Physicists in Mexico have built a system that teaches an artificial intelligence to read one of the strangest currencies of quantum light: the topological charge of entangled photons whose wavefronts corkscrew through space. Reporting in the journal Results in Optics, the team combined three deceptively simple ingredients — a tilted glass lens, a single-photon-sensitive camera and a convolutional neural network — into an automated readout that identifies the orbital angular momentum of entangled photon pairs with about 97.9 percent accuracy, even when the light signal is faint and drowning in noise. Once trained, the network sorts images at a blistering pace of 3,263 per second, a speed that matters because future quantum networks will need to measure the information encoded in photons as fast as it arrives. The work, carried out by researchers supported by Mexico’s SECIHTI, addresses a stubborn bottleneck: producing pairs of orbital-angular-momentum-entangled photons has become routine in laboratories, but reliably and rapidly determining exactly which twisted state each photon occupies has lagged far behind, constraining the technologies that depend on it.

The property at stake is orbital angular momentum, or OAM, a form of angular momentum that light acquires when its phase winds helically around a dark central point known as an optical vortex. Beams shaped this way — the so-called Laguerre-Gauss modes — are labeled by an integer topological charge that counts how many full 360-degree twists the phase makes around a single circuit of the beam’s axis. The sign of that integer records the handedness of the twist. What makes OAM so attractive for quantum technologies is capacity: while the polarization of a photon offers only a two-dimensional space of states, the OAM degree of freedom spans a high-dimensional Hilbert space in which a single photon can, in principle, carry many distinct symbols at once. Ever since photon pairs entangled in OAM were first generated through spontaneous parametric down-conversion more than two decades ago, researchers have pursued applications ranging from quantum communications and quantum key distribution to quantum tomography and the long-envisioned quantum internet, where high-dimensional encoding could multiply channel capacity and strengthen security.

The new study tackles the measurement problem head-on. Entangled pairs are born inside a nonlinear crystal through spontaneous parametric down-conversion, a process in which a high-energy pump photon splits into two lower-energy photons. Because the interaction conserves orbital angular momentum photon by photon, a pump beam carrying no OAM yields pairs whose topological charges are equal and opposite: one photon’s twist perfectly counterbalances its partner’s. Generating such states, the authors note, has become a routine procedure. Knowing what you have made, however, is another matter. The intensity pattern of a Laguerre-Gauss mode is a luminous ring whose radius hints only at the magnitude of the charge and says nothing about its sign, so a photograph alone cannot settle the question. One established technique flattens the phase and projects the photon onto a Gaussian mode through a spatial light modulator coupled to a single-mode fiber, but in doing so it discards the spatial information that makes OAM valuable. Diffraction-based methods, such as passing heralded photons through a triangular aperture, preserve more information yet demand painstaking, time-consuming alignment of the aperture with the center of the vortex — impractical when the whole point is fast readout.

The team’s solution hinges on a classical optics trick with quantum versatility: a biconvex lens tilted slightly away from the beam path. The tilt introduces astigmatism, and the lens performs what opticians call an astigmatic transformation, converting the donut-shaped Laguerre-Gauss mode into a Hermite-Gauss-like pattern of bright lobes arranged along a diagonal. The mathematics is elegant: a mode with radial index zero and topological charge magnitude |ℓ| is reshaped into exactly |ℓ|+1 lobes, while the sign of the charge decides the orientation — +45 degrees for positive values and −45 degrees for negative ones. A charge of +4 thus appears as five lobes slanting one way, and −4 as five lobes slanting the other, a fingerprint of both magnitude and handedness in a single image. Earlier work had used the tilted-lens technique on heralded single photons prepared without OAM entanglement, and separate studies had trained neural networks to classify classical light beams; the new experiment brings the two together on genuinely entangled photon pairs, under low photon flux and high noise, and recovers the sign of the charge, not merely its size.

