Abstract In recent years, a variety of deep learning (DL) models for seismic phase picking have attracted considerable attention and are widely adopted in many earthquake monitoring projects. However, most current DL models pick P and S arrivals trace by trace without simultaneously considering the spatial coherence of seismic phases among different stations in a seismic array. In this study, we develop a generalized neural network named CubeNet based on 3D U-Net to properly consider the spatial correlation of individual picks at different stations and thus improve the picking accuracy. To deal with data acquired by irregularly distributed stations, seismic data are first regularized into data cubes, which are then fed into CubeNet to calculate probability distributions of P arrivals, S arrivals, and noise. In addition, a variable trace resampling method for optimizing the differential sampling points between P and S arrivals in a trace for varying array apertures is also proposed to further improve the picking accuracy. CubeNet is trained by 47,000 microseismic data cubes and then tested by three data sets from different arrays with varying apertures and station intervals. It is found that CubeNet is rather resilient to impulsive noise and can avoid misidentifying most of the abnormal picks, which are challenging for the signal-trace based phase picking methods such as PhaseNet. We believe the newly proposed CubeNet is especially suitable for processing seismic data collected by large-N arrays.