Interpretability and Generalization Bounds for Learning Spatial Physics