High-Performance Liquid Electrolytes for Lithium Metal Batteries Discovered By Machine Learning and High-Throughput Experimentation

Abstract

High-performance liquid electrolytes for lithium metal batteries were discovered via a closed-loop machine learning and high-throughput robotic experimentation workflow. The novel workflow iterates: (1) machine-learning property prediction of a given electrolyte formulation based on a training dataset; (2) machine-learning electrolyte search for optimal candidates; (3) robotic high-throughput sample preparation and testing; and (4) training dataset updates based on new results. As a result, for the electrolytes, the lithium coulombic efficiency at 3 mA/cm2 was increased to 98.2% from the baseline of 76.2%. The full cell capacity was 98.6% at cycle 373 with one improved electrolyte. In comparison, a baseline electrolyte showed a cycle life of 96 cycles. In addition, for machine learning, the electrolyte prediction showed promising prediction accuracy vs. lab-verified data.The advancement of electric drive vehicle batteries is crucial for improving transportation energy efficiency. Batteries using high-capacity lithium (Li) metal anode hold promise due to their high energy densities to extend the driving range and lower costs. Li metal batteries (LMBs) are a technological pathway to below $100/kWh, but with high risk. Currently, an advanced electrolyte is urgently needed to stabilize lithium metal anode for longer battery life.Liquid electrolyte formulation development that discovers effective electrolyte components (including solvents, Li salts, and additives) and optimizes their contents in the electrolyte is a promising approach. The combination of machine learning and robotic high-throughput experimentation shows promise as an effective and efficient development methodology, to tackle the electrolyte formulation search space >1020 formulation candidates too large for researchers to manually study in a timely fashion. Natural language processing was developed to capture latent knowledge from materials science literature without human labelling and demonstrated the capability to recommend materials several years before their discovery. An early prediction model and a Bayesian optimization algorithm were developed to perform closed-loop optimization of fast-charging protocols for LIBs. Using machine learning coupled to a robotic test-stand, hundreds of sequential experiments were performed on aqueous electrolytes containing 6 lithium and sodium salts for conductivities, pHs, and electrochemical responses on platinum, with non-intuitive discoveries.In this work, by using a novel closed-loop machine learning and high-throughput robotic experimentation workflow, liquid electrolytes for lithium metal batteries were optimized. The primary Li CE was improved to 99.7% from the baseline performance of 76.7%. The optimized electrolyte showed high ionic conductivity and oxidation voltage as well. Coin cell validation showed that the cell capacity retention was 98.6% at cycle 373, with the coulombic efficiency of around 100%. Furthermore, 200 mAh NCA pouch cell with the optimized electrolyte showed improved cycle life and thermal stability. Regarding the machine learning algorithm, the correlation between the actual values and the predicted values for the unseen data achieved a high value of 0.986, and the r2 value was 0.79.This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Energy Efficiency and Renewable Energy under Award Number DE-SC0020872. This research used resources of the Molecular Foundry at Lawrence Berkeley Nation Lab, which is a DOE Office of Science User Facility.