By Lauren LeCours

DRESDEN, Germany, Sept. 5, 2025 — With AI, big-data analytics, and cloud computing driving exponential growth in global data volumes, today’s single-mode fiber (SMF) networks are approaching their performance limits. Space-division multiplexing in multimode fiber (MMF) is a leading candidate for the next-generation bandwidth breakthrough; unlike SMFs, a single MMF can carry many orthogonal transverse modes in parallel.

However, random mode coupling during propagation mixes these modes into complex speckle patterns, which complicates signal recovery. Although conventional digital signal processing (DSP) algorithms are theoretically capable of mode demultiplexing, their computational complexity scales rapidly with the number of modes, rendering them impractical for high-capacity MMF networks. To enable the implementation of space-division multiplexing in MMF systems, efficient and accurate mode decomposition is essential.

Jürgen Czarske, a professor at TU Dresden, has devised a solution to the challenge of signal recovery, working with his team from the Chair of Measurement and Sensor System Techniques. The developed solution is a field-programmable gate array (FPGA)-accelerated deep-learning mode-decomposition engine.

Researchers trained the FPGA engine, and, specifically, its custom convolutional neural network on synthetic data sets, and then ran simulations to calculate the contributions of different spatial modes.

According to Qian Zhang, one of the two first authors on the study, the goal of this training was to infer each mode’s amplitude and relative phase directly from a single intensity image, eliminating coherent detection. The researchers used 50,000 data points for three modes, and said that for five or six modes, they would have required the use of 60,000-80,000 data points.

They found that quantizing the network and mapping it onto a low-power FPGA slashed inference latency and energy consumption. The team achieved >100 fps at just 2.4 W, compared with tens of watts for GPU-based solutions.

“FPGA is often [used] for automation techniques, but here you can also really do the evaluation of mode conversion during the propagation of light in the fiber,” Zhang told Photonics Media. “The advantages are at least twofold — less energy and lower latency.”

---

Schematic of the experimental setup: a spatial light modulator (SLM) generates tailored mode superposition. The MMF scrambles the field. A camera is used to capture the speckle patterns emerging from MMF, which is streamed to an on-board FPGA implementing the quantized CNN for real-time mode decomposition. Courtesy of TU Dresden.

---

To further validate their engine, the researchers built a testbed out of a spatial light modulator (SLM), a precision six-axis fiber coupling stage, MMF, and a high-sensitivity IR camera . Real-time FPGA inference reliably extracted the complex field of up to six spatial modes with reconstruction fidelities >97%, paving the way for closed-loop adaptive optics, ultra-dense space-division multiplexing links, and low-latency fiber-sensor interrogators.

Importantly, the work eliminated phase ambiguity by fixing one mode's phase and restricting the relative phase of others. There are two kinds of phase ambiguity: global phase ambiguity, where the absolute phase can be shifted uniformly and still represent the same speckle patterns, and complex conjugate ambiguity, where the same intensity distributions can arise from phases with opposite phase signs. To manage these ambiguities, the team fixed the phase of the first mode to zero, which removed global phase ambiguity. Then, for the second mode, the team members restricted the phase value to a certain range, which helped select one solution out of the two potentially equivalent solutions that differ only by the phase signs.

Simply, by eliminating the global-phase ambiguity that typically plagues intensity-only training data, the researchers’ system and approach ensured a unique and physically meaningful output even when the overall phase drifts. Due to the low latency, the FPGA-based approach is ideal for closed-loop processes, according to the researchers. This technology can be used in communications, specifically free space communications, processing, and energy-efficient communication. Since FPGAs offer a higher throughput compared to traditional processors, the work opens doors for new security techniques.

The researchers are looking to apply their innovation in/for internet systems and biomedical applications, such as endoscopy. Due to the varying applications and implications for its research, the team has a variety of directions planned, especially after receiving a €1.5 million ($1.7M) Reinhart Koselleck Excellence Project award to implement this work into hospitals, specifically, for use in endoscopy procedures.

“At the moment, we are working on AI to help make photonics better,” Czarske told Photonics Media. “But we are also working on the opposite — really, to use photonics to make AI better. We want to see more transfer to applications.”

The research was published in Light: Advanced Manufacturing (www.doi.org/10.37188/lam.2025.031).

---

There are 460 suppliers of Optics in the Photonics Marketplace.

---