Regularized machine learning on molecular graph model explains systematic error in DFT enthalpies.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Materials research aims to discover efficient heterogeneous catalysts for chemical processes.
• Density Functional Theory (DFT) is crucial for modeling catalytic systems, but approximate functionals introduce errors.
• Errors in DFT thermochemistry calculations for gas-phase molecules are significant and poorly understood.

Purpose of the Study:

• To investigate and quantify errors in DFT-calculated enthalpies of formation for gaseous species.
• To develop a model that identifies molecular features responsible for DFT errors.
• To correct DFT thermochemistry for improved accuracy in catalyst design.

Main Methods:

• Calculated enthalpies of formation for 1676 gaseous species using two DFT levels of theory.
• Utilized 'ground truth values' from the NIST database for error assessment.
• Employed graph theory for molecular featurization and a regularized algorithm to model DFT errors.

Main Results:

• Identified specific molecular fragments that significantly contribute to DFT errors in thermochemistry.
• Developed a sparse error model that is statistically robust.
• Demonstrated that the developed model can correct DFT thermochemistry, yielding over an order of magnitude improvement.

Conclusions:

• DFT calculations for thermochemistry contain predictable errors related to molecular structure.
• A data-driven approach using graph theory and sparse modeling can effectively identify and correct these errors.
• The corrected DFT thermochemistry enhances the reliability of computational screening for novel heterogeneous catalysts.