Abstract: For broadening the engineering application of dual synthetic jet (DSJ) in the field of separation flow control, the control mechanism and law of the DSJ were studied numerically, and the RBF neural network model was constructed to describe the relationships between the control parameters and the aerodynamic parameters. Moreover, the optimal aerodynamic parameters that could be achieved under certain constraints and control parameters were searched based on improved particle swarm optimization (PSO) algorithm. In addition, Inception-V3 model was established to identify the control parameters based on the average velocity field with the purpose of adjusting the control parameters of the actuators according to the target flow field, then achieving the optimal control effects that can realize the optimal aerodynamic performance of the airfoil. Results show that, the following issues are contained in the control mechanism of the DSJ on separation flow:momentum injection effect, vortex mixing effect and suction effect. The control parameters have a significant impact on the control effects. When the attack angle is in the range of 16°~21° and 22°~24°, the non-dimensional optimal frequency F⁺ is in the range of 0.5~4.0 and 3~4, respectively. Meanwhile, the larger the separation zone is, the bigger the non-dimensional optimal frequency F⁺ is. A larger momentum coefficient lead to a more significant effects of lift promotion and drag reduction. A great generalization ability is realized by the RBF neural network model, whose maximum test error is less than 17%. The PSO optimization results show that the maximum lift coefficient and the minimum drag coefficient could be realized under different conditions, with the values being 1.793 and 0.013 respectively, with 16° ≤ a ≤ 24°, 0 < F⁺ < 4, and 0 < C_μ < 0.0954. The Inception-V3 model, whose maximum mean square error of test cases is less than 0.1023, has the outstanding ability to predict the control parameters. Apart from that, the average velocity field corresponding to the control parameters predicted by the model is greatly consistent with the original velocity field at small attack angles, but the consistency is bad at large attack angles owing to the great sensitivities to the control parameters.