Reading Light's Topology Without A Picture

Eighty-one. That is how many distinct topological structures a group at the University of the Witwatersrand taught a passive optical network to tell apart, after training it on fifteen examples. The network has no processor in it. It is a stack of transparent layers with a carefully computed phase pattern etched across each one, and the computation happens because light diffracting through those layers ends up somewhere specific. Where it lands is the answer. The work was published in Nature Communications and announced by Wits on 13 October, and the reason it matters has less to do with machine learning than with a problem that has been sitting unsolved underneath an entire proposed communications architecture.

Structured light has been pitched as a way out of the bandwidth problem for twenty years. The idea is straightforward: a beam can carry information not only in its intensity, wavelength and polarisation but in its spatial shape, and because there are many distinguishable shapes, each symbol can carry several bits instead of one. The standard workhorse is orbital angular momentum, where a beam's phase winds around its axis an integer number of times. OAM works as an alphabet because its modes are mathematically orthogonal: a correctly built mode sorter maps each one to a different place, and a receiver can separate them cleanly.

Optical skyrmions are a newer and more interesting proposal, and they break that convenience. A skyrmion is a topological texture — in the optical case, a field in which the local polarisation state varies across the beam cross-section in such a way that, taken together, the polarisations cover the whole Poincaré sphere a whole number of times. That integer is the skyrmion number, N. It is a topological invariant, which is a precise way of saying it cannot change under smooth deformation: squeeze the beam, defocus it, push it through a disturbance, and as long as nothing tears the field, N survives. Textures of this kind are robust in a way that amplitude patterns are not, which is exactly what a communications engineer wants from a symbol travelling through a real link.

The trouble is in how you read it. N is not an orthogonal label. Fields carrying different skyrmion numbers are not necessarily orthogonal, which is why no projection, matched filter or mode sorter is guaranteed to separate them the way an OAM sorter separates winding numbers; the authors call this the non-orthogonal nature of the topological invariant. The conventional workaround is to measure the full field — Stokes polarimetry across the beam, six polarisation projections, a reconstruction of the polarisation at every pixel, then a numerical contour integral to extract N. That works in a laboratory. It is a camera, a computer and a sequence of measurements that the authors describe as highly susceptible to noise, which is not a receiver.

A diffractive optical neural network inverts the usual relationship between optics and computation. In a conventional optical machine-learning system, light carries data and electronics do the work. Here the layers themselves are the work. Each layer is a surface patterned so that every point on it imposes a particular phase delay; light passing through is scattered, and the scattered contributions from the whole surface interfere on their way to the next layer. The transformation from one layer to the next is a multiplication by that phase pattern followed by a propagation, and propagation in free space is itself a linear operation. Stack several layers and the chain of multiplications and propagations is structurally the same thing a feed-forward network does with weights and matrix products — except that the weights are etched, the multiplications are interference, and the whole thing runs at the speed of light through a few centimetres of air with no power consumed beyond the beam itself.

The Wits detector is built from two independent processing channels, each containing five diffractive layers. The design target is unusually clean: map every input topology onto a spatially separated Gaussian spot, so that reading the skyrmion number reduces to noticing which detector lit up. The non-orthogonality that defeats a mode sorter is not solved — it is routed around. The network is not projecting the field onto a basis. It is learning a transformation that happens to send topologically distinct inputs to physically distinct places, which is a weaker requirement and, as it turns out, an achievable one.

That is the step worth dwelling on. A mode sorter is designed from the mathematics of the modes it sorts; it exists only when the modes are orthogonal. A diffractive network is trained from examples, and it does not care whether the thing it is separating has a basis. It only has to be consistently different.

The practical obstacle to diffractive networks has usually been the training itself. A layer is typically parameterised pixel by pixel, which for a realistic aperture means hundreds of thousands of free variables per layer, a training run that only converges with a large dataset, and a fabricated result exquisitely sensitive to the difference between the simulated wavefront and the real one.

The Wits group parameterised the layers in a spatial mode basis instead — expanding each phase screen in a set of smooth optical modes and training the coefficients rather than the pixels. The preprint version of the work puts the saving at around a factor of a thousand in the number of training variables. Two consequences follow. The first is that fifteen training examples are enough, according to the university's announcement, which is why the result can report generalisation to structures the network never saw. The second is subtler and more useful in a laboratory: a phase screen built from smooth modes is a physically gentle object, free of the pixel-scale discontinuities that are precisely where a fabricated element diverges from its simulation.

