of metasurfaces, constructed with either all-dielectric [1][2][3] or plasmonic [4][5][6] nanoresonators, are capable of achieving engineered phase and amplitude control at the element level and thus enable accurate wave front control with subwavelength resolution. The most widely adopted metasurface design approach includes two steps: 1) calculate the amplitude and phase masks necessary for desired functionalities, fitted to square or hexagonal grids, and 2) find meta-atoms with performance closest to the target of each grid for the final design. Accurate and efficient meta-atom on-demand design approaches remain the main challenge with metasurface designs.To design meta-atoms with maximum efficiency and accurate phase gradients, a common method is to consider structures with simple geometric shapes (such as circles, [7,8] rectangles, [9,10] H-shapes, [11,12] and plasmonic thin layers [13,14] ) and perform a parameter sweep over all dimensions to assemble a library covering the full design space. Then best-fit metaatoms are selected from the library to approximate the ideal amplitude/phase map. Beyond this brute-force approach, previous literatures have also reported metasurface designs that based on solid physical considerations, such as waveguiding analysis, [15,16] Huygens surface, [11,17] surface integral equations, [18][19][20] and Pancharatnam-Berry (PB) phase. [4,21] In Metasurfaces have provided a novel and promising platform for realizing compact and high-performance optical devices. The conventional metasurface design approach assumes periodic boundary conditions for each element, which is inaccurate in most cases since near-field coupling effects between elements will change when the element is surrounded by nonidentical structures. In this paper, a deep learning approach is proposed to predict the actual electromagnetic (EM) responses of each target meta-atom placed in a large array with near-field coupling effects taken into account. The predicting neural network takes the physical specifications of the target metaatom and its neighbors as input, and calculates its actual phase and amplitude in milliseconds. This approach can be used to optimize metasurfaces' efficiencies when combined with optimization algorithms. To demonstrate the efficacy of this methodology, large improvements in efficiency for a beam deflector and a metalens over the conventional design approach are obtained. Moreover, it is shown that the correlations between a metasurface's performance and its design errors caused by mutual coupling are not bound to certain specifications (materials, shapes, etc.). As such, it is envisioned that this approach can be readily applied to explore the mutual coupling effects and improve the performance of various metasurface designs.