Physics-Informed Surrogate Neural Network for Optimizing Diode-Based Interfaces in Leadless Pacemaker Energy Harvesting

Efficient AC–DC interfaces are essential for low-power energy harvesters to power intracardiac leadless pacemakers (ICLPs) reliably. For the front-end circuitry, selecting a diode that minimizes both conduction and leakage losses under specified operating conditions is challenging, as the forward voltage drop and reverse leakage current cannot be simultaneously reduced due to the device’s inherent physics. This brief presents a machine learning framework that predicts the DC output of diode-based AC-DC interfaces directly from specific voltage-current (V-I) characteristic points of the diodes and descriptive features of the excitation waveform. A physics-informed surrogate model is trained on SPICE-simulated data of rectifiers driven by experimentally captured cardiac harvester waveforms. Hardware validation with printed-circuit rectifiers powered by the same harvester source exhibits an average error of 0.03 V between prediction and measurement, matching the performance observed on the synthetic set and demonstrating robust generalization to real-world applications. This data-driven approach offers an instant, datasheet-only diode screening without circuit-level evaluation and can be integrated into automated multi-objective optimization flows for the design of low-power energy harvester AC–DC interfaces.