Failure-Aware Evaluation for Lithium-Sulfur Separator Engineering: Measurement and Comparator Design

Abstract

Separator evaluation in lithium-sulfur batteries can look favorable when averaged cycling metrics conceal shuttle-driven degradation, rare short-circuit precursors, or batch-specific defects. This review redesigns measurement and comparison around those consequential failure modes. It considers multimodal AI models that combine voltage-current trajectories, impedance, thermal response, imaging, and separator provenance, then evaluates them with calibrated risk, event-level recall, lead time, false-alarm burden, and performance under chemistry or protocol shift. Comparator design separates the contribution of material modification, preprocessing, sequence modelling, and uncertainty estimation, avoiding baselines that differ in data access or operating conditions. The synthesis also addresses class imbalance, censored cells, delayed labels, and leakage between cycles from the same specimen. An abstaining model is treated as preferable to a confident recommendation outside the validated domain. No new experiment is presented. The article contributes a failure-aware AI evaluation protocol that connects measurement choices to inspection, derating, replacement, and human escalation, making the practical cost of errors visible alongside conventional electrochemical performance.