We propose a method for learning energy-preserving or -dissipating numerical schemes. Our approach admits the universal approximation property.