A software framework for probabilistic sensitivity analysis for computationally expensive models
Introduction
In many fields such as structural reliability [30], [31], [32], material modeling [30], [31], [33], finance etc., mathematical (numerical) models are used for predicting the response of a system. Due to the increasing computer power, the complexity of the model is growing. Generally, the more complex the models are, the larger becomes the uncertainty in the model outputs due to randomness in the input parameters. It is essential to determine how much the model output is changed by the variation in input parameters as well as calibrate and validate the mathematical models. Sensitivity analysis (SA) is a great help for these purposes [1]. Therefore, uncertainty and sensitivity analysis have recently received widespread interest of researcher in many fields such as material modeling and structural design. Numerous SA approaches have been developed to quantify the models with uncorrelated parameters [2]. However, engineering systems are complex and frequently contain correlated input parameters such that if one parameter varies, it results in variations in other parameters. The variation in the output of the models with correlated input parameters (e.g., composition constraints in material modeling [3]) is not only contributed by the variations in input parameter itself, but also contributed by the correlated variations in other parameters [4]. Hence, it is more realistic to estimate the effects of changing more than one parameters on the model outputs simultaneously. It is essential to understand the relations among the uncertain input parameters for designing a SA.
A few methods have been developed to quantitatively assess the effect of correlated input parameters on the model outputs. For instance, Xu and Gernert [4] improved the original Fourier amplitude sensitivity test (FAST) associated with Iman and Conover method [18] – used to generate correlated samples – to properly measure the sensitivity index for a model with correlated input parameters. Then, they developed another method to evaluate the sensitivity index for the uncorrelated and correlated contributions, see [5]. Nevertheless, those methods are limited in estimating the first-order sensitivity indices and the latter is based on a weak assumption that the model output linearly relates to input parameters. Later, SA methods, see [1], [6], [7], were proposed to quantitatively assess the total-effect sensitivity index that is essential for model simplification, however, most of them deal with analytical functions but not for experimental or simulation data.
Hence, a unified framework that links different steps from generating sample, constructing the surrogate model and implementing the sensitivity analysis method is needed. In this article, a review and computer implementation for uncertainty and sensitivity analysis and its application in engineering analysis has been carried out to provide a robust and powerful modeling tool to support for designing uncertainty and sensitivity analysis. We employ a sensitivity analysis (SA) method for the case of correlated parameters [6] whose formulas were derived similarly to Sobol’ formulas for the case of uncorrelated parameters [8]. For the estimation, the sample data is generated from the joint and conditional probability distribution functions of input parameters which are required to account for the constraints in the inputs space. Gaussian copula is used to generate a joint cumulative distribution function (CDF) (multivariate normal distribution) that requires only marginal distributions and covariance matrix of input parameters.
Complex models are often very time-consuming and computationally expensive so that they cannot be used to compute sensitivity indices. Thus, the so-called surrogate-based approach is employed as an approximation of the real model for sensitivity analysis. In [10], the authors presented a penalized spline regression model for a single continuous predictor. Since predictor variables have nonlinear relationships with the model output, the regression models considering multiple smooth functions [11] are adopted in this article to approximate the observed data. Subsequently, the SA indices are computed based on penalized spline regression models.
The objective of this work is to provide a MATLAB toolbox consisting of a set of functions that can be used to randomly generate samples, construct the surrogate model and carry out the SA. The computer implementation has been presented in this article. The support MATLAB code can be found at the website (http://www.uni-weimar.de/Bauing/rabczuk/).
The article is outlined as follows. In the next section, we briefly describe the flow chart and structure of the framework. The sampling technique is shown in Section 3. Surrogate models containing polynomial and penalized spline regression models are presented in Section 4.2. The SA is described in Section 5. Application of the SA method for models with correlated input parameters are presented in Section 6 including two analytical models with additional noise and an industrial example. Finally, we close the manuscript with concluding remarks.
