The mean-squared-error is not enough: Improved prediction of chaotic dynamical systems with scientific machine learning