When using ML models to deal with natural science such as physics, working well+being interpretable actually implies new discovery. Working well only, on the other hand, cannot win you a Nobel Prize.