June 2 - 4, 2011
Keywords of the presentation: Gaussian process; derivative; universal Kriging, nuclear engi- neering
In this work we discuss an approach for uncertainty propagation
through computationally expensive physics simulation
codes. Our approach incorporates gradient information information
to provide a higher quality surrogate with fewer simulation
results compared with derivative-free approaches.
We use this information in two ways: we fit a polynomial or Gaussian process model ("surrogate") of the system response. In a third approach we hybridize the techniques where a Gaussian process with polynomial mean is fit resulting in an improvement of both techniques. The surrogate coupled with input uncertainty information provides a complete uncertainty approach when
the physics simulation code can be run at only a small number
of times. We discuss various algorithmic choices such as polynomial basis and covariance kernel. We demonstrate our findings on synthetic
functions as well as nuclear reactor models.
Ordinary differential equations with uncertain parameters are a vast field of research.
Monte-Carlo simulation techniques are widely used to approximate quantities
of interest of the solution of random ordinary differential equations. Nevertheless,
over the last decades, methods based on spectral expansions of the solution
process have drawn great interest. They are promising methods to efficiently
approximate the solution of random ordinary differential equations. Although global
approaches on the parameter domain reveal to be very inaccurate in many
cases, an element-wise approach can be proven to converge. This poster presents
an algorithm, which is based on the stochastic Galerkin Runge-Kutta method.
It incorporates adaptive stepsize control in time and adaptive partitioning of
the parameter domain.
Keywords of the presentation: Uncertainy quantification, optimzation under uncertainty, functional ANOVA, stochastic collocation
This talk will describe experiences and challenges at Boeing with Uncertainty Quantification (UQ) and Optimization Under Uncertainty (OUU) in conceptual design problems that use complex computer simulations. The talk will describe tools and methods that have been developed and used by the Applied Math group at Boeing and their perceived strengths and limitations. Application of the tools and methods will be illustrated with an example in conceptual design of a hypersonic vehicle. Finally I will discuss future development plans and needs in UQ and OUU.
Keywords of the presentation: Polynomial chaos, Curse of DImensionality, Model Validation, Uncertainty Quantification.
The curse of dimensionality is a ubiquitous challenge in uncertainty quantification. It usually comes about as the complexity of analysis is controlled by the complexity of input parameters. In most cases of practical relevance, the output quantity of interest (QoI) is some integral of the input quantities and can thus be described in a much lower dimensional setting. This talk will describe novel procedures for honoring the low-dimensional character of the QoI without any loss of information. The talk will also describe the range of QoI that can be addressed using this formalism.
The role of UQ as the engine behind the model validation puts a burden of rigor on UQ formulations. The ability to explore the effect of particular probabilistic choices on model validity is paramount for practical applications in general, and data-poor applications in particular. The talk will also address achievable and meaningful definitions of the validation process and demonstrate their relevance in the context of industrial problems.
Industrial product and process designs often exploit physical limits to improve performance. In this regime uncertainty originating from fluctuations during fabrication and small disturbances in system operations severely impacts product performance and quality. Design robustness becomes a key issue in optimizing industrial designs. We present examples of challenges and solution approaches implemented in our robust design tool RoDeO.
Keywords of the presentation: Robust Design Optimization, turbo charger design, polynomial chaos expansions
Deterministic design optimization approaches are no longer satisfactory for industrial high technology products. Product and process designs often exploit physical limits to improve performance. In this regime uncertainty originating from fluctuations during fabrication and small disturbances in system operations severely impacts product performance and quality. Design robustness becomes a key issue in optimizing industrial designs.
We present challenges and solution approaches implemented in our robust design tool RoDeO applied turbo charger design. In addition to the challenges for electricity generating turbines, turbo chargers have to work efficiently for a wide range of rotation frequencies. Time-consuming aerodynamic (CFD) and mechanical (FEM) computations for large sets of frequencies became a severely limiting factor even for deterministic optimization. Further more constrained deterministic optimization could not guarantee critical design limits under impact of uncertainty during fabrication. Especially, the treatment of design constraints in terms of thresholds for von Mises stress or modal frequencies became crucial. We introduce an efficient approach for the numerical treatment of such chance constraints that even do not need additional CFD and FEM calculations in our robust design tool set.
An outlook for further design challenges concludes the presentation.
Contents of this presentation are joint work of U. Wever, M. Klaus, M. Paffrath and A. Gilg.
Keywords of the presentation: climate, Bayesian inference, MCMC, biases
The problem of estimating uncertainties in climate prediction is not well defined. While one can express its solution within a Bayesian statistical framework, the solution is not necessarily correct. One must confront the scientific issues for how observational data is used to test various hypotheses for the physics of climate. Moreover, one also must confront the computational challenges of estimating the posterior distribution without the help of a statistical emulator of the forward model. I will present results of a recently completed estimate of the uncertainty in specifying 15 parameters important to clouds, convection, and radiation of the Community Atmosphere Model. I learned that the maximum posterior probably is not in the same region of parameter space as the minimum log-likelihood. I have interpreted these differences to the existence of model biases and the potential that the minimum log-likelihood, which are often the desired solutions to data inversion problems, are over-fitting the data. Such a result highlights the need for a combination of scientific and computational thinking to begin to address uncertainties for complex multi-physics phenomena.
