Degradation-informed state-of-health estimation of lithium-ion batteries using a Tabular Prior-data Fitted Network

Lithium-ion batteries degrade during service through coupled electrochemical and thermal aging, causing capacity fade and increased operational risk. Accurate state-of-health (SOH) estimation is therefore required for battery health management. When only compact cycle-level descriptors are available, maintaining prediction accuracy together with model interpretability remains difficult. This study proposes a degradation-informed hyperparameter-optimized Tabular Prior-data Fitted Network (HPO-TabPFN) framework for SOH estimation. Each cycle is represented using temporal, energy-related and incremental-capacity descriptors, chronological context is constructed for TabPFN, and inference-related hyperparameters are optimized. Grid search, random search, particle swarm optimization and Bayesian optimization (BO) are evaluated under the same temporal validation protocol. Shapley additive explanations (SHAP) is used to analyze local feature attribution, and Sobol analysis is used to quantify model-output sensitivity and interaction effects. On the evaluated NASA and CALCE cell sequences, hyperparameter optimization reduced the prediction error of TabPFN, and the Bayesian-optimized model achieved average mean absolute percentage error (MAPE) values of 0.2450% and 0.1800%, respectively. Under the same chronological delayed-label protocol, BO-TabPFN gave the lowest error among the unoptimized TabPFN and the evaluated baseline regressors on the tested cells. These results indicate that cycling-derived health indicators can support accurate tabular SOH estimation in the evaluated public datasets, while the NASA B05 case study illustrates how the fitted estimator depends on the selected health factors under the stated attribution and sensitivity settings.