Photonic Reservoir Computing enabled by Active Silicon Micro-Rings

神经形态信号处理（neural morphic signal processing， NSP）是近年来兴起的一种新的信号处理算法和处理方法。其中，光通信系统因其补偿非线性损伤的能力而受益于NSP。NSP采用了一种完全不同的信号处理方法，而不是将处理任务明确地编程到数字信号处理器（DSP）或现场可编程门阵列（FPGA）中。它使用人工神经网络（ANN），训练机器学习处理任务背后的基本物理模型，并采取相应的行动。然而，以高达几百Gb/s的线路速率实现这种ML技术进行实时信号处理是非常具有挑战性的。在未来，当信号线速率（以及相关带宽）进一步呈指数级增长时，这将变得更加困难。目前可以预见，在中期，电子电路的信号带宽将限制在100千兆赫的范围内，最多可达几百千兆赫。因此，将一些信号处理任务转移到光域是可取的，目前已经有更高的多太赫兹带宽可用。光子库计算（RC）具有作为可扩展硬件实现的能力，这在其他人工神经网络中是独一无二的。在rc中，只有（人工神经）网络的输入和输出需要自适应，而不是网络本身。事实上，互连被认为是一个“黑匣子”。由于向高维状态的非线性转换是在储层内部进行的，因此输出变成了一个线性问题。本研究计划的主要目标是分析、制造并演示一种基于硅微环结构的光子储层计算机，以弥补光纤传输系统的缺陷。为了达到指定的目标，将设计和制造一个采用CMOS兼容技术的硅光子学芯片，用于实现水库计算机。为了优化环形谐振腔结构的设计，需要进行大量的数值模拟。该芯片将被表征并集成到实验系统测试平台中。

Neuromorphic signal processing (NSP) has been emerging in recent years as an alternative to classical signal processing algorithms and processes. Among others, optical communication systems can benefit from NSP due to its ability to compensate nonlinear impairments. Instead of programming the processing tasks explicitly into a digital signal processor (DSP) or field-programmable gate array (FPGA) NSP takes a fundamentally different approach to signal processing. It uses artificial neural networks (ANN), where the machine is trained to learn the basic physical model behind the processing task and to act accordingly. However, it is very challenging to implement such ML techniques for real-time signal processing at the required line rates of up to several hundred Gb/s. This will become even more difficult in the future when signal line rates (and associated bandwidths) further scale exponentially. It can be currently foreseen that the signal bandwidth of electronic circuits will be limited in the range of 100 GHz to a maximum of a few hundred GHz in the medium term. Thus, it is desirable to shift some signal processing tasks to the optical domain, where a much higher bandwidth of multiple THz is available already today.Photonic reservoir computing (RC) has the ability to be implemented as scalable hardware, which is unique among other ANNs. In RCs, only the input and output of the (artificial neural) network need to be adaptive and not the network itself. In fact, the interconnections are considered to be a ‘black box’. Since the nonlinear transformation to a higher dimensional state is done inside the reservoir, the output becomes a linear problem. The primary objective of the proposed research project is to analyze, fabricate and demonstrate a photonic reservoir computer based on a silicon micro-ring structure to compensate for the impairments of a fiber-optic transmission system. To achieve the specified objective, a silicon photonics chip using CMOS compatible technology for the realization of the reservoir computer will be designed and manufactured in a commercially available foundry. Intensive numerical simulations are required to optimize the design of the ring resonator structure. The fabricated chip will be characterized and integrated into an experimental system testbed.