Sequence-aware degradation of perovskite photovoltaics: separating rapid light-edge memory from thermo-mechanical cycling fatigue

Stability assessments of perovskite solar cells are commonly summarized by cumulative illumination time, thermal dose, or cycle count. Such descriptors are sequence-blind: they do not distinguish exposure histories with the same integrated stress but different ordering and spacing of light, dark, and temperature transitions. Here, two independent public datasets are re-analysed to determine which sequence-aware model structure is supported beyond in-sample fitting. In a rapid light–dark experiment, devices subjected to identical totals of 10 h light and 10 h dark displayed a 19.50 percentage-point range in normalized power-conversion-efficiency retention across 30–120 s half-cycles. Dose-only, transition-count, nonlinear count, and transition-memory models were compared using leave-one-cadence-out validation. A parsimonious two-parameter edge-memory law was selected, with pooled predictive RMSE of 0.0117 and worst-fold RMSE of 0.0155. A nonlinear transition exponent marginally improved the all-data fit but worsened grouped prediction. A long-period external consistency check showed that the rapid-edge model predicts an essentially negligible penalty for 12 h isothermal cycling, while it failed when light cycling was coupled to 25–55 ∘ C temperature oscillations. A separate thermal-amplitude fatigue law, linear in accumulated cycle number and nonlinear in temperature amplitude, achieved pooled leave-one-temperature-out RMSE of 0.0323 over Δ T = 30 –60 K. The results reject a universal scale-invariant transition term and support a hierarchy comprising rapid light-edge memory, architecture-dependent thermo-mechanical fatigue, and an uncalibrated reversible/irreversible recovery state. The framework is intended for sequence-aware screening and experimental design; it does not provide calibrated orbital lifetime prediction.