Skip to content
Light Does the Math: Diffractive Neural Networks Replace DSP in OAM Communications

Light Does the Math: Diffractive Neural Networks Replace DSP in OAM Communications

University of the Witwatersrand researchers demonstrate a diffractive deep neural network that directly recognizes 16 orbital angular momentum modes at picosecond speeds, eliminating electronic DSP steps like ADC sampling, MIMO equalization, and carrier recovery while achieving over 95% accuracy.

TechVision Expert Circle
TechVision Expert Circle
TechVision Expert Circle

Introduction

On October 13, 2026, a research team from the University of the Witwatersrand (South Africa) published an experimental result: a beam of light passing through a specific optical system directly "read" information encoded in a complex optical field, without first converting the optical signal to electrical and then processing it with digital chips. While optical communication already uses light to transmit data, today most systems follow a "light → electrical → compute → data" flow: fiber delivers light to the receiver, a photodetector converts it to current, and a digital signal processor (DSP) performs heavy equalization, demodulation, and error correction. This compute burden is especially large for orbital angular momentum (OAM) multiplexing, where mode recognition can multiply chip power consumption several times.

1. Where Is the "Compute Black Hole" in Traditional Optical Communication?

A 400 Gb/s coherent link typically consumes 15–25 W at the receiver DSP, with over 40% spent on dispersion compensation and channel equalization. Adding mode multiplexing (e.g., four OAM modes) requires extra MIMO demultiplexing — separating the mixed modes before demodulation. A 4×4 crosstalk matrix demands 16 parallel adaptive FIR filters, each with dozens to hundreds of taps, roughly 3–5× the compute of a single-mode system. The root cause: photodetection collapses the light field's spatial information (amplitude, phase, polarization, OAM) into a 2D intensity distribution. Phase must be recovered via local-oscillator beating; OAM modes need complex digital algorithms. In short, optoelectronic conversion is a dimensionality reduction, and the subsequent compute is essentially "remedial" work.

2. OAM Light Fields: Information Hidden in Rotation

Orbital angular momentum is a physical degree of freedom of light. Unlike a plane-wave laser, an OAM beam has a helical phase front; the topological charge ℓ (an integer) counts the twists. Because different ℓ values are orthogonal, they can be multiplexed like wavelengths. Since 2024, labs have demonstrated simultaneous transmission of over 100 OAM modes. However, more modes increase recognition difficulty: atmospheric turbulence or fiber bending causes inter-modal crosstalk, distorting the spot pattern beyond the reach of traditional correlation detection. Deep learning helps — a CCD captures the received spot and a convolutional neural network (CNN) classifies the mode(s) with >99% accuracy in lab conditions. But a trained CNN inference requires millions of multiply-accumulate operations, with latency in the microsecond-to-millisecond range, a non-trivial cost for real-time communication.

3. Witwatersrand's Approach: Optical Neural Network for Pattern Recognition

The core idea: since a CNN essentially applies a series of linear transforms (convolutions) plus nonlinear activations, can optical elements physically implement these operations? The team built a Diffractive Deep Neural Network (D²NN) , a concept introduced by UCLA's Ozcan group in 2018, and optimized it for OAM mode recognition.

D²NN system diagram
D²NN system diagram

Workflow:

Encoding: A spatial light modulator (SLM) generates beams carrying specific OAM topological charges; each ℓ represents a symbol.

Propagation & distortion: The beam travels through free space or fiber, suffering turbulence- or bending-induced crosstalk that deforms the spot.

Optical inference: The distorted beam passes through multiple diffractive optical layers. Each layer's nanostructured pattern performs a complex-valued matrix multiplication on the incident field. Free-space propagation between layers naturally provides a diffraction effect analogous to convolution.

Decision: The final light field is detected by an array of photodetectors; the region with the highest intensity indicates the recognized OAM mode.

All "computation" occurs at the speed of light with near-zero power (excluding the laser). The diffractive layer parameters are obtained by backpropagation training in a computer using a physical diffraction model instead of mathematical matrix multiplications; once trained, they are fabricated into the optical elements and remain fixed during inference.

4. Optical vs. Electronic Computing: Which Steps Are Eliminated?

Flowchart comparison
Flowchart comparison

The traditional chain takes six steps; the optical chain takes three. Crucially, the most compute-intensive steps — MIMO demultiplexing and channel equalization — are removed. Specifically:

ADC sampling & quantization: Traditional receivers need high-speed ADCs (≥64 GSa/s), consuming 3–5 W per channel. The optical scheme only needs a simple comparator to decide "which region is brightest."

MIMO equalization: The largest compute load. Four-mode multiplexing requires 16 parallel FIR filters performing hundreds of multiply-adds per clock cycle. D²NN implements the equivalent linear transform via diffraction, reducing multiplications to zero.

Carrier recovery: Coherent detection requires a phase-locked loop to track frequency and phase offsets. D²NN uses direct intensity detection, so this step disappears entirely.

Experimental results: the optical system recognizes 16 OAM modes (ℓ = −8 to +7) with >95% accuracy, completing the classification in picoseconds — three to four orders of magnitude faster than electronic DSP.

5. Engineering Gaps Before Deployment

First, non-reconfigurability. Once fabricated, the diffractive structures are fixed. If channel conditions change (e.g., turbulence strength varies), the system cannot adapt like a DSP. Programmable metasurfaces for reconfigurable diffractive layers exist but currently modulate only at kHz rates, far below the GHz needed for real-time communication.

Second, limited precision. Fabrication limits phase precision to about 8 bits (256 levels), whereas digital neural networks typically use 16- or 32-bit floating point, restricting the complexity of tasks the optical network can handle.

Third, cascaded loss. Each diffractive layer introduces insertion loss; five layers yield ~12 dB end-to-end loss, a significant penalty for practical links.

Fourth, the simulation-to-reality gap. Models are trained in simulation, then transferred to physical hardware. Fabrication errors and environmental drift create a mismatch. 2026 work explores in-situ training using measured data to correct fabrication errors, but this adds deployment complexity.

Positive developments: ASML's latest EUV platform supports mass production of freeform optical elements, enabling semiconductor-scale manufacturing of D²NN layers. Silicon photonics processes (e.g., GlobalFoundries GF45SPCLO) can already integrate simple diffractive structures, making on-chip D²NN integration plausible.

6. Conclusion

The value of this work lies not in immediately replacing DSP chips, but in clearly demonstrating a direction: a large portion of communication compute is spent digitally "reconstructing" physical information that the light already carries. If the physics of light can perform the recognition directly, that compute is eliminated at the root. This is especially relevant for 6G visions — terahertz/optical fusion, massive OAM multiplexing, free-space optical links for satellites and drones — where channel complexity and mode counts will exacerbate the electronic compute bottleneck. Letting light "read itself" may become a necessity rather than a luxury. However, moving from lab demo to engineered product requires simultaneously achieving reconfigurability, high precision, low loss, and manufacturability — likely 5–10 more years of process iteration. The direction is set: instead of forcing electronic chips to chase the speed of light, let light do the work.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

optical computing6G communicationsdiffractive deep neural networkDSP offloadingfree-space opticsoptical communicationsorbital angular momentumUniversity of the Witwatersrand
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.