Machine learning model improves battery electrolyte simulations against experimental tests

Researchers reporting in Nature demonstrated a machine learning model of lithium-ion battery electrolytes that reproduced experimental properties and revealed lithium-ion clusters missed by a conventional simulation model. Nuclear magnetic resonance measurements and quantum-chemical calculations supported the predicted structures. The model learns atomic energies from quantum-chemical calculations of gas-phase clusters, then simulates liquid electrolytes. Its calculated association energies agreed with the training-level quantum calculations within chemical accuracy, while the conventional model's errors were 10-100-times larger. The approach costs two to three orders of magnitude more per energy evaluation than the conventional model, but remains many orders of magnitude less expensive than a purely quantum-chemical approach.