On the definition and importance of interpretability in scientific machine learning