Abstract: Dynamic genetic algorithm back propagation(GABP) neural network modeling method is proposed with multiple precision convergence criteria and gradual increase of sample points, aiming at the problem that traditional static surrogate model has low global approximation accuracy and over selected sample points in sheet metal forming optimization. According to the global approximation accuracy and local approximation accuracy of the dynamic model, the supplement sample point adds by the maximum minimum distance and the local optimal solution respectively. Dynamic model is applied to optimization of the NUMISHEET 93 square box forming. Combining grey relational theory to convert multi-objective problems into single-objective problems and constructing iterative schemes for optimization, multi-objective optimization of square box forming is achieved. Optimization results show that the presented method effectively improve the forming quality and the calculation efficiency of square box.