The published title adds a phrase the preprint did not carry — "and their live training" — pointing at the other half of the contribution. A diffractive network trained entirely in simulation has to be correct the first time, because once the layers are made they are fixed. Training in the loop, against the actual optical system, lets the network absorb the system's own imperfections instead of fighting them: a misaligned layer, a slightly wrong propagation distance or an aberrated input stops being an error to be corrected and becomes part of what the weights were trained on.

Three results carry the paper. The network distinguished 81 input topologies, including structures outside its training set, at detection visibilities — equivalently fidelities — better than 93 per cent in the preprint version, with roughly 93 per cent reported for each of the two channels. It continued to work when noise was deliberately added at the detector, where the preprint puts accuracy at 100 per cent across the noise range tested; that matters more than it might appear, because the whole appeal of a topological alphabet is resilience, and a resilient symbol read by a fragile receiver is not resilient. And the group transmitted and received an image encoded in a 14-level topological alphabet with, in the authors' description, no discernible cross-talk — the end-to-end demonstration that turns a classifier into a communications link.

A 14-level alphabet is just under four bits per symbol. That is not, by itself, a competitive data rate, and nobody involved claims it is. Its significance is the ratio between what the receiver does and what it costs: four bits recovered by letting light fall through five passive surfaces onto a small array of detectors, against six polarisation projections, a polarimetric reconstruction and a numerical integral for the same four bits by the conventional route.

The bench numbers are better than one might fear and narrower than a system designer needs. The preprint works at 633 nm with a helium–neon source and reports channel efficiencies above 75 per cent — a passive diffractive stack has no gain, and every layer scatters some light into the wrong place, so three quarters of the launched power arriving at the right detector after five surfaces is a respectable result rather than a marginal one. What is absent is a measured symbol rate. The paper calls the approach high speed without saying how fast, and since the entire argument for an optical receiver is that it reads as fast as a transmitter can write, that is the number the next paper has to carry.

Two structural limits are inherent rather than unreported. The layers are passive and fixed: a network trained for one alphabet is that alphabet's detector and no other, so a system that wants to change its encoding has to change its optics. And diffraction is dispersive, so a stack computed at one wavelength will not perform the same transformation at another — which makes a wavelength-multiplexed version of this receiver a genuinely different design problem rather than a matter of adding channels.

There is also the question of the channel rather than the receiver. Skyrmion topology is robust under smooth deformation, and a turbulent atmosphere is not always smooth; strong scintillation can tear a field rather than bend it. Whether a topological alphabet survives a real free-space path is a separate experimental question from whether a diffractive network can read it on a bench, and Forbes and colleagues — including Cade Peters, a co-author on this paper — have answered part of it elsewhere, sending optical skyrmions over a 270-metre free-space link and finding the topological numbers preserved in every case, at fidelities above 98 per cent in most conditions and 86 per cent in the worst.

For most of the structured-light era the field has been generating alphabets faster than it has been able to read them. Spatial light modulators and metasurfaces made it straightforward to write almost any texture onto a beam; the receiver side stayed stuck with cameras and reconstructions, and the gap between the two is the reason structured light has remained an elegant idea rather than deployed hardware. What this work changes is which side the difficulty sits on. A passive stack that outputs its answer as a position rather than a picture is a receiver architecture, not a measurement procedure, and it arrives for the one class of structured light whose invariant was supposed to be unreadable by projection.

The cost is paid somewhere specific, and it is worth naming for anyone who would have to build this. A diffractive stack trained in the loop is a bespoke optical component: ten surfaces whose phase profiles are not analytic forms but trained solutions, that must be fabricated to the tolerance the training assumed, mounted at the separations it assumed, and aligned to each other and to the beam. That is a demanding specification for an optics shop — smooth, continuous, multi-level phase structures held to interferometric spacing across a stack — and it is exactly the kind of component that stops being exotic once somebody has a reason to make a hundred of them. On the evidence of this paper, the reason is arriving.