Many modern artificial intelligence (AI) applications, such as surgical robotics and real-time financial trading, depend on the ability to quickly extract key features from streams of raw data. This process is currently bottlenecked by traditional digital processors. The physical limits of conventional electronics prevent the reduction in latency and the gains in throughput required in emerging data-intensive services.
The answer to this might lie in harnessing the power of light. Optical computing—or using light to perform demanding computations—has the potential to greatly accelerate feature extraction. In particular, optical diffraction operators, which are plate-like structures that perform calculations as light propagates through them, are highly promising due to their energy efficiency and capacity for parallel processing. However, pushing these systems to operating speeds beyond 10 GHz in practice remains a technical challenge. This is mainly due to the difficulty of maintaining the stable, coherent light needed for optical computations.
To tackle this issue, a research team led by Professor Hongwei Chen from Tsinghua University, China, has engineered a remarkable solution. As reported in Advanced Photonics Nexus , they developed an optical feature extraction engine (dubbed OFE2) that performs optical feature extraction for a variety of practical applications.
A core innovation lies in the OFE2 data preparation module. Providing high-speed and parallel optical signals for optical cores operating in a coherent environment is highly challenging, as using fiber-based components for power splitting and delay introduces strong phase perturbations. The team solved this by developing an integrated on-chip system with tunable power splitters and precise delay lines. This module effectively de-serializes the data stream by sampling the input signal into multiple stable parallel branches. Moreover, an adjustable integrated phase array allows OFE2 to be reconfigured as necessary.
Once the data is prepared, the optical waves pass through the diffraction operator. The process can be mathematically modeled as a matrix-vector multiplication that performs feature extraction. The key to this operation is how the diffracted light forms a focused 'bright spot' at the output, which can be partially deflected toward a specific output port by adjusting the phase of the parallel input lights. This movement and the corresponding changes in output power allow OFE2 to effectively capture features related to the input signal's variations over time.
Operating at a rate of 12.5 GHz, OFE2 can perform a single matrix-vector multiplication in less than 250.5 ps—the shortest latency among similar optical computing implementations. "We firmly believe this work provides a significant benchmark for advancing integrated optical diffraction computing to exceed a 10 GHz rate in real-world applications," says Chen.
The research team successfully demonstrated the capability of the proposed system across diverse tasks. For image processing, OFE2 was able to extract edge features from input images, creating two complementary 'relief and engraving' feature maps. The features generated by OFE2 led to better performance in image classification and increased pixel accuracy in semantic segmentation (such as identifying organs in computed tomography scans). Notably, the AI networks using OFE2 required fewer electronic parameters than a baseline one, proving that optical pre-processing can lead to lighter and more efficient hybrid AI systems.
In addition, the team obtained similar results for a digital trading task, where OFE2 received time-series market data and proposed profitable trading actions based on an optimized strategy. In this task, traders input real-time price signals into the OFE2. After prior training, the optimally configured OFE2 generates output signals that can be directly converted into buy or sell actions through a simple decision process, achieving stable profitability. Since the entire process is executed at the speed of light, it offers a significant latency advantage, allowing profits to be captured with minimal delay.
Taken together, these results point toward a new paradigm in which the most intense computational burdens are shifted from power-hungry electronics to ultrafast, low-energy photonics, leading to a new generation of real-time, decision-making AI systems. "The advancements presented in our study push integrated diffraction operators to a higher rate, providing support for compute-intensive services in areas such as image recognition, assisted healthcare, and digital finance. We look forward to collaborating with partners who have data-intensive computational needs," concludes Chen.
For details, see the original Gold Open Access article by R. Sun, Y. Li, et al., " High-speed and low-latency optical feature extraction engine based on diffraction operators ," Adv. Photon. Nexus 4(5), 056012 (2025), doi 10.1117/1.APN.4.5.056012 .
