A Framework for Refining Error Metrics in Surrogate Models for Engineering Applications

Model-Based Systems Engineering (MBSE) is a growing field in engineering design, enabling rapid prototyping and deployment of concepts. However, the quality of engineering simulations depends heavily on the quality of the models used. As a result, quantifying and reducing model error is critical in MBSE. To do this effectively, examining how model error is measured is crucial. Error metrics reduce the complex relationship between predicted and measured behavior to a single scalar value. This compression can introduce bias, but it is necessary for error quantification and surrogate generation. This paper examines the impact of this compression on model behavior and offers a decision framework for choosing error metrics. While not all uncertainty is reducible, modelers should decide which uncertainties are acceptable and how they are measured.