中心博士生徐绍夫的工作——“用于宽带复杂信号接收的深度学习赋能光子模数转换器”的相关成果近期被Optics Letters期刊接收，该工作得到了国家重点研发计划（2019YFB2203700）、国家自然科学基金（61822508，61571292）的部分资助。该工作基于深度学习赋能光子模数转换器基本概念，提出一种多级下采样结构的全卷积神经网络（即“MU-net”）。利用该神经网络，可以在输入信号为宽带复杂信号的情况下，对光子模数转换器的非线性与失配缺陷进行校正，输出高质量信号。该工作以宽带回波接收为例，实验验证了MU-net对于宽带复杂信号恢复的可行性。结果表明，MU-net可以将信号的无杂散动态范围（SFDR）提升约20dB。并且，输入信号功率大范围变化的情况下，MU-net依然可以正常输出高质量信号。由于宽带复杂信号广泛存在于实际应用环境下，该工作的成功实现将深度学习赋能光子模数转换器概念推广到实际应用中。
摘要: We propose and demonstrate a modified deep-learning-powered photonic analog-to-digital converter (DL-PADC) where a neural network is used to eliminate the signal distortions of the photonic system. This work broadens the receiving capability from simple waveforms to complicated waveforms via implementing a modified deep learning algorithm. Thus, the modified DL-PADC can be applied in real scenarios with wideband complicated signals. Testing results show that the trained neural network eliminates the signal distortions with high quality, improving the spur-free dynamic range by ~20 dB. An experiment for echo detection is conducted as an example which shows that the neural network enhances the quality of detailed target profile detection. Furthermore, the modified DL-PADC only comprises a low-complexity photonic system which obviates the requirement for redundant hardware setup while maintaining the processing quality. It is expected that the modified DL-PADC can perform as a promising photonic wideband signal receiver with low hardware complexity.