Compact early-to-middle-cycle descriptors reveal empirical charge–discharge imbalance as a predictor of long-horizon lithium–sulfur battery degradation

Long-horizon prognosis of lithium–sulfur batteries is difficult because capacity fading alone does not capture non-monotonic sulfur conversion, evolving polarization, and charge–discharge asymmetry. We developed a compact, interpretable artificial-intelligence framework that uses cycles 11–400 to predict cycle-1000 retention and cycle-500-to-cycle-1000 loss in 75 target-valid cells drawn from 120 long-term coin-cell datasets. Across 50 repeated leave-one-cell-out evaluations, the compact empirical representation reduced mean absolute error from 12.53 to 7.97 percentage points for cycle-1000 retention and from 9.17 to 7.29 percentage points for late-stage loss relative to a capacity-history baseline. Feature-group replacement identified empirical charge–discharge imbalance as the strongest incremental contributor, whereas capacity-morphology and knee-related descriptors provided smaller gains. Lithium–sulfur-specific plateau, polarization, and differential-capacity descriptors did not consistently reduce average error but added electrochemical context and ranking information. Shorter observation windows showed a progressive testing-time versus accuracy trade-off, while experimental-series holdout and negative-control analyses supported within-resource robustness. These results show that compact, chemistry-aware data representations can support interpretable long-horizon battery prognosis without relying on increasingly complex models.