In this work, a refined Bayesian neural network (BNN) based approach with six inputs including the proton number, mass and engineered features associated pairing effect, shell isospin ``abnormal" shape staggering effect of $^{181,183,185}$Hg, is proposed to accurately describe nuclear charge radii. The new able well radii atomic nuclei $A\ge40$ $Z\ge20$. standard root-mean-square (rms) deviatio...