Physics-Informed and Brain-Inspired Set-Routed Neural Networks for Lithium-Ion Battery Life-Cycle Assessment in Medical Devices

This work presents a dual-purpose study. The first objective is to generate a realistic synthetic dataset for battery health monitoring. The second objective is to validate both the realism of the dataset and the performance of a novel neural architecture, termed the Set-Routed Neural Network (SetNet), for regression-based estimation of State of Health (SOH) in medical-grade energy storage devices. The dataset comprises 20 complete life cycles, each containing approximately 50,000-time steps with recorded electrical, thermal, and state parameters including Current (A), Voltage (V), Temperature (°C), State of Charge (SOC), and SOH. To ensure realism, statistical validation and frequency-domain analyses confirmed physically plausible behavior, with a mean voltage of 3.89 V ± 0.08, mean temperature of 25.08 °C ± 0.20, and SOC distributions consistent with operational ranges. In the learning phase, SetNet was benchmarked against a standard Multilayer Perceptron (MLP) using an 80/20 train-test split over all life cycles. Results show that SetNet achieved an RMSE of 0.0496, MAE of 0.0412, and an R2 score of 0.2639, outperforming the MLP baseline (RMSE 0.0872, MAE 0.0679, R2 0.275). The smooth convergence curves and stable mini-batch loss observed in SetNet indicate superior generalization and robustness. These findings highlight SetNet’s ability to route features selectively across modular subnetworks, enabling efficient modeling of multi-domain interactions (electrical–thermal–chemical) inherent in medical battery systems. Overall, the proposed architecture establishes a biologically inspired, interpretable foundation for predictive maintenance and reliability assessment in future medical energy systems.