Fast machine learning models of electronic and energetic properties consistently reach approximation errors better than DFT accuracy

   Abstract