In a free form optimization of a one-dimensional metagrating deflector, an electromagnetic simulation based on rigorous coupled-wave analysis (RCWA) is performed to calculate the deflection efficiency η of a given metagrating structure. In this work, we demonstrate an effective surrogate simulator architecture based on U-Net that predicts the adjoint gradient of deflection efficiency as a Figure of Merit (FoM). Our network was found to be effective, solving the forward problem with errors within 2% and the inverse problem with errors within 0.5% while increasing the computational speed by ~1000 times. Based on its high efficiency, our surrogate simulator could be reliably used to find the optimal solution of high complexity inverse design problems in periodic structures with reduced computational cost.