To suppress the generation of separation flow on the airfoil surface and the instability of vortex shedding, a deep reinforcement learning （DRL）strategy was adopted to optimize the control of flow separation around an airfoil under low-turbulence conditions. Specifically, during the training of an active flow control policy based on the soft actor-critic （SAC） algorithm, a novel reward function was proposed. This reward function incorporated the Liutex statistic as a key metric to quantify the rotational intensity of vortex structures and vortex core size within the flow field, and was ultimately constructed by combining aerodynamic performance metrics. The training results show that the reward function based on vortex-driven has significant advantages in effectively eliminating large vortices and optimizing aerodynamic performance. The combination of the vortex-driven reward function and the DRL framework effectively improves the aerodynamic performance and minimizes fluctuations of drag and lift coefficients, demonstrating the application potential of this method in advanced flow control strategies.