Active learning aided four-mode fiber design with equalized zero dispersion for short-reach MDM optical communications

In this work, we demonstrate an active learning method for the optimized design of a few-mode fiber (FMF) with equalized zero dispersion between four modes, which can be used for short-reach mode-division-multiplexed (MDM) optical communication without multi-input-multi-output (MIMO) processing and chromatic dispersion compensation (CDC). To obtain the desired FMF, a multi-parameter design of a complex fiber structure is needed, which is usually very difficult, inaccurate, and time-consuming. The proposed active learning can utilize fewer data than the neural network to achieve improved prediction performance by selecting more valuable data. By balancing zero dispersion, equalized dispersion, and manufacturing feasibility, structure parameters of the four-ring step-index FMF supporting four modes are predicted by the active-learning-based inverse design. The standard deviation of four-mode dispersion of the designed fiber is 0.016. The total dataset is significantly reduced to 400 by using active learning and equalized zero dispersion is obtained. The equalized zero dispersion performance is characterized by using an optical parametric amplification (OPA) modal which is highly sensitive to dispersion. The broad OPA gains with high pump power and low amplification cross talk indicate that the designed FMF has low dispersion near to zero, low nonlinearity, and weak coupling for all four modes, which is highly suitable for high-speed MIMO-less and CDC-less MDM optical communications.