<p>Computer model simulations are widely used to represent and predict the behavior of physical processes in many fields of engineering. To improve simulation predictions by inferring unknown input model parameters and model bias, computer models are often calibrated through comparing their predictions against validation experiments&nbsp;. Bayesian model calibration can combine prior knowledge about unknown parameters and experimental data and quantify the uncertainty in the calibration results. In Bayesian calibration, the posterior distributions are typically estimated using Markov Chain Monte Carlo sampling in which a large number of model realizations are needed. Hence, it is often necessary to approximate the simulation model using a surrogate model to reduce the computational cost. Training such surrogates requires realizations of the original model at inputs that explore the parameter space. In some cases, computer models with various levels of fidelity may be available, such as reduced-order physics or coarser grids in lower-fidelity models. A higher-fidelity model more adequately represents the physical process than a lower-fidelity model, but is more computationally expensive. Although the lower-fidelity model is less expensive, using it in model calibration can result in significant error. Therefore, considering the available computational budget, a methodology for efficiently combining data from the lower- and higher-fidelity models is needed to train a computationally inexpensive surrogate model with the highest possible accuracy and predictive capability. To maximize the information gained from the computer realizations for training this hybrid surrogate model, and to improve the prediction of discrepancy inferred from the model calibration, a batch sequential experiment design strategy is utilized. This strategy efficiently augments an initial batch of experiments. The design criterion used to maximize the information content of the future experiments is the expected improvement for predictive stability (EIPS). This metric quantifies the stability of the inferred discrepancy based on the Kullback–Leibler distance between the current and future distribution of the discrepancy. For optimization of EIPS design criterion, the Federov exchange algorithm is used. The proposed algorithm chooses additional design points from the higher- or lower-fidelity model to maximize expected improvement and minimize computational costs at the same time. This approach is illustrated through a simple engineering example: a one-mass oscillator subject to vibration isolation.</p>

Multi-Fidelity Model Calibration With Quantified Uncertainties

Paper Proceedings: https://asmedigitalcollection.asme.org/VVS/VVS2021/volume/84782

ASME 2021 Verification and Validation (VVS2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.