The partial projections of a sample set to the PLS (partial least square) space with a less noise is used as the input elements of the BPN (back propagation network) to build a “balance” neural network structure, which circumvents an overfitting shortcoming of the usual BPN to a great extent. The samples designed from an optimal region of the PLS sub-space using a nonlinear inverse mapping technique are predicted by the PLS-BPN and are selected based on their predicted target values. This method is applied to hydrogen-storage-battery materials and several samples with a probable better initial capacity are designed.