Neuro-inspired all-optical signal equalization

Short-reach optical communication systems including those in data centers, Fiber-to-the-Home (FTTH), and 5G networks are under growing pressure to support higher data rates while staying within tight power and cost budgets. Today, this scaling heavily depends on DSP, which can account for 30–50% of transceiver power and scales poorly as data rates rise. As the ICT sector already consumes around 1,000 TWh of electricity per year, and the short-reach transceiver market grows toward $25B by 2029, the need for a scalable low-power alternative is becoming increasingly urgent.

We propose a compact photonic machine-learning equaliser that shifts signal correction from electronics into optics. Instead of relying on power-hungry DSP, the chip uses modular, physics-based photonic neural networks to process distorted optical signals directly in the optical domain. Within a compact footprint below 2 cm², our solution will integrate multiple impairment-compensation functions, including dispersion compensation, and SPM/XPM mitigation.

By leveraging the native parallelism of photonics, the same hardware will be able to support WDM signals, coherent modulation formats, and multi-level modulation formats without duplicating the architecture. The result will be an adaptable and interpretable photonic equalisation platform that can deliver robust performance, orders-of-magnitude lower power than DSP, lower cost and complexity than coherent solutions, and natural scalability to higher data rates