Machine learning assisted two-dimensional beam-steering for integrated optical phased arrays

Abstract

Optical phased arrays (OPAs) offer promising beam-steering solutions for ultra-small light detection and ranging applications and free-space optical communication systems. The paper presents a new machine learning assisted two-dimensional beam formation and steering technique, which calibrates phase errors via a convolutional neural network architecture to mitigate phase errors emanating from fabrication imperfections, thermal crosstalk and wavelength tuning effects. The technique is experimentally validated on an eight-channel OPA, and the results demonstrate strong control predictive capabilities with a very low mean squared error of 0.0026 and a high correlation coefficient of 0.97 after training. Reconstructed signals derived from a subset of predicted control inputs yielded a maximum lateral peak intensity deviation of a single pixel resolution (0.15°) from the ground truth, and no observable longitudinal displacement error, for a full-width at half-maximum beam-width of approximately 1.5°.