Silicon-graphite composite anodes deliver higher
energy density than graphite-only designs but degrade through
distinct mechanisms that conventional machine learning models
cannot disentangle. This study introduces material-specific degra-
dation diagnosis, disentangling silicon loss (LAMNE Si) from
graphite loss (LAMNE Gr), using only seven BMS-compatible,
cycle-level summary features, without requiring destructive tear-
down or high-resolution electrochemical data. Unlike prior XG-
Boost/SHAP applications on aggregate metrics or deep learning
on synthetic curves, this paper presents the first interpretable
approach, to the authors’ knowledge, to leverage simple cycle-
level features for material-specific loss prediction. Two XGBoost
models achieve R-squared of 0.887 and 0.874 on unseen test cells.
SHAP analysis reveals a silicon capacity threshold (approximately
4250 mAh) where high capacity transitions from indicating
active silicon participation to deactivated behavior; graphite
shows monotonic relationships throughout. A significant negative
residual correlation (correlation coefficient of -0.334, p-value
of 0.015) provides data-driven evidence of competitive lithium
inventory dynamics. This framework enables field-deployable
diagnosis on standard BMS hardware.

Lithium-ion battery, silicon-graphite anode, bat- tery degradation, XGBoost, SHAP, interpretable machine learn- ing, degradation mode analysis, LAMNE Si, LAMNE Gr, field-ready diagnostics