A Forecasting Metric for Predictive Modeling Sezer Atamturktur1, François Hemez2, and Cetin Unal5 1Department of Civil Engineering, Clemson University, South Carolina, U.S.A., sez@clemson.edu 2XCP 1-Division, Los Alamos National Laboratory, New Mexico, U.S.A., hemez@lanl.gov 3CCS DO-Division, Los Alamos National Laboratory, New Mexico, U.S.A., cu@lanl.gov 1 INTRODUCTION As the complexity of engineering systems increases, their performance becomes more difficult to predict through modeling and simulation. This paper investigates simulation models used to forecast predictions of the performance of engineering systems in support of high-consequence decision-making. Specifically this paper directs its attention to the validation of the simulation models for certification purposes. Instead of relying on virgin models, i.e. models that are not calibrated or bias-corrected, we envision certification to be applied through a combined experimental and numerical campaign that relies on simulation models calibrated and bias corrected against experimental measurements. We are particularly interested in the quantification and control of errors associated with the forecasting predictions of these calibrated and bias corrected simulation models. In certification, the purpose of simulation models is to reduce the number of required experiments, and is best illustrated by considering two extreme cases. (1) Purely empirical certification: the absence of a sound simulation model where certification is only obtained based on experimental measurements. (2) Purely model-based certification: the availability of a ‘perfect’ simulation model where certification needs practically no experimental measurements. In a purely empirical approach, forecasting is commonly achieved by constructing a function that best fits the data produced from an experimental campaign. Then, the best-fitted function is exercised to make forecasting predictions at untested settings. Both the experimental uncertainty and the uncertainty in the curve-fitting process can be considered by making forecasting predictions that are “best estimates” with quantified uncertainties. In purely empirical certifications, the number of experiments necessary to train the best-fitted function can rapidly become prohibitive. Purely empirical certifications are further challenged with the fact that experiments are typically time-consuming and expensive and thus, even in T. Proulx (ed.), Linking Models and Experiments, Volume 2, Conference Proceedings of the Society for Experimental Mechanics Series 5, 431 DOI 10.1007/978-1-4419-9305-2_33, © The Society for Experimental Mechanics, Inc. 2011
RkJQdWJsaXNoZXIy MTMzNzEzMQ==