Accurate and stable prediction of the State-of-Health (SOH) of lithium-ion batteries is essential for extending battery lifetime, preventing catastrophic failures, and enabling predictive maintenance in electric mobility and energy storage applications. Conventional machine-learning approaches rely heavily on historical feature engineering, whereas deep recurrent models struggle to represent long-range temporal dependencies and produce smooth degradation trajectories. This paper proposes a unified transformer-based approach for cycle-resolved SOH estimation using physics-informed deep learning. Three transformer variants are evaluated on two widely used public datasets: the Oxford Battery Degradation Dataset and the NASA battery Aging Dataset. A baseline iTransformer is introduced for direct sequence modeling of voltage, current, temperature, and cycle time. A time-informed iTransformer integrates positional encodings to explicitly model temporal patterns in capacity fade. Finally, a physics-informed variant incorporates a smoothness regularization term based on second-order discrete derivatives to enforce physically meaningful degradation behavior. Experimental results demonstrate that the physics-informed iTransformer achieves superior accuracy, smoother prediction curves, and improved cross-dataset generalization. Compared with the baseline, The physics-informed model reached a MAE of 0.309, RMSE of 0.188, and R2 of 0.95 on the Oxford dataset. Cross-dataset experiments on the NASA benchmark further confirm that physics guidance significantly reduces overfitting and improves real-world robustness. The proposed approach provides a scalable foundation for coming generation physics-consistent deep learning strategies for electrochemical modeling and battery management systems.