Abstract: Feedforward neural networks are used for modeling the behavior of complex systems. However, most of the published papers dealt with small-or medium-scale systems. One of the possible reasons is that it is too slow or impossible for the learning algorithims of large-scale neural networks to converge. In this paper, a neural network algorithm based on modified quasi-Newton method is introduced, aiming at enhancing the neural network's ability to solve the modeling problem of large scale systems. Compared with quasi Newton method, this algorithm needs less memory and the convergence speed is almost the same. The proposed methodology is applied to modeling of industrial product quality with 32-input. Simulation results show the effectiveness of the approach.