SAND2025-11240 SANDIA REPORT SAND2025-11240 Printed September, 2025 Challenges in quantifying unparameterized spatial uncertainties in deep geologic repositories for nuclear waste Teresa Portone & Dusty M. Brooks & Laura P. Swiler Prepared by Sandia National Laboratories Albuquerque, New Mexico 87185 Livermore, California 94550 Issued by Sandia National Laboratories, operated for the United States Department of Energy by National Technology & Engineering Solutions of Sandia, LLC. NOTICE: This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government, nor any agency thereof, nor any of their employees, nor any of their contractors, subcontractors, or their employees, make any warranty, express or implied, or assume any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represent that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government, any agency thereof, or any of their contractors or subcontractors. The views and opinions expressed herein do not necessarily state or reflect those of the United States Government, any agency thereof, or any of their contractors. Printed in the United States of America. This report has been reproduced directly from the best available copy. Available to DOE and DOE contractors from U.S. Department of Energy Office of Scientific and Technical Information P.O. Box 62 Oak Ridge, TN 37831 Telephone: (865) 576-8401 Facsimile: (865) 576-5728 E-Mail: reports@osti.gov Online ordering: http://www.osti.gov/scitech Available to the public from U.S. Department of Commerce National Technical Information Service 5301 Shawnee Road Alexandria, VA 22312 Telephone: (800) 553-6847 Facsimile: (703) 605-6900 E-Mail: orders@ntis.gov Online order: https://classic.ntis.gov/help/order-methods NT OF E ME N RT ER A DEP GY β€’ β€’ IC A U NIT ER ED ST M A TES OF A 2 ABSTRACT Spatially heterogeneous uncertainties are prevalent in geophysical modeling applications, such as probabilistic post-closure performance assessment (PA) of deep geologic repositories for nuclear waste. Such uncertainties are often highly influential to model outputs, so it is desirable to identify the most important mechanisms by which they influence model predictions. However, these uncertainties are often unparameterized in the sense that there is no set of parameters that can be specified to yield a particular realization of the uncertainty. Additionally, the uncertainty is not intrinsically endowed with a parameterization that captures a realization’s mechanistic influence on model outputs. Therefore, in this work we present a novel methodology to develop and assess a set of proxy variables that aim to represent this influence. We show how they can be computed, downselected, and incorporated into the construction of statistical surrogate models mapping model inputs to outputs. We present our methodology in the context of a motivating application problem in deep geologic repository PA and discuss the challenges in capturing the effects of these spatial heterogeneities in uncertainty analyses. 3 4 CONTENTS 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2. Application Problem: Nuclear Waste Repository in Fractured Crystalline Rock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 3.1. Polynomial Chaos Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 3.2. Sobol’ indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.2.1. Binning-based main effects indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 3.2.2. PCE-based indices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 4. Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 4.1. Proxy variables for heterogeneous spatial uncertainties . . . . . . . . . . . . . . . . . . . . . . . . 23 4.1.1. Proxy variable selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 4.1.2. Incorporating proxy variables in PCE construction . . . . . . . . . . . . . . . . . . . . . 27 4.2. Regression PCE approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 4.3. Accuracy checks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 5. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 A. Expanded numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 5 6 LIST OF FIGURES Figure 2-1. A cutaway of an example permeability field for a crystalline reference case simulation. Permeability of the background rock is filtered out so that fractures are depicted. The repository lies in the upstream half of the domain. A glacial aquifer sits at the top of the computational domain, where 129 I concentrations are tracked at observation points (marked in green) and as a maximum over the aquifer region. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Figure 2-2. To generate a DFN realization, probability distributions for fracture centroid, ra- dius, and orientation are sampled repeatedly until a target fracture density 𝑃32 is achieved. The DFN is then mapped to an equivalent continuous porous medium, where heterogeneous, anisotropic material properties are defined within each cell. The streamwise component of the permeability tensor is shown here for a uniform grid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 Figure 4-1. An illustration of mapping DFNs to graphs, then extracting topological proper- ties from the graphs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Figure 4-2. Pearson correlation coefficients between repository performance QoIs (rows) vs. candidate proxy variables (columns). Proxy variables with at least one correlation coefficient greater than .32 (an 𝑅 2 β‰ˆ 0.1) are: average degree, density, number of intersections with the repository, number of intersections, logRSTT, and mean log permeability of the ECPM. . . . . . . . . . . . . . . . . . . . . . . . . 