Sparse Polynomial Chaos expansion for advanced nuclear fuel cycle sensitivity analysis
Introduction
Nuclear fuel cycle simulators are being extensively used to evaluate different strategies mainly in mid and long term with the aim of assessing, among others, the possibility of implementation of advanced fast reactor technologies or the sustainability potential of the fuel cycle in the sense of safety, non-proliferation, economic competitiveness and waste management (Salvatores et al., 2008; Bianchi, 2010; OECD/NEA, 2013; IAEA, 2018). The fuel cycle scenarios, which cover all the technologies involved from the front-end to the back-end, are usually described as a collection of parameters that, given the correct operations (irradiation, reprocessing, decay, etc.), provide a result in terms of a set of indicators, such as the amount of waste, the gallery length required in the final repository or the amount of natural resources needed for fuel fabrication.
In recent studies, a special effort has been done so that the scenario analyses do not only provide the result of the assessment but also an estimation of the uncertainties in the output indicators, due to the assumptions made in the models, the inherent uncertainties in the input parameters or nuclear data. These uncertainties may be crucial for the evolution of the fuel cycle scenario since they can lead to the unavailability of certain materials when they are needed, introducing unexpected changes in the fuel cycle with strong consequences in terms of technology implementation or economic competitiveness.
Under this topic, in order to propagate the uncertainties from the input specifications to the desired observables as well as to quantify the impact of each variable, the research done in the recent years cover different degrees of approximation (OECD/NEA, 2017; Thiollière et al., 2018; Skarbeli and Álvarez-Velarde, 2019). The simplest techniques, involving parametric studies and local methods, have a narrow range of validity since they make strong assumptions about the nature of the underlying mathematical dependence that the input parameter space has with a certain output indicator (Cacuci, 2003). In the case of parametric studies, linear dependency is assumed between the input and the output parameters, while results of local methods, based on a first order approximation, are constrained to the vicinity of the reference values (true linear dependence does not exist when there is irradiation of nuclear fuel). The obvious benefit of these techniques is that their calculation only require a few evaluations. On the other hand, global theories (like the Sobol’s variance decomposition) do not impose any restriction, allowing the capture of non-linear relationships along the whole input parameter space as well as interactions (Saltelli et al., 2007).
However, the excessive computational cost these global theories require usually makes them non affordable from a practical point of view, so only a small subset of variables can be selected for this kind of studies. This variable selection can be done relying again on local approximations, like performing a sensitivity search over the whole input parameter space for finding the most relevant variables in a first order approximation. Nevertheless, the computational demand for estimating the Sobol indices with this hybrid methodology is still very high. For example, in the EULER cluster available at CIEMAT (consisting on Xeon E5450 and X5570 quadcores with 2 GB of RAM per core that can execute approximately 100 cases in 20 min), the estimation of the Sobol indices (with more than 100000 simulations) required around two weeks of calculations for an advanced nuclear fuel cycle scenario.
One possibility for dealing with this situation is to approximate the computational exhaustive simulation with a simpler model. This approach, known as surrogate model or metamodel, include splines, artificial neural networks and support vector machine approximations, among others (see for example de Boor, 2001, Basheer and Hajmeer, 2000, Clarke et al., 2004). Another way of generating the surrogate model is by using the Polynomial Chaos expansion (Ghanem and Spanos, 1991, Xiu and Karniadakis, 2003, Sudret, 2008). The advantage of this choice, as it will be explained in Section §4, is that it allows for the easy deduction of the statistical information as it has already been proven by previous works within the nuclear community. In Williams (2006), they are used for solving a stochastic transport equation. Transient scenarios have also been studied with this methodology by means of the point kinetics approximation. For example, Gilli et al. (2012) studies a reactivity insertion which was lately explored by Wu and Kozlowski (2017) too but for performing inverse uncertainty quantification. In Perkó et al. (2014), the case of an unprotected loss of flow is analyzed by using an adaptive basis which is constructed depending to the contributions to the variance.