The problem of estimating uncertainties in climate prediction is not well defined. While one can express its solution within a Bayesian statistical framework, the solution is not necessarily correct. One must confront the scientific issues for how observational data is used to test various hypotheses for the physics of climate. Moreover, one also must confront the computational challenges of estimating the posterior distribution without the help of a statistical emulator of the forward model. I will present results of a recently completed estimate of the uncertainty in specifying 15 parameters important to clouds, convection, and radiation of the Community Atmosphere Model. I learned that the maximum posterior probably is not in the same region of parameter space as the minimum log-likelihood. I have interpreted these differences to the existence of model biases and the potential that the minimum log-likelihood, which are often the desired solutions to data inversion problems, are over-fitting the data. Such a result highlights the need for a combination of scientific and computational thinking to begin to address uncertainties for complex multi-physics phenomena.
A team at the Lawrence Livermore National Laboratory is
currently undertaking an uncertainty analysis of the Cummunity Earth
System Model (CESM), as a part of a larger effort to advance the
science of Uncertainty Quantification (UQ). The Climate UQ effort has
three major phases: UQ of the Cummunity Atmospheric Model (CAM)
component of CESM, UQ of CAM coupled to a simple slab ocean model, and
UQ of the fully coupled CESM (CAM + 3D ccean). In this poster we
describe the first phase of the Climate UQ effort; the generate of CAM
ensemble of simulations for sensitivity and uncertainty analysis.
Keywords of the presentation: kriging, standard error, mean squared error, global optimization
Kriging response surfaces are now widely used to optimize design parameters in industrial applications where assessing a design's performance requires long computer simulations. The typical approach starts by running the computer simulations at points in an experiment design and then fitting kriging surfaces to the resulting data. One then proceeds iteratively: calculations are made on the surfaces to select new point(s); the simulations are run at these points; and the surfaces are updated to reflect the results. The most advanced approaches for selecting new points for sampling balance sampling where the kriging predictor is good (local search) with sampling where the kriging mean squared error is high (global search). Putting some emphasis on searching where the error is high ensures that we improve the accuracy of the surfaces between iterations and also makes the search global.
A potential problem with these approaches, however, is that the classic formula for the kriging mean squared error underestimates the true error, especially in small samples. The reason is that the formula is derived under the assumption that the parameters of the underlying stochastic process are known, but in reality they are estimated. In this paper, we show how to fix this underestimation problem and explore how doing so affects the performance of kriging-based optimization methods.
In this work, we studied sensitivity of physic processes and simulations to parameters in climate model, reduced errors and derived optimal parameters used in cloud convection scheme. MVFSA method is employed to derive optimal parameters and quantify the climate uncertainty. Through this study, we observe that parameters such as downdraft, entrainment and cape consumption time have very important impact on convective precipitation. Although only precipitation is constrained in this study, other climate variables are controlled by the selected parameters so could be beneficial by the optimal parameters used in convective cloud scheme.
We analyze stochastic two-stage optimization problems with a stochastic dominance constraint on the recourse function. The dominance constraint provides risk control on the future cost. The dominance relation is represented by either the Lorenz functions or by the expected excess functions of the random variables. We propose two decomposition methods to solve the problem and prove their convergence. Our methods exploit the decomposition structure of the expected value two-stage problems and construct successive approximations of the stochastic dominance constraint.
Keywords of the presentation: Shock profiles in random media
Many issues in uncertainty quantification, as they emerge from the
perspective of large scale scientific computations of increasing complexity,
involve dealing with stochastic versions of the basic equations modeling the
phenomena of interest. A common reaction is to generate samples of solutions
by choosing parameters randomly and computing solutions repeatedly. It is
quickly realized that this is much too computationally demanding (but not entirely useless).
Another common reaction is to do a sensitivity analysis by varying parameters in the
neighborhood of regions of interest, leading to adjoint methods and computations that
are not much more demanding than the basic one for which we want to find error bars.
One does not have to be a sophisticated probabilist or statistician to realize that there
is room for some interdisciplinary research here. My experience in studying waves and diffusion in
random media motivated me to look into uncertainty quantification and to address some
of the emerging issues. One such issue is the study of the propagation of shock profiles in
random (turbulent) media. I will introduce this problem and analyze it from the point of
view of large deviations, which is a regime that is particularly difficult to explore numerically.
This problem is of independent interest in stochastic analysis and provides an example
of how ideas from this theoretical research area can be used in applications.
This is joint work with J. Garnier and T.W. Yang.