26 Figure 4-3. Pearson correlation coefficients between candidate proxy variables that are cor- related with at least one repository performance QoI. Density and number of intersections are significantly correlated with average degree, and mean log10 permeability is correlated with logRSTT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Figure 4-4. Density fits to the proxy variables, compared to histograms with 10, 20, and 30 bins to show the robustness of the empirical distribution. . . . . . . . . . . . . . . . . . . . . 28 Figure 5-1. Empirical first-order Sobol’ indices for repository performance QoIs computed using 1000 samples (blue bars) and the 95% confidence interval for 100 bootstrap samples (orange). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 Figure 5-2. A comparison of results for 2𝑛𝑑 -order PCEs without and with cross-validation and a 3π‘Ÿ 𝑑 -order PCE with cross-validation. PCE results are presented colum- nwise. Scatterplots depict the PCEs evaluated at the build points vs. the build points. Bar charts depict the PCE-based Sobol’ indices (green) vs. those com- puted using the build points (blue) and the 95% confidence interval for 100 bootstrap samples (orange). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 Figure 5-3. A histogram comparing the 2𝑛𝑑 order, no cross-validation PCE predictions of peak 129 I concentration in the aquifer, vs. the build points. . . . . . . . . . . . . . . . . . . . 33 7 Figure 5-4. Input-output scatterplots for the peak 129 I concentration in the aquifer. The DFN indices have been sorted according to QoI variance from smallest to largest to make the dependence more clear. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Figure A-1. Scatterplots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 Figure A-2. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). . . . . . . . . . . . . 41 Figure A-3. Scatter plots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 Figure A-4. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). . . . . . . . . . . . . 43 Figure A-5. Scatter plots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 Figure A-6. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). . . . . . . . . . . . . 45 8 LIST OF TABLES Table 2-1. Performance QoIs for the original and simplified CRCs. Mathematical definitions and methods of computation for MdRT, FMF, and MTT are discussed in [22] (FMF is the ratio of the net mass flux and the instantaneous mass described therein). 16 Table 5-1. Descriptions and distributions for the parameters included in the UQ studies. Uniform (and log-uniform) distributions with lower and upper bounds 𝑙𝑏 and 𝑒𝑏 are denoted U [𝑙𝑏, 𝑒𝑏] (and log U [𝑙𝑏, 𝑒𝑏]). Normal distributions with mean πœ‡ and standard deviation 𝜎 are denoted N (πœ‡, 𝜎 2 ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 9 10 1. INTRODUCTION Probabilistic post-closure performance assessment (PA) of systems for deep geologic disposal of nuclear waste seeks to evaluate the performance of a repository system and provide reasonable assurance that it will meet relevant requirements for the protection of humans and the environment over a prolonged period. The intergovernmental Nuclear Energy Agency has an extensive set of guidance documents on safety assessments [25, 26, 27]. PA is key to the design and approval of nuclear waste repositories and involves characterization of the many sources and effects of uncertainty in these systems [12, 30, 6]. Modeled phenomena include: multicomponent, multiphase flow and transport; heat conduction and convection; biogeochemical reactions; geomechanics, and isotope decay and ingrowth. A key challenge recently faced in this application area is the presence of unparameterized heteroge- neous spatial uncertainties, meaning there is no set of parameters that can be specified to produce a particular realization of the uncertainty. This can occur, e.g., through the stochastic generation of heterogeneous subsurface properties. Lack of parameterization for such sources of uncertainty make them challenging to incorporate into uncertainty analyses typical to PA, especially global sensitivity analysis (GSA), which measures the importance of different sources of uncertainty to model outputs. Novel in this work is the development of proxy variables that aim to capture characteristics of the spatial variability in the subsurface that drive performance of the repository system. Such proxy variables can enable incorporation of spatial heterogeneity into GSA, as well as enable incorporation of spatial variability into surrogate models, which