Recently, multi-objective genetic algorithms have been applied to real-world problems and the good results were derived. When evaluation function needs a huge calculation time to derive fitness value, high calculation cost becomes a problem. One solution to this problem is to perform the search with a small population size. With this solution, however, the diversity of the solutions is often lost. This means that Pareto optimal solutions with high diversity may not be obtained. To resolve this issue, we proposed a new diversity maintenance method using Artificial Neural Network (ANN). In this method, the converged solutions on certain points are relocated uniformly in the objective space by inverse analysis using ANN. In this paper, we improved the proposed method using clustering method before inverse analysis. This mechanism improves the approximation ability of ANN. The proposed method was introduced to NSGA-II, and its effectiveness was examined on mathematical test functions. In some test functions, the proposed method provided solutions with a high degree of diversity, even when the search is performed with a small number of solutions on the high dimensional problems. The effectiveness of the clustering before inverse analysis was also discussed through the experiments.