Scientific Achievement
This work has demonstrated that quantum chemically informed machine learning can be used to successfully predict energies of large organic molecules with sizes beyond those in the training set at a much lower cost in computer time.
Significance and Impact
The results of this study demonstrating that machine learning methods can be accurately extended beyond the size of molecules in the training set means they can be used to explore a vast exploration space of molecules for discovery of energy storage electrolytes. The ability to quickly make these predictions greatly expands the capability of the JCESR Electrolyte Genome.
Research Details
- This work is based on a set of 191 molecules with 10-14 non-hydrogen atoms (i.e., larger than those in the training set) having accurate experimental enthalpies of formation.
- The best-performing ML method investigated in this paper, FCHL-Δ, gave atomization energies for the 191 of organic molecules within about 0.4 kcal/mol of very accurate quantum chemical energies calculated by the G4MP2 method.
- This machine learning approach combined with ones being developed for anions, cations, and solvation. energies will enable fast and accurate screening of properties needed to discover molecules for advanced energy storage systems
