Multiscale equations with scale separation can be approximated by corresponding homogenized equations with slowly varying homogenized coefficients (G-limits). The inverse problem of recovering the G-limits given multiscale data is a difficult taskas it is typically ill-posed. In this work, we develop an efficient Physics-informed neural networks (PINNs) algorithm for recovering the G-limits from simulated multiscale data. We demonstrate that our approach could produce desirable approximations to theG-limits and, consequently, homogenized solutions via a small amount of data. Besides, we demonstrate the robustness of our method via several benchmark examples with both periodic and non-periodic multiscale coefficients.