The thermal field in the seeded directional solidification (DS) of quasi-single crystalline (QSC) silicon needs precise control to preserve the seed and maintain a flat solidification interface. Compared with adjusting the furnace configuration and materials properties, optimizing controlling recipe parameters of the DS process is more convenient and efficient in reality for improving the thermal field. However, controlling elements in the furnace interact with each other and the cooperation effect cannot be estimated by qualitatively deducing, which adds difficulties to the design and improvement of controlling recipe. In this study, an optimization system was built to optimize the controlling recipe during crystal growth and an industrial G6 DS furnace was taken as an application example. Flattening the solidification interface, reducing the thermal stress in the growing crystal and shortening the casting time were chosen as three optimization targets. Two independent heaters and a heat gate were the three controlling elements whose recipe parameters were the optimization variables. A 2D heat transfer model along with a thermal stress model was developed and an Artificial Neural Network (ANN) was trained to study the mapping relationship between optimization variables and targets in order to save computing resource and time. Genetic Algorithm (GA) was employed to yield the optimal recipes of the three controlling elements. The results showed that the optimal recipe achieved better crystal quality compared with the original recipe. The effectiveness and efficiency of this optimization system indicates that it can achieve the optimal result that theoretic analysis may hardly access.