We develop an adaptive machine-learning framework that addresses cross-operation-condition battery lifetime prediction, particularly under extreme conditions. This framework uses correlation alignment to correct feature divergence under fast-charging and extremely fast-charging conditions. We report a linear correlation between feature adaptability and prediction accuracy. Higher adaptability generally leads to better prediction accuracy, aiding efficient feature engineering. Our analysis shows that the first 120 cycles provide sufficient information for lifetime prediction, and extending data to the first 320 cycles only marginally improves prediction accuracy. An early prediction using only one feature at the 20th cycle produces a 93.3% accuracy, saving up to 99.4% computation time and repetitive tests. Our quantitative adaptability evaluation enhances prediction accuracy while reducing information redundancy via proper feature and cycle selections. The proposed framework is validated under another unseen complex operation condition with a 90.3% accuracy without prior knowledge. L ithium-ion batteries (LIBs) have been broadly deployed in consumer electronics, 1 electric vehicles, 2 battery energy storage systems, 3 and smart grid applications 4 due to their high energy density, 5 wide working temperature range, 6 and mature technology ecology. However, such batteries continuously degrade during cycling, leading to severe issues such as capacity drop, 7 temperature rise, 8 cell-to-cell inconsistency, 9,10 and shortened lifetime. For safety concerns, it is therefore essential for battery management systems to accurately predict the state of health (SOH) and the remaining useful life (RUL) of batteries. In addition, accurate knowledge of the SOH and RUL helps to evaluate the batteries for next-stage decision-making, such as repurposing in second life 11 and recycling routine selection 12 at the end of life. Therefore, prediction of the SOH and RUL is critically important throughout a battery's life, while it remains challenging due to the constantly changing operation conditions. Much previous research reported mechanism-driven and semiempirical prediction methods. For the mechanism-informed methods, a pseudo-two-dimensional model, 13 a single-particle model, 14 electrochemical impedance spectros-copy, 15,16 distribution of relaxation time, 17,18 an equivalent circuit model, 19 incremental capacity analysis, 20 and differential voltage analysis 21 are advantageous in accurately predicting microscopic degradation, such as lithium plating, 22 solid-electrolyte-interphase (SEI) formation, 23 loss of lithium inventory (LLI), 24 and loss of active materials (LAM). 25 However, the diverse operation conditions, such as dynamic charging and discharging protocols, 26 state of charge, 27 and ambient temperatures, 28 can cause significant divergence in the primary degradation mechanisms, leading to poor performance in practical use. In contrast to the mechanism-driven method, semiempirical methods are developed by assuming equivalent circuit model 29 and empirical battery degradation patterns by deliberately fitting the historical usage parameters into the