Phononic materials with effectively scale-separated hierarchical features using interpretable machine learning