Deep potentials for molecular dynamics (MD) achieve first-principles accuracy at much lower computational cost. However, their use in large length- and time-scale simulations is limited by their lower speeds compared to analytical atomistic potentials, primarily due to network complexity and long embedding time. Here, based on the moments theorem, we develop a chemical-bonding-aware embedding for neural network potentials that achieve state-of-the-art accuracy in forces and local electronic density of states prediction with an ultrasmall 16x32 neural network resulting in significantly lower computational cost.