require explicit input-output sample pairs for their construction. In this report we discuss recent investigations into the efficacy of such proxy variables to rep- resent spatially heterogeneous sources of uncertainty and diagnostic methods to assess how well they capture the effects of the spatial heterogeneities on model outputs. The report proceeds as follows: in Chapter 2, our motivating application problem for deep geologic repository PA in fractured crystalline rock is introduced; in Chapter 3, an overview of relevant sampling-based and surrogate-based methods is provided; in Chapter 4, we discuss the development of proxy vari- ables representing spatially heterogeneous properties and their incorporation into surrogate model construction; in Chapter 5, results are presented and discussed; conclusions and future work are discussed in Chapter 6. 11 12 2. APPLICATION PROBLEM: NUCLEAR WASTE REPOSITORY IN FRACTURED CRYSTALLINE ROCK In this work we consider a generic nuclear waste repository modeled in a crystalline rock formation, called the crystalline reference case (CRC). The CRC is introduced and discussed in detail in [23, 38]. It is a demonstration case to build understanding of the performance of nuclear waste repositories in crystalline rock and is not based on any physical site. We summarize the problem here, with an illustration of problem domain in Figure 2-1. In the CRC, a repository is placed hundreds of meters deep in a highly impermeable, fractured rock, where flow is dominated by pathways along the fractures. Nuclear waste packages placed within the repository region can breach upon corrosion of the waste package [23]. Upon breach, radionuclides such as 129 I are released from the waste packages and transported through the surrounding repository regions and host rock. Potential exposure to humans occurs by radionuclides reaching the biosphere, e.g., groundwater used for drinking water or crop irrigation. We take as a key repository performance quantity of interest (QoI) for PA the concentration of 129 I in an aquifer region. We take additional QoIs that shed light on system properties in the vicinity of the repository and the far field, such as mass flow rates from the repository and out of the downstream plane of the domain. Figure 2-1. A cutaway of an example permeability field for a crystalline reference case simulation. Permeability of the background rock is filtered out so that fractures are depicted. The repository lies in the upstream half of the domain. A glacial aquifer sits at the top of the computational domain, where 129 I concentrations are tracked at observation points (marked in green) and as a maximum over the aquifer region. 13 Discrete fracture networks (DFNs) are used to represent the fractured host rock. DFNs are networks of two-dimensional shapes, typically ellipses or rectangles, distributed in a three-dimensional do- main (or lines distributed in a two-dimensional domain). They are commonly used to simulate isothermal fluid flow and particle transport in fractured rock and have been used to assess the per- formance of proposed radiocative waste repositories in crystalline rock at Aspo [33] and Formsark, Sweden [11]. It is not possible to perfectly characterize heterogeneous subsurface properties, for example, to identify each fracture in a km3 domain. Instead, DFNs are often modeled statistically and multiple DFN instantiations are generated to represent uncertainty in the subsurface properties. Fracture centroids, sizes, and orientations are represented as random variables with their probability distributions calibrated to agree with observations at site, as in [5]. To create a DFN realization, fractures are generated by sampling the probability distributions repeatedly until a specified frac- ture density (𝑃32 ) is reached [15]. The repeated sampling of these statistical models to generate a single DFN realization means we do not have a single set of parameters that maps to a particular DFN realization; the DFNs are unparameterized in the previously described sense. The statistical representation for the DFNs used in the CRC is adapted from properties of the Forsmark site in Sweden (Table 2, [35]) and is described in detail in [24, 41]. A DFN must be discretized to perform simulations. One approach is to mesh the fractures directly (fractures are assigned volume through their aperture, often assumed a function of fracture size [15]) and evolve the governing equations on the meshed DFN. This approach is not amenable to PA modeling because, unless coupled to a continuum, it cannot simulate heat conduction through the rock matrix (nor the resulting thermally-driven fluid fluxes), and it cannot simulate coupled chemical and thermal processes. Since nuclear waste releases a significant amount of heat, this is problematic for simulations of nuclear waste repository sites. A desirable alternative to directly meshing the fractures is to derive an equivalent continuous porous medium (ECPM) through upscaling, wherein a mesh is defined over the entire computational domain. In cells which contain a fracture, equivalent continuum properties associated