In this paper, neural networks are used to predict core characteristics, and the predicted results are used to screen poor loading patterns in order to improve optimization efficiency. The radial peaking factor, cycle length, and maximum burnup through the cycle depletion calculations were evaluated by the neural network, and these core characteristics were used for screening. The screened loading patterns were evaluated by the core calculation code as ordinary in-core optimizations. The calculation results of the test problem indicated that the loading pattern screening using the neural network effectively improves the optimization results. Since the computation time for a cycle depletion calculation with the neural network is quite short, the computation load for the screening is negligible. Since the neural network is periodically retrained using the latest evaluation results of the core calculation code, its prediction accuracy is continuously improved during the optimization. The typical prediction accuracies of the radial peaking factor, cycle length, and maximum burnup in the latter part of the optimizations were 3 to 4%, 0.01 to 0.02 GWd/t, and 0.2 GWd/t, respectively, in the test problem. These accuracies are satisfactory for loading pattern screening.