Uncertainty-Aware Decision Support for Thermowettability in Graphene Channels: Failure Modes across the Evidence Chain

Abstract

AI-supported decisions about graphene-channel thermowettability can fail at multiple points: biased molecular simulations, poorly calibrated temperature or contact measurements, oversimplified representations, distribution shift, and control thresholds detached from physical risk. This review follows those failures from evidence generation to action. It evaluates probabilistic surrogate models, ensemble uncertainty, and residual-based anomaly detection as ways to reveal conflicting signals rather than average them away. Failure-aware testing includes altered surface condition, thermal gradients beyond training support, measurement dropout, model misspecification, and rare transport regimes. The synthesis emphasizes calibration by regime, asymmetric error costs, and a separate estimate of whether the input itself is trustworthy. Provenance links predictions to simulation settings, specimens, sensors, preprocessing, and model versions; high epistemic uncertainty prompts new measurement or human review. The article is a structured synthesis and claims no new transport result. Its contribution is a defensible AI decision process in which uncertainty is traced to its source and every failure mode has an associated diagnostic, escalation, or revalidation response.