The search for the optimum molecules for use as electrolytes in lithium-ion batteries involves an examination of billions of potential candidates. The challenge, according to scientists at the U.S. Department of Energy’s Argonne National Laboratory, lies in the tradeoff between molecular modeling accuracy and computational cost. Fortunately, artificial intelligence and machine learning may be able to help.
One of the scientists’ tools is a computationally intensive model called G4MP2. Using this tool, the scientists accurately modeled tens of thousands of small organic molecules, as described in a recent paper. “Energies for the ˜133,000 molecules in the GDB-9 database, containing organic molecules having nine or less atoms of carbon, nitrogen, oxygen, and fluorine as well as hydrogen atoms, have been calculated at the G4MP2 level of theory,” the authors write.1 (GDB-9 refers to a database available at the Wolfram Data Repository of molecular quantum calculations describing geometric, energetic, electronic, and thermodynamic properties.)
However, those 133,000 molecules represented only a small subset of 166 billion total molecules, including large ones, the scientists wanted to evaluate as potentially suitable for use as electrolytes. Applying the G4MP2 model to that huge number of molecules would be a computationally impossible task, even for the BEBOP supercomputing cluster at Argonne’s Laboratory Computing Resource Center, which the scientists used in their research.
That’s where machine learning comes in. The authors explain that the G4MP2 energies of the GDB-9 molecules will be useful in future investigations of the application of machine learning to quantum chemical data.
A second paper describes the details. The researchers applied a less computationally intensive quantum-mechanical modeling framework based on density functional theory, which is less accurate than G4MP2. But by using the G4MP2 results, they could train the density-functional-theory model to improve its accuracy while keeping compute costs down.
“Our resulting models learn the difference between low-fidelity, B3LYP, and high-accuracy, G4MP2, atomization energies and predict the G4MP2 atomization energy to 0.005 eV (mean absolute error) for molecules with less than nine heavy atoms (training set of 117,232 entries, test set 13,026) and 0.012 eV for a small set of 66 molecules with between 10 and 14 heavy atoms,” the authors of the second paper write.2
“When it comes to determining how these molecules work, there are big tradeoffs between accuracy and the time it takes to compute a result,” said Ian Foster, Argonne Data Science and Learning division director and an author of one of the papers, as quoted at Newswise.3 “We believe that machine learning represents a way to get a molecular picture that is nearly as precise at a fraction of the computational cost.”
“The machine-learning algorithm gives us a way to look at the relationship between the atoms in a large molecule and their neighbors, to see how they bond and interact, and look for similarities between those molecules and others we know quite well,” added Argonne computational scientist Logan Ward, an author of one of the studies. “This will help us to make predictions about the energies of these larger molecules or the differences between the low- and high-accuracy calculations.”
Machine learning has demonstrated its usefulness in application areas ranging from banking to medicine. It’s applicability to the search for stable, safe electrolytes for lithium-ion batteries represents yet another.
REFERENCES
1. Narayanan, Badri, et al., “Accurate quantum chemical energies for 133,000 organic molecules,” Chemical Science, June 27, 2019.
2. Ward, Logan, et al., “Machine learning prediction of accurate atomization energies of organic molecules from low-fidelity quantum chemical calculations,” MRS Communications, September 2019.
3. “Building a better battery with machine learning,” Newswise, Nov. 26, 2019.