Error correction of calculated heat of formation

Abstract

Despite of their success, the results of first-principles quantum mechanical calculations contain various errors. These errors, systematic or random, are caused by inadequate treatment of electron correlation, incompleteness of basis sets, the Born-Oppenheimer approximation, relativistic effects or approximated exchange-correlation functionals. We propose here a Neural Network based method to reduce drastically these errors. As a demonstration, the method is applied to improve the calculated heats of formation for 156 organic molecules. We find that the systematic deviations are eliminated and the remaining random numerical errors are substantially reduced. As a result the root mean square error is reduced from 25 kcal/mol to 4.7 kcal/mol.