In this work, the Polynomial Chaos expansion has been applied to an advanced European nuclear fuel cycle with pressurized water reactors (PWR) and accelerator driven subcritical systems (ADS) focused on the advanced management of the transuranic inventories. The expansion coefficients have been determined by regression, in which a regularization term was included in order to build an adaptive sparse basis. The Sobol indices have been calculated with this technique, and the results have been compared with previous studies in order to show the improvements of this new methodology.
Section snippets
Cycle description
The different techniques used for the computation of the Sobol indices (by direct integration, and by the Polynomial Chaos expansion) have been tested using the fuel cycle scenario that was modeled in previous works. It is based on the first one defined in the PATEROS project (Partitioning and Transmutation European Roadmap for Sustainable nuclear energy) framed on the Framework Program of Euratom (Salvatores et al., 2008; Martínez-Val, 2008). This advanced scenario is centered in the
Sensitivity analysis
Given the mathematical complexity of the fuel cycle simulation, the system can be understood as a black-box function, where an input vector produces the solution y. When dealing with uncertainties, the input will be a random vector , and the observable will be therefore a random function Y. Let us assume now that , a d-dimensional function, is integrable over the hypercube . According to the Sobol expansion, its variance can be decomposed into increasing order summands (Sobol, 1993)
General remarks
The Polynomial Chaos expansion, originally introduced by Wiener for studying Gaussian variables back in the 30’ (Wiener, 1938) and lately extended to other families of random variables (Xiu and Karniadakis, 2002), has in the last decades gained attention in the field of uncertainty quantification given its ability for modeling random responses (Ghanem and Spanos, 1991, Xiu and Karniadakis, 2003, Sudret, 2008). Since a detailed description of the mathematical background can be found on the
Scenario results with Sobol method
As previously mentioned, the objective of this work is to show the benefits of the Polynomial Chaos Expansion. Therefore, we will only focus on the sensitivity analysis of the fuel cycle. A detailed discussion of the scenario outcomes can be found on Skarbeli and Álvarez-Velarde (2019). Nevertheless, the key aspects relevant for this work will be described in the following.
Four output indicators where chosen for measuring the quality of the transmutation scenario as it was mentioned in Section §
Conclusions
The exhaustive computational cost the Sobol indices estimation demands as a consequence of the high-multidimensional integrals that emerge from their definitions, make them generally unaffordable for nuclear fuel cycle global sensitivity analyses. In this work, instead of solving the integral expressions, an alternative methodology to deduce the Sobol indices with smaller cost has been applied. This methodology (which relies on the Polynomial Chaos expansion theory for building a surrogate
CRediT authorship contribution statement
A.V. Skarbeli: Methodology, Software, Formal analysis, Investigation. F. Álvarez-Velarde: Conceptualization, Software, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was supported in part by the Spanish national company for radioactive waste management ENRESA, through the CIEMAT-ENRESA agreement on Transmutación de residuos radiactivos de alta actividad and the Spanish Programa Estatal de I+D+i Orientada a los Retos de la Sociedad project SYTRAD2 (ENE2017-89280-R).
References (33)
- et al.
Artificial neural networks: fundamentals, computing, design, and application
J. Microbiol. Methods
(2000) - et al.
Adaptive sparse polynomial chaos expansion based on least angle regression
J. Comput. Phys.
(2011) - et al.
Performing uncertainty analysis of a nonlinear point-kinetics/lumped parameters problem using polynomial chaos techniques
Ann. Nucl. Energy
(2012) - et al.
Importance measures in global sensitivity analysis of nonlinear models
Reliab. Eng. Syst. Saf.
(1996) Relaxed lasso
Comput. Stat. Data Anal.
(2007)- et al.
Large scale applicability of a fully adaptive non-intrusive spectral projection technique: sensitivity and uncertainty analysis of a transient
Ann. Nucl. Energy
(2014) - 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 on advanced fuel cycle scenario simulations applying local and global methods
Ann. Nucl. Energy
(2019) Global sensitivity analysis using polynomial chaos expansions
Reliab. Eng. Syst. Saf.