Mathematical modeling of industrial applications often yields time-dependent
systems of differential algebraic equations (DAEs) like in the simulation of
electric circuits or in multibody dynamics for robotics and vehicles. The
properties of a system of DAEs are characterized by its index. The DAEs
include physical parameters, which may exhibit uncertainties due to
measurements, for example. For a quantification of the uncertainties, we
replace the parameters by random variables. The resulting stochastic model can
be resolved by methods based on the polynomial chaos, where either a
stochastic collocation or the stochastic Galerkin technique is applied. We
analyze the index of the larger coupled system of DAEs, which has to be solved
in the stochastic Galerkin method. Moreover, we present results of numerical
simulations, where a system of DAEs corresponding to an electric circuit is
used as test example.
Keywords of the presentation: scenario generation, Quasi-Monte Carlo, scenario tree, electricity portfolio, risk-averse
We review some recent advances in high-dimensional numerical integration, namely,
in (i) optimal quantization of probability distributions, (ii) Quasi-Monte Carlo (QMC) methods, (iii) sparse grid methods. In particular, the methods (ii) and
(iii) may be superior compared to Monte Carlo (MC) methods under certain
conditions on the integrands. Some related open questions are also discussed.
In the second part of the talk we present a model for optimizing electricity
portfolios under demand and price uncertainty and argue that electricity
companies are interested in risk-averse decisions. We explain how the stochastic
data processes are modeled and how scenarios may be generated by QMC methods
followed by a tree generation procedure. We present solutions for the risk-neutral
and risk-averse situation, discuss the costs of risk aversion and provide several possibilities for risk aversion by multi-period risk measures.
Read More...
Keywords of the presentation: model selection, calibration, parameter estimation
This talk compares three approaches for model selection: classical least squares methods, information theoretic criteria, and Bayesian approaches. Least squares methods are not model selection methods although one can select the model that yields the smallest sum-of-squared error function. Information theoretic approaches balance overfitting with model accuracy by incorporating terms that penalize more parameters with a log-likelihood term to reflect goodness of fit. Bayesian model selection involves calculating the posterior probability that each model is correct, given experimental data and prior probabilities that each model is correct. As part of this calculation, one often calibrates the parameters of each model and this is included in the Bayesian calculations. Our approach is demonstrated on a structural dynamics example with models for energy dissipation and peak force across a bolted joint. The three approaches are compared and the influence of the log-likelihood term in all approaches is discussed.
The need for accurate predictions arise in a variety of critical
applications such as climate, aerospace and defense. In this work two
important aspects are considered when dealing with predictive
simulations under uncertainty: model selection and optimal
experimental design. Both are presented from an information theoretic
point of view. Their implementation is supported by the QUESO library,
which is a collection of statistical algorithms and programming
constructs supporting research into the uncertainty quantification
(UQ) of models and their predictions. Its versatility has permitted
the development of applications frameworks to support model selection
and optimal experimental design for complex models.
A predictive Bayesian model selection approach is presented to
discriminate coupled models used to predict an unobserved quantity of
interest (QoI). It is shown that the best coupled model for prediction
is the one that provides the most robust predictive distribution for
the QoI. The problem of optimal data collection to efficiently learn
the model parameters is also presented in the context of Bayesian
analysis. The preferred design is shown to be where the statistical
dependence between the model parameters and observables is the highest
possible. Here, the statistical dependence is quantified by mutual
information and estimated using a k-nearest neighbor based
approximation. Two specific applications are briefly presented in the
two contexts. The selection of models when dealing with predictions of
forced oscillators and the optimal experimental design for a graphite
nitridation experiment.
Keywords of the presentation: Model Validation, Calibration, Uncertainty Quantification, Gaussian Process Emulator, Bayesian Statistics
Model calibration, validation, prediction and uncertainty quantification have progressed remarkably in the past decade. However, many issues remain. This talk attempts to provide answers to the key questions: 1) how far have we gone? 2) what technical challenges remain? and 3) what are the future directions? Based on a comprehensive literature review from academic, industrial and government research and experience gained at the General Electric (GE) Company, we will summarize the advancements of methods and the applications of these methods to calibration, validation, prediction and uncertainty quantification. The latest research and application thrusts in the field will emphasize the extension of the Bayesian framework to validation of engineering analysis models. Closing remarks will offer insight into possible technical solutions to the challenges and future research directions.
Keywords of the presentation: Polynomial chaos, numerical methods.
Uncertainty quantification has been an active fields in recent years, and many numerical algorithms have been developed. Many research efforts have focused on how to improve the accuracy and error control of the UQ algorithms. To this end, methods based on polynomial chaos have established themselves as the more feasible approach. Despite the fast development from the computational sciences perspective, significant challenges still exist for UQ to be useful in practical systems. One prominent difficulty is the simulation cost. In many practical systems one can afford only a very limited number of simulations. And this prevents one from using many of the existing UQ algorithms. In this talk we discuss the importance of such a challenge and some of the early efforts to address it.