Modern artificial intelligence systems, such as robotic surgery and high-frequency trading, require real-time processing of large data streams. Rapid extraction of key features is becoming critically important, yet traditional digital processors face physical limitations and cannot deliver the necessary reduction in latency or increase in throughput for today’s demanding applications.

Optical Computing as a Solution

Researchers are exploring optical computing as a promising way to overcome these constraints. Using light instead of electricity to perform complex calculations can significantly boost the speed and efficiency of data processing. One of the most intriguing approaches involves optical diffractive operators—thin plates that perform mathematical operations as light passes through them. Such systems can process multiple signals simultaneously with low energy consumption. However, maintaining stable coherent light at frequencies above 10 GHz has proven to be a major challenge.

A Breakthrough in Optical Data Processing

To address this issue, a team led by Professor Hunwei Chen from Tsinghua University (China) has developed the Optical Feature Extraction Engine (OFE2). In an article published in Advanced Photonics Nexus, they present a new method for high-speed optical feature extraction suitable for a range of practical applications.

The key innovation of OFE2 is its advanced data preparation system. Delivering fast, parallel optical signals to core components without losing phase stability is a major challenge in this field. Fiber-based systems often introduce unwanted phase fluctuations during light splitting and delay. The Tsinghua team solved this by creating a fully integrated chip-based system with adjustable power splitters and precise delay lines. This architecture converts sequential data into several synchronized optical channels. Additionally, the integrated phase matrix allows OFE2 to be easily reconfigured for different computational tasks.

How OFE2 Works

After preparation, the optical signals pass through a diffractive operator that performs feature extraction. This process is similar to multiplying a matrix by a vector: light waves interact and form focused “bright spots” at specific output points. Fine-tuning the phase of the input light directs these spots to selected output ports, enabling OFE2 to detect subtle changes in input data over time.

OFE2 operates at a frequency of 12.5 GHz and completes a single matrix-vector multiplication in just 250.5 picoseconds—the fastest result for this type of optical computation. The device has been tested in various fields.

Applications and Effectiveness

In image processing, OFE2 successfully extracted edge features, generating paired “relief and engraving” maps that improved the accuracy of image classification and organ recognition in CT scans. Systems using OFE2 required fewer electronic parameters than standard AI models, demonstrating the efficiency of optical preprocessing for hybrid neural networks.

OFE2 was also applied in digital trading, where it processed market data in real time to generate trading decisions. After being trained with optimized strategies, the device converted incoming price signals directly into trading actions, ensuring consistent profits. Thanks to its high computational speed, traders could respond almost instantly.

Future Prospects

These achievements mark significant progress in the field of computing. Shifting the most resource-intensive stages of AI processing from electronic chips to ultra-fast photonic systems like OFE2 could usher in a new era of energy-efficient, real-time artificial intelligence. Such technologies have the potential to support computationally demanding services in image recognition, medicine, and digital finance.