A research team at South Africa's University of the Witwatersrand (Wits University) has demonstrated an optical device that reads information embedded in light without reconstructing an image of its polarization distribution. The university presented the work as a "brain powered by light" on October 13. The peer-reviewed paper, by Hadrian Bezuidenhout and colleagues, was published in Nature Communications on October 7. To use special structures of light as a way to carry information that resists external disturbance, it is not enough to encode information into the light; the receiving side also needs an efficient way to read it. This work uses the propagation and diffraction of light to classify information, an attempt to reduce the image reconstruction and electronic computation that were previously required.
Why the pattern can be read without photographing it
The team worked with a light structure called an "optical skyrmion." In this kind of light, the polarization state, which describes the direction and trajectory of the electric field's oscillation, varies with position across the beam's cross-section. If this spatial distribution is mapped onto the Poincaré sphere, which represents polarization states as points on a sphere, a value called the "skyrmion number" can be defined by how many times the distribution covers the whole sphere. It shares a mathematical structure with the skyrmions studied in magnetic materials, but here the information is carried by light itself.
The skyrmion number has the property that its value does not change when the light distribution undergoes certain deformations. Because the information can still be read even if the light pattern changes somewhat, applications in communications are anticipated. However, it cannot withstand every kind of disturbance, and there are limits to the conditions under which the skyrmion number is conserved. For communications use, it must be confirmed not only that the information survives transmission but also that the receiver can read it accurately.
Traditionally, finding the skyrmion number meant measuring the polarization state at each position, for example with Stokes polarimetry, and reconstructing the entire light field. The spatial variation of the polarization state was then differentiated and integrated to calculate the skyrmion number. This is effective for analyzing the structure of light in detail, but if communications only needs the number that represents the information, the amount of data to measure and compute is excessive. The new paper identifies reducing this readout burden as its challenge.
The team therefore focused on the fact that, for the optical skyrmions studied here, the skyrmion number can be obtained from the orbital angular momentum (OAM) indices of the horizontally and vertically polarized components. Orbital angular momentum is a physical quantity related to the helical structure of a light wavefront. If the two polarization components are separated and the OAM index of each is read, the skyrmion number can be calculated without reconstructing the whole light field. The mechanism simplifies readout by extracting only the necessary features directly.
The readout is handled by a "diffractive optical neural network." A spatial light modulator (SLM) adjusts the phase of the light, and repeated propagation and diffraction guide the light to a position corresponding to its OAM index. In the experimental setup, five layers of phase patterns were placed for each of the horizontal and vertical polarizations. By measuring where the light concentrates on the output plane, the index of each polarization component can be determined, and the original information can be read from their combination.
Classifying 81 input states into 15 kinds of information
For each polarization component, the OAM indices to be identified are nine values, from −4 to 4. Combining horizontal and vertical polarization gives 9×9, or 81 input states in total. The team ran experiments classifying these into 15 skyrmion numbers, from −7 to 7. However, the 81 inputs do not correspond directly to 81 independent communication symbols. Different light patterns can have the same skyrmion number, and the receiver treats them as the same information. Being able to extract a common number from visually different light is a distinguishing feature of this classification scheme.
Nor was the network trained by memorizing the 81 input patterns individually. According to the Methods section of the paper, training used nine basis modes for each polarization component, 18 in total. If each OAM index can be identified, the skyrmion number can be obtained from their combination. The design can therefore handle combinations not given directly during training, and its foundation is a physical property of the light being targeted.
Training the network itself, on the other hand, uses an electronic computer. The phase pattern of each layer is represented by 121 Fourier components, and the coefficients are updated based on the output measured by a camera. Each phase pattern has 242 coefficients counting real and imaginary parts, and because five layers are placed on each of the two polarization paths, the total number of coefficients adjusted is 2,420. In other words, an electronic computer is used for training, but once trained, the optical system handles the conversion and classification of input light.
When evaluating identification performance, different metrics need to be distinguished. The paper reports a value of about 93% for a metric evaluating how much of the light concentrates at the correct output position. Meanwhile, in experiments with classical laser light, where the average intensity of many photons can be measured, the target skyrmion numbers were reportedly judged 100% correctly by setting a threshold or selecting the strongest output. However, this 100% is classification accuracy in an experiment with classical light, and does not mean the same success rate would be obtained with single photons.