The experiment begins with a 10-millimeter type-II PPKTP nonlinear crystal pumped by a continuous-wave laser at 405 nanometers, producing pairs of orthogonally polarized photons at 810 nanometers in a collinear geometry, with the pump focused to a waist of 250 micrometers to optimize the pair-generation rate. After filtering, a polarizing beam splitter cleanly separates the partners, which travel the same path: vertically polarized photons serve as the signal, horizontally polarized ones as the idler. The signal is magnified by a telescope and directed onto a reflective liquid-crystal spatial light modulator — a Holoeye Pluto 2.1 with 1920 × 1080 pixels — that displays holograms projecting the photon into whichever Laguerre-Gauss mode is chosen; a second telescope matches the spot to the core of a single-mode fiber, which together with the modulator acts as a mode filter. Detection of the signal photon by an avalanche photodiode produces an electronic trigger, delayed through nuclear instrumentation modules, that gates an intensified CCD camera — an Andor iStar 334T — to image the heralded idler. The idler is routed through a 28-meter optical delay line, taking about 90 nanoseconds, to compensate the camera’s internal timing, then through a biconvex lens of 250-millimeter focal length that can sit either perpendicular to the beam or rotated by 22.5 degrees to perform the astigmatic transformation. Measurements were recorded at pump powers of 10, 5 and 1 milliwatt with five-second exposures, deliberately yielding images of varying photon flux and noise.

At the heart of the analysis sits a convolutional neural network trained from scratch on the experimental images. The network accepts 128 × 128-pixel grayscale images and passes them through four convolutional layers — the first two with 64 filters and the next two with 128, each using 3 × 3 kernels, rectified linear activations and same padding — alternated with 2 × 2 max-pooling layers that shrink the spatial dimensions, and batch-normalization layers inserted after the second and fourth convolutions to stabilize training. The extracted features are flattened and fed to a fully connected layer of 512 neurons, followed by a dropout layer that randomly silences 20 percent of the connections to guard against overfitting, and finally a softmax output with 20 units — one for each topological charge from −10 to +10, excluding zero. Training used the Adam optimizer with an initial learning rate of 0.001, mini-batches of 64 images and a maximum of 20 epochs, with the learning rate reduced tenfold whenever progress stalled for three epochs and training halted early after five consecutive epochs without improvement.

The dataset comprised 8,266 images of the lobe patterns spanning all charges from −10 to +10 except zero, captured at the three pump powers, and was divided into 70 percent for training, 15 percent for validation and 15 percent for testing — 1,240 unseen test images with 62 examples per class. On a workstation equipped with an Intel Xeon 2.00-gigahertz processor, 64 gigabytes of memory and an NVIDIA RTX 3090 graphics card, training took just 55 seconds. The results were striking. During training the network reached a mean squared error of 0.0014 with perfect accuracy; on validation data the error was 0.0383 with roughly 99 percent accuracy; and on the independent test set it registered an error of 0.0712 with about 97.9 percent accuracy. The narrow gap between those figures indicates the model generalizes rather than memorizes. In deployment, the trained network classified all 1,240 test images in 0.38 seconds on average — a throughput of 3,263 images per second — producing a nearly diagonal confusion matrix in which true and predicted charges almost always coincided.

The errors, though few, were informative. For charges above four in magnitude, the confusion matrix showed at most three off-diagonal misclassifications, a shortfall the authors trace to physics rather than software: as the charge grows, the pattern must display more lobes, which become narrower and more closely spaced until they strain the fixed transverse resolution of the telescopes and the camera’s 13-micrometer pixels. Each individual lobe also dims as its energy is spread thinner, degrading the signal-to-noise ratio. Both limitations, the team notes, have straightforward remedies — refining the optical system, raising the pump power or extending the camera’s exposure to accumulate more photon-detection events into a cleaner image. What stands out most is the network’s tenacity at the edge of visibility: in test images acquired at the lowest pump power, where the lobe structure is barely discernible beneath the noise, the CNN still assigned the correct charge, demonstrating precisely the low-flux robustness that real quantum experiments demand.

The achievement reads as a template for what machine learning may bring to quantum optics. A readout that is fast, automated, tolerant of noise and sensitive to the sign of the topological charge removes one of the practical impediments to using the full spatial richness of photons as information carriers — in quantum imaging systems, in high-dimensional quantum key distribution and, eventually, in the quantum internet, where entanglement distributed between distant nodes must be verified at speed. The authors frame their system as a step toward the generalized integration of convolutional neural networks in real-world quantum protocols, a vision in which trained algorithms shoulder the pattern-recognition burden that manual optical alignment once made slow and fragile. With a tilted piece of glass, a camera that can see single photons, and a neural network that learned its craft in under a minute, measuring the twist of entangled light has moved from a laboratory patience test toward a routine, real-time operation — a quiet but consequential advance on the road to technologies that will treat every degree of freedom of a photon as a usable channel.