Optical Neural Network Reads Hidden Light Patterns

Researchers at the University of the Witwatersrand in Johannesburg have built a computer that runs on light instead of electricity. The optical neural network, developed in the university's Structured Light Laboratory, can recognize information hidden inside complex laser beam patterns without first taking a detailed picture of them. The team describes the device as a light-powered brain, and the research was published this month in Nature Communications.

The optical neural network was led by MSc student Hadrian Bezuidenhout and conceived and supervised by Dr Isaac Nape and Professor Andrew Forbes, with contributions from Cade Peters and Ram Nandan Kumar, according to the university. The group set out to solve a bottleneck in optical communication: shaping light into patterns is easy, but reading those patterns back is not. Their goal was a receiver that keeps pace with the transmitter.

Why shaped light is hard to read

Light can be sculpted into elaborate spatial patterns that behave like letters in an alphabet, and each pattern can stand for a different symbol. Because there is no hard limit on how many patterns can be created, a single pulse of light could in principle carry far more information than a conventional signal. The catch is that these patterns rarely keep their appearance while they travel. Turbulence, distance, and the physics of propagation distort them, so a receiver that depends on recognizing an exact picture can fail even when the message itself arrived intact. Engineers have treated these alphabets as a route to packing more data into each flash of light, which is why the reading side has become the bottleneck.

The Wits team focused on a family of patterns called optical skyrmions, whose identity is stored in a mathematical property known as topology. Topology survives distortion: the number of twists in a ribbon stays the same even when the ribbon is crumpled, and the same principle applies to the twist encoded in these beams. Reading the twist instead of the picture means the optical neural network can recover the message without reconstructing every detail of the beam.

How the optical neural network decodes a beam

Conventional approaches would photograph the beam and analyze the image with heavy computation. The new optical neural network skips that step. Incoming light passes through trained optical layers that steer different components of the beam toward different detectors, and the spot where the light lands reveals the pattern's topological identity. The device pulls out the label it needs directly, the researchers report, instead of rebuilding the full light field from multiple measurements.

That shortcut is the point. A large alphabet of light patterns is only useful if a receiver can read patterns as fast as a transmitter can create them, and the usual imaging-and-analysis route is too slow and too computationally expensive. The optical neural network does the recognition with the light itself, which keeps the process fast and simple.

What this kind of hardware could change

The clearest payoff for the optical neural network is in optical communications. Because the method relies on a property of the beam that survives distortion, it could suit laser links through open air, such as between buildings or up to satellites, where turbulence scrambles beams. The researchers also point to sensing, imaging, and quantum communications, and TechCentral notes that a trained detector could eventually be built into fixed optics that draw no power on their own.

There are still practical limits. The demonstration relied on computer-assisted training, and the tests were run under controlled conditions rather than through real atmospheric turbulence. Scienmag notes that optical losses and noisy real-world channels remain untested, so the step from the laboratory bench to working links is still ahead of the team. The open questions are whether the optical layers can be trained faster and whether the recognition holds up once real turbulence is introduced.

For now, the result is a proof that optics alone can handle the reading side of structured light, with no software reconstruction step. As Nape put it in the university's announcement, building ever larger alphabets of light patterns only matters if receivers can keep up, and this optical neural network is a way to let light do the reading itself.