Speaker
Description
With the rapid adoption of Digital Twins in recent years, simulation models designed to replicate real-world physical systems have become increasingly common. To achieve accurate representations, it is typically necessary to update model parameters based on observations collected from sensors or measurements of the physical asset. However, no model can fully capture the infinitely complex nature of reality. As a result, quantifying the uncertainty in model predictions is essential for reliable decision-making. Bayesian updating frameworks provide an appealing approach for parameter calibration, inherently accounting for such uncertainties.
One often-overlooked source of error is model form uncertainty. This type of uncertainty arises from the fundamental discrepancies between the model and reality, stemming from the assumptions and simplifications made during model construction. Ignoring model form uncertainty can lead to overly confident predictions that fail to accurately reflect sensor observations. To address this, we propose an embedded model form uncertainty framework that attributes the model variability to a stochastic extension of the model's latent parameters. This approach enables the quantification of uncertainties that can be represented by a variation in the model parameters. Of particular interest are scenarios involving noisy observations or additional discrepancies that cannot be directly integrated into the model.
By incorporating uncertainty through the parameters, this method not only quantifies uncertainty in predictions but also propagates model form uncertainty to other Quantities of Interest (QoI) that rely on the same model or its parameters. Consequently, QoI computations yield more reliable values, accounting for the potential uncertainties introduced by imperfect models during parameter updating. Moreover, this approach facilitates a more comprehensive statistical analysis of QoI distributions, offering deeper insights into the model's reliability and highlighting areas for potential improvement. By incorporating model form uncertainty, decision-makers can achieve a more robust and nuanced understanding of system behavior and prediction quality.
