Interpretable models for extrapolation in scientific machine learning