Fig. 2: The sigma posterior distributions obtained after performing the Bayesian calibration without and with the model discrepancy correction.
From: Bayesian calibration of irradiated graphite property models under high temperatures
Sigma posterior distributions using the (a) individually fitted and (b) Bradford unified physics-based models for predicting IIDC with the NBG-18 grade. The σ posterior was obtained using the standard Bayesian approach and captures the uncertainty due to model discrepancy and experimental noise. In contrast, σε posterior was obtained using the KOH Bayesian approach and supposedly captures the uncertainty due to experimental noise only, as it accounts for the model discrepancy via a GP term. It is therefore intuitive that σε is on average less than σ. Additionally, in each figure there are two σ and σε posteriors corresponding to the two experimental datasets from AGC8 and INNOGRAPH10. The μσ is the variance posterior across both the experimental groups and, therefore, is more diffuse than the group-level variance posteriors.
