Programmable Photonic Chip Tames Modal Dispersion in Record-Breaking Fiber Transmission

Global data traffic is growing at a pace that threatens to outstrip the capacity of the optical fibers that form the backbone of the internet. For decades, engineers have squeezed ever more information into single-mode fibers by combining wavelength-division multiplexing with digital coherent detection, but that strategy is now bumping against the nonlinear Shannon limit, a fundamental ceiling on how much data a single spatial channel can carry. One of the most promising escape routes is spatial division multiplexing, or SDM, in which multiple parallel spatial channels ride through a single multicore or multimode fiber. A multimode fiber offers the highest spatial information density of all, because several distinct propagation modes can each carry independent data streams through the same glass. Yet the approach has long been haunted by a stubborn physical problem: the different modes travel at different speeds, and the resulting distortion becomes brutally expensive to undo at the receiving end.

The distortion in question is known as differential mode delay, or DMD. Because each mode follows a different path through the fiber, a pulse of light launched into all modes simultaneously arrives at the far end smeared across time, with the earliest and latest components separated by hundreds of nanoseconds on long-haul links. Separating the interleaved data streams requires multiple-input multiple-output digital signal processing, MIMO-DSP, whose computational burden scales as the product of the square of the number of modes and the width of the dispersion window. For long-distance transmission, that scaling quickly becomes impractical, demanding enormous digital equalizers that consume power, silicon area and money. Previous attempts to soften the blow relied on fixed mode scramblers or static mode permutations, but these passive tricks offer limited flexibility, require laborious trial-and-error tuning, and cannot adapt as fiber conditions change or as the number of modes grows.

Now a team of researchers at NTT in Japan has demonstrated a fundamentally different way of attacking the problem, reported in Nature Photonics. Instead of accepting the fiber’s transmission matrix as fixed and compensating for the damage entirely in the digital domain, they installed programmable photonic unitary processors at intermediate nodes along the transmission line. These processors directly reshape the optical signals in flight, optimizing the transmission channel itself rather than merely cleaning up after it. The team calls the scheme parameterized SDM transmission, and in a proof-of-concept experiment they used it to carry high-quality data over 1,300 kilometers of three-mode fiber, a distance at which conventional approaches would face crushing digital processing loads.

The heart of the demonstration is a telecom-grade programmable photonic processor built on a silica-based planar lightwave circuit platform. The chip implements an 8-by-8 Clements-type mesh of Mach-Zehnder interferometers, measuring 50 by 25 millimeters, and performs a unitary transformation on the spatial modes passing through it. Crucially, the device was engineered to meet the harsh requirements of real optical networks: it achieves a fiber-to-fiber insertion loss of just 2.1 decibels, operates across the entire telecom C band, works identically for both polarizations, and maintains a matrix fidelity above 96 percent across the full band. Those figures stand in sharp contrast to typical photonic processors, which often suffer more than 10 decibels of loss and function only at a single wavelength, making them useless for wavelength-division-multiplexed traffic.

Finding the right unitary transformation to apply at each node is itself a formidable mathematical challenge, and this is where the team’s second innovation comes in. They built a differentiable model of the entire SDM transmission system, essentially a digital twin in which every step of computing the accumulated modal dispersion, including the singular value decomposition used to extract group delays, is expressed in differentiable linear algebra. That property allows gradient-based optimization, the same backpropagation machinery that powers modern machine learning, to systematically tune the phase parameters of the interferometer mesh and drive the differential mode delay toward its minimum. Once the optimal unitary matrix is found in the digital twin, it is transferred to the physical photonic circuit at each transmission span.

Numerical simulations revealed how the approach behaves under different physical conditions. In fibers with no or weak intermodal coupling, the optimized unitary transformations suppressed the accumulation of modal dispersion far better than random mode mixing, which merely slows dispersion growth to a square-root-of-distance scaling. In strongly coupled fibers, the simulations showed, no effective unitary conversion exists and the optimization converges toward the random-mixing behavior, an honest boundary of the method. The team also found that deploying the processors more densely along the route further reduces the residual dispersion, and that the optimization landscape contains multiple equivalent minima, with the gradient-based search reliably finding whichever optimum lies nearest to its starting point.

The experimental system circulated 12-gigabaud polarization-multiplexed quadrature-phase-shift-keying signals through a recirculating loop containing 51.2 kilometers of graded-index three-mode fiber per span. Including the fiber loss, the mode multiplexers and the photonic processor, each span imposed about 17.4 decibels of total insertion loss, which was compensated by erbium-doped fiber amplifiers, meaning the processor slots naturally into existing amplifier-based link budgets. By sweeping a rotation angle that continuously morphs the applied matrix from the identity to a mode permutation, the researchers showed that the in-line optical processor exerts direct, predictable control over the accumulated modal delay, with experimental measurements tracking the digital twin’s predictions closely.

When the optimized unitary matrix was loaded onto the chip, the accumulated differential mode delay over the full link dropped from roughly 25 nanoseconds to about 10 nanoseconds across three representative wavelengths spanning the C band, confirming wideband, high-fidelity operation. Because the dispersion window sets the required length of the digital MIMO equalizer, that reduction translates directly into lower post-processing complexity. After 1,331 kilometers of transmission, the normalized generalized mutual information of every spatial and polarization mode remained comfortably above the 0.836 error-free threshold, with cleanly clustered constellations demonstrating that the low-loss, polarization-insensitive processor had preserved signal quality end to end.

The researchers also weighed the energy economics of their approach. Their current thermo-optic prototype draws roughly 6.4 watts for a full 8-by-8 mesh, but a three-stage multiplane-light-conversion design with optimized thermal isolation could bring that down to about 1 watt, in line with the budget of typical network nodes that consume tens of watts. Measured as energy per multiply-accumulate operation, the optimized design would achieve around 9.6 femtojoules, with silicon or silicon-nitride platforms potentially reaching below 2 femtojoules, two to three orders of magnitude better than state-of-the-art electronic processors performing equivalent matrix operations. If fiber characteristics never change, the optimized transformation could even be frozen into a passive circuit requiring zero standby power.

The implications reach beyond a single laboratory demonstration. The optimization framework is not tied to compensating modal delay alone; the same differentiable model can be extended to target mode-dependent loss, another chronic limiter of long-haul multimode transmission, potentially minimizing digital cost and maximizing capacity simultaneously. Unused ports in the interferometer mesh could be repurposed as an optical switching engine, letting a single device serve as both transmission optimizer and router. While further loss reduction, toward roughly 1 decibel or below, will be needed for longer distances and higher mode counts, the work establishes a scalable blueprint for embedding photonic computation directly into the optical layer, hinting at networks where the fiber plant itself becomes a programmable, self-optimizing machine.

Subject of Research: Programmable photonic unitary processors for modal dispersion compensation in long-haul spatial division multiplexed fiber transmission

Article Title: Programmable photonic unitary processor enables parameterized differentiable long-haul spatial division multiplexed transmission

Article References: Programmable photonic unitary processor enables parameterized differentiable long-haul spatial division multiplexed transmission. (n.d.). https://doi.org/10.1038/s41566-026-01997-x

Image Credits: AI Generated

DOI: 10.1038/s41566-026-01997-x

Keywords: spatial division multiplexing, photonic unitary processor, differential mode delay, multimode fiber, MIMO digital signal processing, planar lightwave circuit, Mach-Zehnder interferometer, gradient-based optimization, differentiable model, optical communications, wavelength division multiplexing, Nature Photonics