Beam engineering is one of the most important functionalities in light detection and ranging (LiDAR). In this work, a silicon optical phased array (OPA) is employed to control the beam profile. Machine-learning-based genetic algorithm optimization is utilized to suppress the sidelobes of the far field pattern assuming the random distribution of aperiodic arrays. The optimized mainlobe position versus wavelength relationship in two-dimensional aperiodic arrays is distinctly different from prior works. Analysis was performed to show the effect of fabrication error of arrays on the side mode suppression ratio. Our study provides an effective pathway to optimize the random distributed OPAs within a controllable time frame among the vast number of parameters.