A BP neural network of 4-12-4 was established in this paper. Taking the prediction error of training data as fitness function, the genetic algorithm with a global optimization ability was used to search for the optimal initial weights and thresholds of the BP neural network. The input parameters of the BP model consist of arc gap, flow rate, welding current and welding speed of TIG welding, while the outputs of the model include welding seam sizes, that is, the front height, front width, back height and back width. The optimized BP network model shows good generalization and prediction ability, and the prediction precision is improved significantly compared with the BP model without optimization.