When Zero-Shot Transfer Fails: Parameter-Efficient Few-Shot Adaptation of a Selective State-Space Battery RUL Model to an Unseen LFP Fleet

<title>Abstract</title> <p>Cross-chemistry zero-shot transfer has become a common claim in deep-learning battery prognostics: a model trained on one cell family is reported to generalise to another without retraining. The fragility of that claim, together with its repair, is characterised here on the HUST lithium iron phosphate (LFP) cycling dataset — 77 usable cylindrical cells and 127,179 prediction windows — using a selective state-space (Mamba) remaining-useful-life (RUL) model for which a zero-shot R² of 0.911 had previously been reported on the Oxford lithium nickel manganese cobalt oxide (NMC) pouch dataset. That earlier figure was subsequently found to be leakage-dependent (Section 1.2) and is quoted only as the claim whose limits this study documents. Applied zero-shot to HUST, the same checkpoint reaches a median per-cell R² of only 0.157, and 26 of 77 cells score below zero, which places them behind a cell-mean predictor; nominal 90% conformal intervals attain 41.2% empirical coverage. A parameter-efficient adaptation that freezes the three Mamba backbone blocks and updates 80,641 of 812,024 parameters (9.9%), trained on 61 cells and evaluated on 16 permanently held-out cells, raises the median per-cell R² to 0.610 under a strictly causal end-of-life (EOL) denominator that is available at inference. Because the adaptation normalises RUL by each cell’s own EOL, converting a prediction back to cycles requires a denominator, and the choice of denominator materially changes the reported score. Three regimes are therefore ablated: an oracle EOL, which is not available at inference and yields 0.985; a constant fleet-median prior, which yields 0.886; and a causal EOL projected only from capacity observed up to the current cycle, which yields 0.610. The causal value is reported as the result of this work. The 0.985 figure is an upper bound that cannot be achieved in deployment and is stated only to bound the gap. Under the causal regime no held-out cell falls below R² = 0.112, against a zero-shot floor of − 0.943 with 9 of 16 cells below zero. The oracle denominator is not a free win: supplied to the un-adapted model it lowers the median per-cell R² from − 0.246 to − 0.448, so the improvement is attributable to adaptation rather than to a leaked label, although adaptation and normalisation remain entangled in the headline number and must be reported separately. Conformal intervals recalibrated on the target fleet narrow from ± 236 to ± 145.3 cycles, while measured coverage moves from a dangerously optimistic 41.2% to a conservative 99.5%. Three contributions follow: a negative result that bounds when cross-chemistry zero-shot transfer may be claimed; a demonstration that roughly 10% of the parameters and 61 cells suffice to repair such a failure; and an explicit denominator ablation that is argued to belong with any RUL result obtained under per-cell target normalisation.</p>