with that fracture are defined; see [34] for the upscaling procedure used here. ECPMs allow the representation of fractured and background media in the same domain, capture the behavior of radionuclide transport in fractured rock, and allow simulation of coupled heat and fluid flow, including heat conduction through the matrix of the fractured rock, while honoring the intrinsic heterogeneity and anisotropy of the underlying DFN [34, 11]. The process of generating a DFN and mapping it to an ECPM is depicted in Figure 2-2. DFN realizations arising from the same statistical representation can result in significantly different repository performance properties. For the CRC, DFN realizations are a primary source of vari- ability in repository performance QoIs [39, 37]. However, the lack of a parametric representation of a DFN makes it challenging to identify the characteristics of a particular DFN realization that drive performance of the nuclear waste repository. This lack of a parametrization also makes it challenging to perform global sensitivity analysis or construct surrogate models for UQ. We thus investigate the development of scalar proxy variables to represent DFN characteristics to enable more detailed uncertainty analysis with DFNs. The CRC is simulated using PFLOTRAN, an open-source massively parallel subsurface flow and reactive transport simulator [20, 21, 9, 10]. PFLOTRAN models subsurface flow using a porous medium continuum approach, which includes capabilities for multicomponent systems, multiphase 14 Figure 2-2. To generate a DFN realization, probability distributions for fracture centroid, radius, and orientation are sampled repeatedly until a target fracture density 𝑃32 is achieved. The DFN is then mapped to an equivalent continuous porous medium, where heterogeneous, anisotropic material properties are defined within each cell. The streamwise component of the permeability tensor is shown here for a uniform grid. flow and transport, heat conduction and convection, biogeochemical reactions, geomechanics, and radionuclide decay and ingrowth. The CRC is computationally challenging, with multiple spatial and time scales present in the problem: it resolves on the order of 107 degrees of freedom, and requires approximately 1000 core hours per simulation on average to run. Other generic repository simulations are even more taxingβ€”for example, a model problem in a shale host rock required 15000 core-hours per simulation [18]. The high computational cost of PA simulations makes typical GSA algorithms using model eval- uations computationally intractable. Surrogate-based approaches are an attractive alternative for enabling GSA in this context, but they face their own challenges for this application problem, namely the inability to directly incorporate unparameterized source of uncertainty (like DFN real- izations) in their construction. This has motivated our investigations into the development of proxy variables to represent the influence of the DFNs on repository performance QoIs. Several QoIs are considered to study the performance of the simulated nuclear waste repository site. These so-called β€œperformance QoIs” include measures of repository retention, 129 I concentration, and bulk flow and transport characteristics, and are defined and described in Table 2-1. As discussed in detail in [40], a nested sampling structure is employed for the CRC in order to separate practically reducible and irreducible uncertainties. It will not be feasible to measure and model the true fracture network at a potential repository site, so it practically remains an irreducible uncertainty. By employing a nested sampling structure, we can separate and quantify the effects of β€œouter loop” DFN uncertainty and β€œinner loop” parameter uncertainty. Other studies may vary different types of uncertainty on the inner and outer loops, depending on the problem structure. We note that an outer sampling loop over epistemic uncertainties and inner sampling loop over aleatory uncertainties is common but not exactly what is done here. In this study, sampling comprises an outer loop with 25 DFN realizations and an inner parameter loop with 40 random samples of all other parameters. The inner-loop random samples differ for each DFN realization in order to produce more independent samples for surrogate construction. The outer loop also sets a random seed for the random sampling of waste package degradation properties, so that the order in which 15 Type of QoI Performance QoI Description Median Residence Time Scalar. The time at which half of a passive Repository (MdRT) of Repository tracer has been flushed from the reposi- retention [yr] tory. Fractional Mass Flux Function of 𝑑. The instantaneous frac- (FMF) from Repository tional loss rate of a passive tracer from [yrβˆ’1 ] the repository region at time 𝑑. Peak 129 I Concentration Scalar. The maximum 129 I concentration in Aquifer [M] in the aquifer, taken over space and time. 129 I Peak 129 I π‘₯ and 𝑦 location Scalars. The π‘₯ and 𝑦 location at which the Concentration [m] peak occurs. Time of Peak 129 I Con- Scalar. The time at which the peak