Scientific Achievement
This work introduces a new method to improve quantum chemical predictions of energies of large organic molecules that can also be used in combination with machine learning to predict molecular energies at a much lower cost of computer time.
Significance and Impact
This study significantly enhances the accuracy by which quantum chemical methods combined with machine learning methods can be used to screen energy properties of molecules, such as reaction energies and ionization potentials. The ability to rapidly make such predictions with greater accuracy will greatly expand computational capabilities for exploration of potential electrolyte molecules.
Research Details
- Investigated the cause of the increase in error compared to experiment in the quantum chemical method G4MP2 when applied to larger organic molecules with 10 or more non-hydrogen atoms.
- In order to address this problem we added an atom-specific correction to the method, which is dependent on the type of atom, not on the bond type. This makes the method significantly more accurate with no additional computational time.
- The new method, G4MP2A, has been used in a Δ-learning ML method to predict molecular energies with improved accuracy and reduced computational cost.
