Constrained LLM-Guided Model Selection for Cross-Chemistry Early-Life Battery Prognostics

Data-driven battery prognostics can fail when models trained on one chemistry are deployed on another. This study investigates early-life end-of-life (EOL) classification under chemistry-driven domain shift and introduces a schema-constrained large language model (LLM) configuration layer as the primary methodological contribution. The LLM does not predict EOL directly and does not introduce a new transfer-learning or uncertainty-estimation algorithm. Instead, it maps structured chemistry context and numerical dataset evidence to an executable architecture, reliability head, and bounded set of training hyperparameters selected from an approved model zoo containing one-dimensional CNN, self-attention, WideResNet, OpenMax, and Spectral-normalized Neural Gaussian Process (SNGP) options. Full degradation trajectories define five EOL categories using an 80% state-of-health threshold, whereas model inputs are restricted to early-cycle observations. Across six directed Argonne/CAMP chemistry-transfer tasks, the target-only deterministic CNN achieved a mean balanced accuracy of 68.9%, compared with 78.3% for the LLM-selected CNN with self-attention and SNGP, corresponding to an average improvement of 9.4 percentage points. Fixed non-LLM transfer alternatives achieved 70.0% for the deterministic CNN, 71.7% for the OpenMax-enforced CNN, 66.5% for deterministic WideResNet, and 73.3% for WideResNet-SNGP. Context ablation further reduced LLM-selected accuracy from 78.3% with history and chemistry context to 73.5% without history, 71.6% without chemistry, and 67.1% when both were removed. These results support the use of constrained LLM guidance as an auditable configuration-selection layer, while cross-chemistry transfer learning and uncertainty-aware EOL prediction provide the experimental setting in which that selection mechanism is evaluated.