(2008)Polynomial chaos functions and stochastic differential equations
Ann. Nucl. Energy
(2006)
Inverse uncertainty quantification of reactor simulations under the bayesian framework using surrogate models constructed by polynomial chaos expansion
Nucl. Eng. Des.
Modeling uncertainty in flow simulations via generalized polynomial chaos
J. Comput. Phys.
CP-ESFR SP2-WP1: proposal of an ESFR reference scenario
Sensitivity & Uncertainty Analysis
Valuation of mortgage-backed securities using brownian bridges to reduce effective dimension
J. Comput. Finance
Analysis of support vector regression for approximation of complex engineering analyses
J. Mech. Des.
Cited by (6)
A variance deconvolution estimator for efficient uncertainty quantification in Monte Carlo radiation transport applications
2024, Journal of Quantitative Spectroscopy and Radiative TransferRecent research in advanced fast reactors and fuel cycle strategies in Spain
2024, Nuclear Engineering and DesignBiophysical parameter estimation of crops from polarimetric synthetic aperture radar imagery with data-driven polynomial chaos expansion and global sensitivity analysis
2022, Computers and Electronics in AgricultureCitation Excerpt :In this study, one of the attractive metamodeling methods, namely polynomial chaos expansion (PCE), was considered to handle these two strong assumptions. PCE is one of the most popular techniques, which establishes the functional relationship between the input data, including their interaction, and the output data using orthogonal polynomials, in uncertainty quantification, specifically in metamodeling for engineering (Hariri-Ardebili and Sudret, 2020; Skarbeli and Álvarez-Velarde, 2020; Yuzugullu et al., 2017; Yuzugullu et al., 2015). Besides the frequent use of PCE in uncertainty quantification, studies show that the PCE can be used as a data-driven machine learning technique for regression analysis (Celik and Erten, 2021; Torre et al., 2019; Oladyshkin and Nowak, 2012), and its direct relationship with sensitivity analysis based on orthogonal polynomial coefficients makes the model popular where there is a high dimensional input feature vector.
Preliminary uncertainty and sensitivity analysis of the Molten Salt Fast Reactor steady-state using a Polynomial Chaos Expansion method
2021, Annals of Nuclear EnergyCitation Excerpt :Moreover, we opted for a non-intrusive approach, which treats the reference model as a “black-box”, thus requiring no modifications of the underlying equations. These characteristics make the adopted PCE approach particularly attractive for nuclear applications (see, for example, Fichtl and Prinja, 2011; Gilli et al., 2012, 2013; Perkó et al., 2014b; Turati et al., 2018; Peng et al., 2019; Skarbeli and Álvarez-Velarde, 2020), where the reference models are typically coupled multi-physics simulations often implemented in difficult to manipulate legacy codes and – in general – are computationally expensive. The published literature on UQ/SA approaches applied to molten salt reactor systems is limited.
Automated selection of nuclides and reactions of interest in a depletion simulation. Precision loss estimation for multiple outputs
2021, Progress in Nuclear EnergyCitation Excerpt :This is the case for the depletion module of CLASS which is really close to the Runge–Kutta method of SMURE. For meta-model using predictors directly for the isotopic composition of the spent fuel, like the sparse polynomial chaos expansion available in TR_EVOL (Skarbeli and Alvarez-Velarde, 2020) or the neural networks available in CYCLUS (Bae et al., 2020), no decrease of running time inside the fuel cycle simulator will probably occur. Other codes avoid direct use of meta-models specific to the fuel cycle simulator but use external codes chosen for its simplicity and small running time.
PRINCIPAL COMPONENT ANALYSIS BASED POLYNOMIAL CHAOS EXPANSION REGRESSION OF LEAF AREA INDEX FROM POLSAR IMAGERY
2021, International Geoscience and Remote Sensing Symposium (IGARSS)