Robust to background light; atmospheric turbulence remains a future challenge
The trained receiver was also put through a noise-tolerance test in which external white light was mixed in. Even as incoherent background light entering the detection plane was increased, classification with classical light reportedly maintained 100% accuracy until the noise level, as defined in the paper, reached −0.22 dB. However, this result applies to "isotropic noise" that broadly overlaps the detected light, and does not show tolerance to every kind of optical disturbance.
In the conventional approach of reconstructing the polarization image, fluctuations in light intensity at each position tended to lead to errors in calculating spatial derivatives. The new method does not need spatial derivatives and reads information by comparing the light intensity in each output region. Because it only needs to correctly identify the strongest signal, the classification result is less likely to change when background light is added. However, when the intensity differences between outputs become small, how accurately the camera can measure those differences becomes important. Indeed, classification accuracy dropped under conditions with even stronger background light.
The effects of atmospheric turbulence, meanwhile, are a different problem from noise caused by background light. A prior study by Guo, Peters and colleagues showed that the skyrmion numbers of classical and quantum light remained relatively stable even under experimentally reproduced turbulence. However, the study also reported that correlations such as quantum entanglement could degrade. Even if a specific topological quantity such as the skyrmion number is preserved, other properties of the light are not necessarily maintained.
For the present receiver too, the effects of mode mixing, phase disturbance, and atmospheric turbulence on readout performance are listed as subjects for future verification. Even if conservation of the skyrmion number has been confirmed under particular conditions by earlier studies, it does not follow that connecting a new receiver will read the same information accurately. At this stage, it has not been demonstrated that light that has passed through a disturbed transmission path can be read stably by this receiver.
The team also applied the method to image transmission. Color values were assigned to skyrmion numbers, light corresponding to each pixel of an image was sent in sequence, and the image was reconstructed from the information read at the receiving end. The paper reports 100% image reconstruction in this experiment. However, this is a proof of principle showing that skyrmion-number readout can be applied to image transmission, and it does not verify performance in high-speed, long-distance communication in a real noisy environment. Judging its practicality as a communications technology requires evaluating the whole system, from transmission through the channel to reception.
Optical loss and speed remain obstacles to a practical communications receiver
In considering the device's performance, the difference between light-routing efficiency and transmittance is also important. The "over 70%" figure reported in the paper represents the fraction of light reaching the CCD that was concentrated in the designated detection regions. By contrast, transmittance relative to the light input to the device was about 15±3%, meaning most of the incident light was lost before reaching the detection plane. The main figures given in the Results of the October 7 paper are summarized below.
| Metric | Value reported in the paper | Denominator and meaning |
|---|---|---|
| Output routing efficiency | Over 70% | Fraction of total light reaching the CCD that entered the designated detection regions |
| Transmittance before adding diffraction grating loss | About 30±3% | Fraction of input light reaching the CCD after reflections at the SLM and mirrors |
| Transmittance including diffraction grating loss | About 15±3% | Fraction of input light reaching the CCD, including diffraction grating loss |
These are figures measured and evaluated on the same experimental setup against different baselines, and a performance gap cannot be found by simply dividing one by another. Even if the performance of routing light to the correct detection position is high, losing much of the light on the way weakens the signal that can be received. To use it as an actual communications receiver, overall optical loss in the device must be reduced while maintaining routing accuracy.
In addition, training the networks for the two polarization paths took about 45 hours on the actual device. This time was needed to adjust each layer's phase pattern on the physical hardware, and it does not mean that 45 hours are needed each time a trained network judges information. Although using light propagation for classification has advantages, how far this device can raise communication speed will need to be measured separately.
As for power consumption, the high efficiency of the optical computation cannot simply be taken as the energy efficiency of the receiver as a whole. The authors point out that evaluation needs to include not only the laser and the SLM but also the power consumed by the camera and control circuitry. They also suggest replacing the trained phase patterns with a fixed-structure metasurface, which could simplify the device configuration if realized. However, this experiment used an SLM, and a receiver built with metasurfaces has not been completed.
The light used in this demonstration was classical laser light generated by combining specific Laguerre-Gaussian modes with horizontal and vertical polarization. Application to light using single photons or quantum entanglement remains a subject for future research. It also remains to be verified whether information can be read accurately over transmission paths affected by atmospheric turbulence and whether the required light levels and processing speeds fall within a practical range. If these challenges can be solved, the approach could lead to a new optical communications receiver that extracts only the necessary information directly, without reconstructing the polarization distribution of light in detail.