Multi-channel photonic sampled ADC with hybrid deep-learning for distortion recovery

Abstract

Analog-to-digital converter (ADC) is indispensable components in modern information systems, and photonic technologies have emerged as promising avenues for overcoming bottlenecks in ADC performance. This paper introduces a novel approach: a photonic sampled ADC integrated with a neural network that combines Convolutional Neural Networks (CNNs) and Transformers for data recovery. The proposed system utilizes a multi-wavelength optical sampling pulse train interleaved in the time domain through fiber dispersion. Sampled signals are quantified in parallel channels by electronic ADC (EADC). Through supervised training, the hybrid deep neural network extracts both local and global features of photonic system defects, enabling the recovery of distorted data. Experimental validation showcases the effectiveness of this architecture, achieving successful reconstruction of radio frequency (RF) signals at a sampling rate of 12 Giga-samples per second (Gs/s) with an effective number of bits (ENOB) of 7.13. The results underscore the potential of the proposed architecture in achieving high conversion accuracy in multi-channel photonic sampled ADCs.