Based on the back-propagation principle of feedforward neural networks, we design three charge density generators considering physical factors and network structure to make the residual between theory and experiment reverse flow from the root mean square charge radii to the nuclear charge distributions. It is found that about 96% of the nuclei on the validation set falling within 2 standard deviations of the predicted charge radii for CDG-3, which shows the impressive predictive power. In particular, CDG-3 reduces significantly the error interval for radii of A Dy. For A Tb, the predictions are larger than the experimental average by almost 0.1 fm and different from DDHF. According to the Hohenberg-Kohn theorem, the charge densities are further mapped to the matter densities and binding energy, which introduces the correlation of the different observables' residuals and enhances interpretability to the residual analysis. The corrections basing charge radii residual almost provide a better description for the binding energy of the Ca isotopes excluding 47,48,49 Ca, which supports an indispensable beyond-mean-field effect near 48 Ca.