Parsimony as the ultimate regularizer for physics-informed machine learning