Gaussian Processes are a line of Bayesian probabilistic models, which are popular in regression with uncertainty problems. The requirement of manually selected kernel function and extensive computation prevent the framework scale to universal and large-scale problems. In this work, we aim to propose a universal kernel design to handle versatile time-series data named grid spectral mixture kernel. We rely on a set of kernels sampled from kernel hyper-parameter space and learning the weights to combine them. Using an ADMM solver, we achieve the sparse combination of kernels, which in turn provides an interpretation of the dataset.