Transition metal-phosphide electrocatalysts for lithium-sulfur batteries: experimental and machine learning study for performance improvements and management strategies

Abstract

The commercial potential of lithium-sulfur batteries (LSB) is not yet fully achieved due to slow kinetics and polysulfide shuttling affecting capacity and stability. In this study, experimental and machine learning methods were used to minimize the gap between the current state of lithium-sulfur battery technology, to its theoretical performance limit by utilizing transition metal phosphide (TMP) catalysts. A thermally synthesized nanocrystalline FeP was characterized for its electrocatalytic potential in LSB. The lithium polysulfide (LiPS) adsorption capability of the catalyst mitigates shuttle effect and increases sulfur utilization with superior catalysis of short-chain LiPS during charging and discharging. The LSB with FeP-coated interlayer achieves initial capacity of 1006 mAh/g at 0.1C, and better stability, keeping 660 mAh/g at 0.5C for 100 cycles compared to catalyst-free LSB which is attributed to the less sulfur shuttle to the anode as indicated by the post-mortem analysis. Utilizing machine learning to predict remaining useful life and identify factors that maximize it, neural network models were developed from cycling data with the gated recurrent unit model (GRU) performing over long short-term memory model with respective root mean squared errors (RMSE) of 9.07 and 11.66 cycles. The coulombic efficiency and the high-voltage to low-voltage capacity ratio in each cycle are most important in the GRU model. The features also depend on the presence of TMP catalyst, with the corresponding features improving the RMSE from the model with no catalyst features and importance analyses further showing positive contribution to the RUL of the phosphide catalysts with Fe, Ni, and Cu atoms. Machine learning is also used to identify when the first voltage drop is expected to occur within a single discharge cycle. The GRU model achieves a test RMSE of 5.9 mAh/g with the model working well on healthy batteries, and for high-rate batteries with corresponding error of less than 10 s. This allows for better utilization of the stored energy especially for voltage-sensitive applications. Combining the catalytic properties of TMP with machine learning, speeds-up catalyst development and battery assembly optimization for more specific uses. The potential can be further maximized by making the models robust to extreme cases and novel materials.