Bowen Du, Zhi Chen, Zhongrui Wang, Xiaojun Liang, Weihua Gui, Wen Gao, Dianyuan Fan, and Han Zhang; "Fluorescence Matrix-Vector Multiplication: Realization of In-Memory-Display Computing"; Optica 12(7), 968-977 (2025); https://doi.org/10.1364/OPTICA.555491