All-solid-state battery is fundamentally different from liquid electrolyte batteries. To design interface stabilities, the solid system must be described by the constrained thermodynamic ensemble rather than the conventional unconstrained ensemble.

Li dendrite penetration, and associated microcrack propagation, at high current densities is one main challenge to the stable cycling of solid-state batteries. The interfacial decomposition reaction between Li dendrite and a solid electrolyte was recently used to suppress Li dendrite penetration through a novel effect of “dynamic stability”. Here we use a two-parameter space to classify electrolytes and propose that the effect may require the electrolyte to occupy a certain region in the space, with the principle of delicately balancing the two property metrics of a sufficient decomposition energy with the Li metal and a low critical mechanical modulus. Furthermore, in our computational prediction prepared using a combination of high-throughput computation and machine learning, we show that the positions of electrolytes in such a space can be controlled by the chemical composition of the electrolyte; the compositions can also be attained by experimental synthesis using core–shell microstructures. The designed electrolytes following this principle further demonstrate stable long cycling from 10 000 to 20 000 cycles at high current densities of 8.6–30 mA/cm 2 in solid-state batteries, while in contrast the control electrolyte with a nonideal position in the two-parameter space showed a capacity decay that was faster by at least an order of magnitude due to Li dendrite penetration.

It is a common intuition from battery experts that many shape features in the voltage profile image contain abundant information related to battery performance. However, such features are often too subtle for a human to extract by eye inspection and further correlate with battery performance. Using long cycling data from hundreds of large-format pouch cells and a total of 2 million cycles tested over 1000 days, we demonstrate here for the first time that it is advantageous to accurately predict the capacity and remaining useful life in real time by learning battery voltage profile images rather than voltage values. A strategy of end-to-end performance prediction of large-format battery cells is thus demonstrated to be feasible using only a few of the previous cycles at any given time point during the cycling test. Our work paves the way toward the application of machine learning for real-time battery performance prediction and regulation for electric vehicle applications.