Kolmogorov-Arnold network for efficient equalization in short-reach IM/DD systems

Neural networks have been widely employed for nonlinear equalization in short-reach intensity-modulated direct-detection (IM/DD) systems. However, the computational complexity is a major problem which limits its real-time application. To overcome the computational inefficiency of conventional neural-network-based equalizers, this paper proposes a novel Kolmogorov-Arnold network-based equalizer (KAN-E) for efficient equalization in short-reach IM/DD systems. C-band directly modulated laser (DML)-based 20-km and 10-km IM/DD optical links are experimentally demonstrated, which transmit pulse amplitude modulated (PAM)-6 and PAM-8 signals of various data rates. The results demonstrate that the KAN-E achieves up to 65.9% parameter reduction while achieving superior bit error rate (BER) performance (supporting >40 Gbaud PAM-6/8 signal transmission) compared with the feedforward neural network equalizer (FNN-E) and the convolutional neural network equalizer (CNN-E). Building on the KAN-E architecture, we also propose a fast KAN-based equalizer (FKAN-E) to further optimize equalization efficiency. By replacing 3-order B-spline basis activation functions with Gaussian radial basis functions (GRBFs), FKAN-E eliminates complex interpolation computations, thereby significantly reducing computational overhead. Experimental results show that FKAN-E can maintain comparable BER to KAN-E while achieving substantial computational efficiency improvements. For the 20-km PAM-6 signal transmission, compared with KAN-E, FKAN-E helps reduce the training and inference time by 19.9% and 58.0% respectively. For the 10-km PAM-8 signal transmission, the corresponding reduction ratios of 27.0% and 48.4% can be achieved. The proposed high-performance low-complexity KAN/FKAN-based equalization scheme can be a promising candidate for future IM/DD receiver implementation in short-reach optical interconnects.