Inverse-designed subwavelength metamaterial 3×3 reconfigurable power distributor for high-density programmable photonic computing

Inside the silicon chips that power our modern world, light is increasingly replacing electricity as the carrier of information. This shift, known as photonics, promises computers that are faster and more energy-efficient because light can move data without the heat and resistance that slow down electronic signals. However, a fundamental bottleneck has held back the full potential of these optical systems. For decades, the basic building blocks used to route light on a chip have been limited to simple switches that can send a beam to only one of two possible paths. This two-way constraint forces engineers to build complex, sprawling networks of components to perform even modest calculations, consuming valuable space and limiting how much computing power can be packed onto a single chip. As the demand for high-speed data processing grows, the need for a more compact and flexible way to direct light has become urgent.

A team of researchers at Huazhong University of Science and Technology has now demonstrated a new approach that breaks this two-way limit. They have created a tiny, reconfigurable device that can split a single beam of light into three separate paths and adjust the amount of light sent to each path at will. Unlike previous devices that were fixed in their behavior once manufactured, this new component can be dynamically tuned using electrical signals. By applying specific voltages to tiny heating elements on the chip, the researchers can change the phase of the light entering the device, effectively deciding exactly how much power goes to each of the three output channels. This capability transforms the device from a simple splitter into a programmable tool that can perform complex mathematical operations directly with light, a crucial step toward building dense, high-performance optical computers.

The secret to this achievement lies in a design strategy called inverse design, combined with a material structure known as a metamaterial. Instead of starting with a familiar shape and tweaking it, the researchers used a computer algorithm to work backward from the desired result. They asked the computer to find the most efficient arrangement of microscopic silicon features to achieve a specific goal: taking light from three inputs and distributing it to three outputs in any ratio the user chooses. The resulting structure is a metamaterial, which is a man-made material engineered to have properties not found in nature. In this case, the device is covered in a pattern of incredibly fine grooves, smaller than the wavelength of the light itself. These grooves act like a sophisticated lens, bending and shaping the light waves with extreme precision. The entire active area of the device is so small that it would fit comfortably on the tip of a human hair, occupying a footprint of only 4 by 1.9 micrometers.

To prove that this theoretical design works in the real world, the team fabricated the chip and tested its performance. They fed light into the device and adjusted the voltage on the heating electrodes to see if they could control the output. The results were striking. The device successfully routed light to all three output ports, and by changing the voltage, the researchers could continuously adjust the power distribution. They were able to send anywhere from a tiny fraction of a percent to nearly the entire beam to any single port, with the remaining light distributed among the others. For instance, they could direct 91.8% of the light to the middle port while sending very little to the sides, or balance the light almost equally across all three paths. This level of control was achieved without significant loss of signal strength, proving that the light was being reshaped efficiently rather than just absorbed or scattered.

Beyond simply moving light around, the researchers showed that this device could act as a brain for a simple optical computer. They used the chip to classify different types of iris flowers based on their physical measurements, a classic problem in machine learning. By converting the flower data into specific voltage settings for the device, the light exiting the chip naturally sorted itself into the correct category. When tested on a set of flower samples, the device correctly identified the type of iris in nearly 89% of the cases. This demonstrated that the hardware itself could perform the logic of a neural network, making decisions based on the physical properties of light rather than digital code. The team also simulated a larger version of the device with four input and output ports. In these computer simulations, the four-port version successfully recognized handwritten numbers with an accuracy of 97%, suggesting that the design can be scaled up to handle more complex tasks.

The implications of this work extend far beyond flower classification. By proving that light can be split and tuned into three or more paths within such a tiny space, the researchers have removed a major barrier to scaling up optical computing. Traditional optical systems rely on bulky components that take up too much room to be practical for large-scale integration. This new device, with its ultra-compact size and ability to be reprogrammed on the fly, offers a path to packing millions of these logic units onto a single chip. It suggests a future where optical computers can be built with a density and flexibility that rivals, and potentially surpasses, the most advanced electronic processors. The study confirms that by combining advanced design algorithms with the unique properties of metamaterials, it is possible to create a new generation of photonic tools that are not only smaller but also smarter, capable of performing complex calculations directly in the flow of light.