The interaction between turbulence and a solid surface plays crutial roles in determining momentum, heat and mass transport between the fluid and solid, and therefore its accurate prediction is of particular importance in engineering designs. With increasing the Reynolds number, however, the ratio of the spatio-temporal scales between the large-scale energy-containing eddies and the microscopic near-wall turbulence structures becomes significantly large. This makes resolving all the essential spatio-tempral scales of turbulence prohibitively expensive even with the state-of-the-art computer resources [1]. Large-eddy simulation (LES), where only large-scale eddies are directly resolved, while the effects of sugbrid-scale eddies are modelled, can be considered as a promissing approach to maintain the prediction accuracy, while reducing the computational costs. However, a universal subgrid model adapted to various flow regimes has not been established yet. In addition, subgrid models that demonstrate high accuracy in a priori test often encounter numerical instability in a posteriori test [2]. In the present study, we propose a novel approach to achieve high prediction accuracy equivalent to that of well-resolved simulation, while significantly reducing computational costs by combining turbulence big data obtained by hyperformance computing (HPC) and artificial intelligence (AI). We first conduct under-resolved and well-resolved simulations of wall turbulence and store the instantaneous snapshots of the velocity fields at different distances from the wall as training data. Then, a deep network discriminator is trained so that it can distinguish between the flow fields obtained by the under-resolved and well-resolved simulations. Once the discriminator is trained, it is integrated to under-resolved simulation and the instantaneous flow field is modified by an addtional forcing term so that the resulting velocity field is indistinguishable from those obtained by well-resolved simulation. It is shown that, even though the present inner-layer forcing model modifies only the fluctuating component of the velocity field at a certain distance from the wall, it makes the mean velocity profile within the entire boundary layer approach to that obtained by well-resolved simulation. In addition, it is also found that the present model trained at a lower Reynolds number can be applied to higher Reynolds numbers, and still maintains the superior prediction performance.