Neural network approach to quantum-chemistry data: Accurate prediction of density functional theory energies