Machine learning assisted abstraction of photonic integrated circuits in fully disaggregated transparent optical networks

Optical networks are fast evolving towards full disaggregation and softwarization down to layer-0: the data transport layer. Moreover, network elements are progressively exploiting photonic integrated circuits (PICs) to perform complex functions at the photonic level. Thanks to the advanced simulation tools, also the behavior of photonic integrated circuits can be abstracted and used within the SDN paradigm for network planning and management, permitting a full network disaggregation and softwarization down to below layer-0. To this aim, one of the main issues is the need for exact knowledge of the physical parameters of integrated circuits. In this work, we use machine learning techniques to deliver an augmented knowledge of the physical parameters of integrated circuits to be used for their full and accurate softwarization. We consider a performance prediction problems applied to a switching component. Overall results as well as data sets for machine-learning training are obtained by leveraging the integrated software environment of the Synopsys Photonic Design Suite.