Accelerating Electrolyte Discovery for Energy Storage through Machine Learning

Mean absolute deviations (MAE) between accurate G4MP2 energies and two machine learning (ML) methods (SchNet Delta and FCHL Delta) and density functional (B3LYP) energy calculations. Results are shown for molecules outside of the ML training set, a holdout set of 13 K of the 133 K molecules in the QM9 set.

Scientific Achievement

Utilized high performance computing to generate a database of highly accurate quantum chemical energies of 133 K organic molecules and used machine learning to enable prediction of energies from low fidelity, low cost quantum chemical calculations.

Significance and Impact

This work has established methods based on machine learning that will be able to make fast and accurate predictions of energy storage properties (redox behavior, solubility, etc) for discovery of electrolytes from the vast chemical space of millions of molecules for the JCESR Electrolyte Genome.

Research Details

  • Developed a accurate energy database of 133 K organic molecules using the G4MP2 quantum chemical method, which has an accuracy of 0.05 eV for molecular energies.
  • Identified two state of the art machine learning approaches for predicting the energies of organic molecules. The highest accuracy was achieved from approaches that learned the difference between low cost, but less accurate density functional energies and G4MP2 energies.
  • These machine learning approaches are able to predict G4MP2 energies to within 0.005 eV/mol for molecules, which will enable fast and accurate screening of properties to discover desirable molecules for advanced energy storage systems.

DOI: 10.1557/mrc.2019.107

Download this highlight 

Latest Updates

See All