With these assertions comes a caveat, however. While in this study, a non-zero discrepancy bias is considered as an indicator of missing physics, it may also be an indicator of inadequate numerical implementation. Within the scope of this manuscript, we will assume the a priori verification of simulation codes to produce converged solutions throughout the domain of applicability. Moreover, the concepts of zero or constant discrepancy bias would only be valid provided that the experiments are well spread to sufficiently explore the domain of applicability. This statement naturally leads to the following question: how much of the domain of applicability must be explored by the experiments? Reference [7] attempts to answer this question by investigating the influence of “coverage” on the predictive maturity of a simulation model. It should be noted that the above-mentioned merits rely heavily on the decisions made to facilitate model calibration. Perhaps the first decision that must be made early in the process is defining the domain within which the system is expected to operate reliably and safely. This decision also constitutes the domain within which the simulation model is executed to make forecasts. Subsequent necessary steps include the selection of the appropriate model parameters for calibration. Therefore, completing thorough model-based sensitivity analysis, quantifying uncertainty and developing a Parameter Identification and Ranking Table are necessary requirements that must precede model calibration. The approach described herein relies on a foundational premise: the physics or engineering principles involved in the application of the code remains unchanged when transitioning from tested to untested settings. This migration is best controlled by understanding the underlying phenomena and by exploring the domain of applicability through coverage (i.e. spreading the experiments to explore the domain of applicability). Therefore, a satisfactory understanding of underlying phenomena and a sufficient coverage of the domain of applicability with experiments are prerequisites for successful forecasting. In the case study examples, the forecasting errors are calculated as the distance between a probability distribution and a scalar value. This approach can be further refined by taking the experimental variability into account. BIBLIOGRAPHICAL REFERENCES [1] M. Kennedy and A. O’Hagan, “Predicting the output from a complex computer code when fast approximations are available,” Biometrika. 87 (2000) 1–13. 445
RkJQdWJsaXNoZXIy MTMzNzEzMQ==