Please come and join "The 2nd Annual KAIST NQE Distinguished Lecture Series". Find the details below.

1. Topic: Predictive Modeling: Combining Computations with Experiments to obtain Optimally Predicted Results with Reduced Uncertainties

2. Date and Time: June 2(Tue) 2015, 4:30 pm

3. Place: N13-1 Building (Chang Young Shin Student Activity Center)/ Woolim Hall (Rm.101)

4. Speaker: Dr. Dan Gabriel Cacuci, SmartState Endowed Chair Professor University of South Carolina, USA.

• South Carolina’s SmartState Endowed Chair Professor in Advanced Materials and Nuclear Power and Director of the Center of Economic Excellence in Nuclear Science and Energy, University of South Carolina (USC), USA

• Editor, Nuclear Science and Engineering (a research journal of the American Nuclear Society), and The Handbook of Nuclear Engineering

• Principal Research Fellow, Department of Earth Sciences and Engineering, Imperial College London, UK

5. Abstract:

Discrepancies observed in practice between experimental and computational results provide the basic motivation for performing quantitative model verification, validation, and model calibration through data assimilation. Furthermore, numerical simulations of ever increasing fidelity and complexity demand a broad multidisciplinary research on scalable algorithms and models, including hardware, architecture, system software, libraries, workflows, performance, verification, and application software. “Predictive modeling” incorporates all of these activities, aiming at predicting “best-estimate” values for model responses and parameters, along with reduced predicted uncertainties for these quantities.

This Lecture will present the principles underlying an original (2014) methodology for predictive modeling of large-scale nonlinear coupled multi-physics systems. Illustrative applications of this new methodology to large scale (millions of model parameters) reactor physics and thermal-hydraulics systems will also be highlighted, demonstrating the reduction in the predicted uncertainties for various fundamental design and operational parameters (e.g., effective multiplication factor, reaction rates, time-dependent void fractions). The Lecture will also sketch the principles underlying a newly (2015) developed generalization of the “adjoint sensitivity analysis methodology” for nonlinear systems, originally developed by the author (1981) and widely applied since then (e.g., earth and atmospheric sciences, econometrics). This generalization enables the exact and most efficient computation of arbitrarily high-order response sensitivities to a large number of model parameters. In turn, the availability of such high-order sensitivities enables the computation of high-order moments (e.g., skewness and kurtosis) of model response distributions, which subsequently enables the quantification of non-Gaussian features (asymmetries, “long tails” characterizing rare events, etc.) of model results.

The Lecture will conclude by highlighting the main directions of ongoing applications (e.g., design of a small fast LBE-cooled reactor, proliferation detection) and ongoing research aimed at extending the author’s predictive modeling methodology, from second to fourth-order, incorporating computational and experimental covariance, skewness and kurtosis information. Successful completion of these ongoing developments is expected to provide a paradigm-changing methodology for predicting “best-quantified” design and operational parameters for characterizing the features of large-scale non-linear systems.