Neural ODE keeps SiC power module health checks accurate across failure modes

Data-driven health estimators for silicon-carbide (SiC) power modules are usually validated on a single failure mode, leaving their real-world reliability unclear. In a new arXiv paper (2608.08365), researchers compared five standard prognostics and condition-monitoring methods against a physics-informed Neural ODE (NODE) using two power-cycling campaigns with structurally different failure mechanisms: solder-layer fatigue and wire-bond lift-off. They used a per-module k-fold protocol to ensure fair benchmarking. The NODE was tested with two input representations—baseline electrical precursors and a custom set of cumulative thermoelectric features—while keeping the rest of the pipeline identical.

Results showed that every reference method lost accuracy on the wire-bond campaign, with average errors increasing and precision dropping relative to their soldered-campaign performance. In contrast, the NODE fed with cumulative features maintained its soldered-campaign metrics across both mechanisms, with variation within fold-to-fold variance. However, the same NODE architecture using baseline precursors fell back to the reference-method cluster. The authors conclude that input representation contributes at least as much as model architecture to failure-mechanism transferability, and the paper is accepted at IECON 2026.