Gas-Informed Machine Learning Framework for Stage Classification and Early Forecasting of Battery Degradation.

Battery health prediction typically relies on electrical signals such as capacity, voltage, and impedance, which are chemically nonspecific and can lag the onset of failure. Here we show that operando gas evolution measured by online electrochemical mass spectrometry (OEMS) provides SOC-resolved chemical signatures that enable earlier degradation-stage identification and long-horizon capacity forecasting. Using intermittently sampled CO, CO2, and C2H4 profiles, we develop two complementary models: (i) a label-free PCA-K-means stage classifier that distinguishes linear and accelerated nonlinear degradation, and (ii) a single-shot forecaster that fuses a short capacity/SOH history with one OEMS-measured cycle to predict future capacity trajectories without autoregressive rollout. On held-out test cells, the stage classifier exhibits a 3.6% misclassification rate (5/136 OEMS-measured cycles) and flags the nonlinear transition 32.0 ± 17.2 cycles before the Bacon-Watts knee (18.0 ± 8.1% of lifetime). On held-out cells closest to the training protocol, the forecaster achieves 1.7-2.9% RMSE at 100-cycle horizons and reduces long-horizon drift relative to capacity-only baselines. These results establish gas evolution as a chemically grounded complement to electrical monitoring and motivate future integration with compact off-gas sensing for cycle-level battery health management.