Abstract: Welding process has a direct effect on welding fume formation rate. Establishing a prediction model of fume formation rate related to welding parameters and predicting fume formation rate of a particular welding process is of great significance to control and reduce emission of welding fumes. Due to complex factors affecting fume formation rate and their highly nonlinear characteristics, a predication model of fume formation rate of gas metal arc welding (GMAW) based on neural network was put forward. Through experimental data of fume formation rate of flux-cored wire E501T-1, back propagation (BP) neural network and Elman neural network models were established respectively, and genetic algorithm (GA) was used to optimize the two neural networks. Through verification of 15 groups of measured data, the calculated results showed that after genetic algorithm optimization, qualified prediction values of BP neural network model and Elman neural network model were increased by 6.7% and 13.4%, respectively. The mean square error of GA-BP model was about 586.21, and average absolute percentage error was about 3.01%. Both of them were the smallest among the four models, and prediction results of GA-BP were more accurate and reliable. Based on date predicted by GA-BP model, fume formation rate of different welding current and arc voltage was predicted. Under a certain welding speed and shielding gas flow rate, the minimum welding fume formation rate was obtained when welding current was about 170 A and arc voltage was about 26 V.Highlights: (1) Neural network model was applied to numerical prediction of fume formation rate, and weights and thresholds of neural network were optimized by genetic algorithm, which improved accuracy and reliability of prediction.(2) According to prediction results of the optimized model, influence of welding current and arc voltage on fume formation rate was analyzed, which provided a useful guidance for further controlling fume formation rate.