Section snippets
Matlab toolbox: A flowchart and structure of the code
A framework including sampling of correlated input parameters, construction of surrogate model, and implementation of sensitivity analysis are respectively described in Fig. 1. Also, the structure and purpose of the code are briefly depicted in Table 1. In the first step, sampling technique is used to randomly generate correlated input values which are then inserted into the computational model to obtain the model response in the second step. In the third step, the regression model is used to
Sampling method
Latin Hypercube Sampling (LHS) [14], [15] is an improved sampling strategy that enables a reliable approximation of the stochastic properties even for a small number of samples N. LHS is used to provide the design points which are spread throughout the design space. The LHS can be summarized as:
- •
Divide the cumulative curve into N equal intervals on the cumulative distribution of each parameter;
- •
A probability value is then randomly selected from each interval of the parameter distribution
Polynomial regression model
Given a computational model with d predictor variables, the general multiple regression model is
The model response Y of N observed data points can be shown as:in which Y is an N × 1 vector, ϵ is an N × 1 error vector and β are the regression coefficients. By minimizingthe least-squares estimate of β is determined aswith . To test the model, we simply compute either the root mean square error (
Extension of Sobol’ approach for models with correlated inputs (ESACIs)
We consider a computational model f(X1, X2, …, Xd) defined in with a realization matrix . If the realization matrix X is partitioned into and the total variance of f(X1, X2, …, Xd) can be written as follows:
We can estimate the first-order of the group y byand the total-effect index of the group z by
Similarly, the total-effect index of the group y can be computed by
Application
In this article, we use the presented method applied to analytical models for the first two examples and industrial data for the last one. If the input parameters follow a correlated multivariate normal distribution, an application of the method is straightforward. In the general case, arbitrary multivariate data need to be transformed to a correlated multivariate normal distribution through copula-based methods. With regards to the first two analytical, a random noise will be added to model
Conclusions
A unified method for probabilistic sensitivity analysis for computationally expensive models with correlated inputs has been presented. The framework presents sequential steps consisting of generating random sample data by LHS, transforming an arbitrary distributions into a multivariate normal distribution, approximating the observed data via surrogate models and estimating the sensitivity indices for the model with correlated parameters. In summary, the computer implementation involves the
Acknowledgements
We gratefully acknowledge the support from National Basic Research Program of China (973 Program: 2011CB013800), NSFC (41130751), the Ministry of Science and Technology of China (SLDRCE14-B-31), Science and Technology Commission of Shanghai Municipality (16QA1404000), IRSES-MULTIFRAC. The support from the Alexander von Humboldt Foundation in the framework of the Sofja Kovalevskaja Award endowed by the Federal Ministry of Education and Research is acknowledged by X. Zhuang.
References (33)
- et al.
Variance-based sensitivity indices for models with dependent inputs
Rel. Eng. Syst. Safety
(2012) - et al.
Variance based sensitivity analysis of model output. design and estimator for the total sensitivity index
Comput. Phys. Commun.
(2010) - et al.
Uncertainty quantification for multiscale modeling of polymer nanocomposites with correlated parameters
Comp. Part B: Eng.
(2015) - et al.
Uncertainty and sensitivity analysis for models with correlated inputs
Rel. Eng. Syst. Safety
(2008) - et al.
Estimation of global sensitivity indices for models with dependent variables
Comput. Phys. Commun.
(2012) Variance-based sensitivity analysis in the presence of correlated input variables
Proceedings 5th international conference on reliable engineering computing (REC), Brno
(2012)- Ruppert D.....
- et al.
Stochastic predictions of bulk properties of amorphous polyethylene based on molecular dynamics simulations
Mech. Mat.
(2014) - Surjanovic S., Bingham D.. Virtual library of simulation experiments: test functions and datasets. Retrieved September...
- et al.
An importance quantification technique in uncertainty analysis for computer models
Proceedings of ISUMA, first international symposium on uncertainty modelling and analysis
(1990)
Extending a global sensitivity analysis technique to models with correlated parameters
Comput. Stat. Data Anal.
Sensitivity analysis for non-linear mathematical models
Math. Model. Comput. Exp.
A unified framework for stochastic predictions of mechanical properties of polymeric nanocomposites
Comput. Mat. Sci.
Semiparametric regression
DACE A MATLAB Kriging toolbox
Informatics and mathematical modeling
Cited by (542)
Methods for enabling real-time analysis in digital twins: A literature review
2024, Computers and StructuresModeling via peridynamics for crack propagation in laminated glass under fire
2024, Composite StructuresDeep learning uncertainty quantification for ultrasonic damage identification in composite structures
2024, Composite StructuresEffect of local openings on bearing behavior and failure mechanism of shield tunnel segments
2024, Underground Space (new)Effects of water contents on stick-slip mechanism at the granite-basalt interface
2024, Tribology International