Lightweight scale-enhanced neural network-based optimized detection for bandwidth-limited IM/DD systems

In this Letter, we propose a lightweight and efficient optimized detection scheme based on a lightweight scale-enhanced neural network (LSE-NN) for receiver digital signal processing (DSP) in bandwidth-limited intensity modulation and direct detection (IM/DD) transmission systems. The proposed LSE-NN-based scheme comprises three main components: input preprocessing via 2-tap scale-enhanced low-pass filtering, an NN-based lookup table (NN-LUT), and a neural network-based log-maximum à posteriori (MAP) decoder with a fixed number of surviving states (NN-MAP). The scheme is implemented in a 122-Gbps optical IM/DD system employing 4-level pulse amplitude modulation (PAM-4) over a 2-km transmission distance, with a 10-dB bandwidth of 15 GHz. Experimental results demonstrate that the proposed LSE-NN-MAP scheme improves receiver sensitivity by 2.4 dB while reducing computational complexity by more than 98.44% compared to conventional optimized detection, achieving the 7% forward error correction (FEC) threshold. To the best of our knowledge, this is the first demonstration of a lightweight and efficient neural network architecture employed for optimized detection in IM/DD systems.