Formation-to-IR-to-capacity: Diagnostic-informed capacity trajectory prediction for lithium-ion batteries

Accurate early prediction of battery lifetime is essential for manufacturing quality control, yet conventional approaches often require extended cycling or additional post-formation diagnostics. We propose a diagnostic-informed two-stage framework linking formation signals, state-of-charge (SoC)-resolved internal resistance (IR), and the full capacity-degradation trajectory. In Stage 1, a Random Forest–based virtual hybrid pulse power characterization model predicts the 1-, 3-, and 10-s pulse-resistance responses at 24 state-of-charge (SoC) points from formation-phase voltage, current, and capacity time series. A one-RC model is then fitted to the predicted resistance triplet at each SoC to derive a 24-point RC-R1 profile; a prescribed aging-temperature indicator was used in the primary Weng analysis, and an additional temperature-free sensitivity analysis was performed to evaluate prediction without post-formation condition information. The inferred RC-R1 is treated as a diagnostically structured surrogate associated with HPPC-derived resistance behavior rather than as a direct replacement for measured resistance. In Stage 2, a multilayer perceptron combined with a long short-term memory–based residual refinement module predicts the degradation trajectory and derives end-of-life (EOL) using a 70% state-of-health threshold. On the 40-cell Weng dataset, the proposed model achieved an R² of 0.876 for capacity-trajectory prediction and an EOL R² of 0.713 with a mean absolute percentage error of 5.93%. External representation transfer was further evaluated on an independent Cui dataset comprising 182 cells formed under 62 protocols. The frozen Stage-1 representation showed strong agreement with measured low-SoC RC-R1 values (Spearman ρ = 0.956). After Stage 2 was retrained within the external dataset, the transferred representation yielded full-cycle and EOL R² values of 0.739 and 0.538, respectively. These results support the feasibility of formation-inferred RC-R1 as a diagnostically structured surrogate for trajectory-level battery prognosis. Broader validation remains necessary before industrial deployment.