oc- centration [yr] curs. Maximum value is the end of the simulation, 100 kyr. Mean Travel Time Function of 𝑑. Mean travel time from (MTT) at Peak 129 I repository to location of maximum 129 I Location [yr] concentration in aquifer at time 𝑑. RockAq / AqEb Mass Function of 𝑑. The ratio of mass flow rates Bulk flow & Flow Rate (MFR) Ratio [kg/yr] from rock to aquifer (RockAq) and transport [-] from aquifer to east boundary (AqEb). characteristics Computed using the total mass of water crossing the planes as a function of time. AqEb / RockEb MFR Function of 𝑑. The ratio of mass flow Ratio [-] rates [kg/yr] from aquifer to east boundary (AqEb) and from rock to east boundary (RockEb). Table 2-1. Performance QoIs for the original and simplified CRCs. Mathematical definitions and methods of computation for MdRT, FMF, and MTT are discussed in [22] (FMF is the ratio of the net mass flux and the instantaneous mass described therein). waste packages breach differs for each DFN realization. 16 3. BACKGROUND Here we introduce the surrogate modeling approaches deployed in this report, variance-based sensitivity analysis (VBSA) used for GSA, and sampling- and surrogate-based methods for VBSA. Throughout the chapter we discuss these methods in the context of a model quantity of interest, 𝑄, which depends on 𝑑 uncertain input random variables z. We denote a matrix of 𝑁 random realizations of z as Z ∈ R𝑁×𝑑 . 3.1. Polynomial Chaos Expansions Polynomial chaos expansions are stochastic expansions whereby the model output is expanded in a series of basis polynomials that are orthogonal with respect to the probability densities of the uncertain inputs. A total-order PCE expansion of order 𝑝 for model output 𝑄(z) is defined as: ( 𝑑 ) βˆ‘οΈ βˆ‘οΈ 𝑑 𝑄(z) β‰ˆ 𝛼i Ξ¨i (z), i = [𝑖1 , . . . , 𝑖 𝑑 ] ∈ Z+ , I𝑝 = i : π‘–π‘˜ ≀ 𝑝 , (3.1) i∈I𝑝 π‘˜=1 where i are multi-indices whose sums do not exceed 𝑝, and the basis functions Ξ¨i are the product of order 𝑖 π‘˜ one-dimensional orthogonal polynomial basis functions in each input dimension [36]. The number of terms (coefficients) required for a total-order PCE of order 𝑝 with 𝑑 input parameters is (𝑑 + 𝑝)! π‘π‘‘π‘’π‘Ÿπ‘šπ‘  = . (3.2) 𝑑! 𝑝! The number of terms required to achieve a good approximation of 𝑄(z) increases rapidly with the complexity of 𝑄 (e.g., higher-order behavior requiring larger 𝑝) and the number of inputs 𝑑. The coefficients of the expansion are determined either by projecting model responses onto the basis functions or by solving a regression problem. Projecting model responses onto basis functions requires a multidimensional weighted integral over the domain of z, where the weighting function is the probability distribution for z, 𝑝(z): ∫ βŸ¨π‘„, Ξ¨i ⟩ 2 βˆ’1 𝛼i = = ⟨Ψi ⟩ 𝑄(z)Ξ¨i (z) 𝑝(z)dz. ⟨Ψi2 ⟩ For common types of random variables, quadrature rules can be used to efficiently approximate the integral, but these quadrature rules require a structured set of samples that can’t be reused for other UQ tasks. 17 Determining the coefficients via a regression problem does not impose any structure on the input samples. For a set of 𝑁 random input samples Z, let Q ∈ R𝑁 be the vector of model outputs evaluated for each sample in Z, and let 𝚿 ∈ Rπ‘Γ—π‘π‘‘π‘’π‘Ÿ π‘šπ‘  be the matrix of π‘π‘‘π‘’π‘Ÿπ‘šπ‘  basis functions Ξ¨i evaluated at the sample input values in Z. A typical least-squares regression problem for PCE coefficients 𝜢 ∈ Rπ‘π‘‘π‘’π‘Ÿ π‘šπ‘  is defined as πœΆβˆ— = min βˆ₯𝚿𝜢 βˆ’ Qβˆ₯ 2 . (3.3) 𝜢 The collocation ratio π‘Ÿ 𝑐 = 𝑁/π‘π‘‘π‘’π‘Ÿ π‘šπ‘  corresponds to whether the regression problem is over-determined (π‘Ÿ 𝑐 > 1), uniquely determined (π‘Ÿ 𝑐 = 1), or under-determined (π‘Ÿ 𝑐 < 1). Different regression ap- proaches are appropriate depending on π‘Ÿ 𝑐 , spanning ordinary least squares (π‘Ÿ 𝑐 β‰₯ 1) and regularized regression approaches (π‘Ÿ 𝑐 < 1) such as Least Absolute Shrinkage Operator (LASSO) [16]. If the output response is β€œnoisy” with respect to the inputs zβ€”that is, if there is variability in 𝑄 beyond what is caused by zβ€”it may be advantageous to use cross-validation methods and select π‘Ÿ 𝑐 ≫ 1 to avoid overfitting to noise in the regression problem [16]. 3.2. Sobol’ indices VBSA is a powerful tool for identifying input-output dependencies in complex models and pri- oritizing modeling resources. VBSA measures the fraction of model output variance that can be attributed to each input, taking into account interactions between inputs. The most important inputs can be identified by ranking them in terms of their contribution to the output variance. This information can be used to interpret the modelβ€”do the most important inputs agree with expert understanding of the physical dependencies for this model output? It can also be used to prioritize data collection and further modeling efforts to reduce uncertainty in these inputs. Sobol’ indices are the most common VBSA method. Detailed explanations of Sobol’ indices and methods to compute them are provided in [31, 29], but the basic concepts are described here for completeness. The Sobol’ main effect index for a model output 𝑄(z) and one of its uncertain parameters 𝑧𝑖 is defined as  V𝑧𝑖 EzβˆΌπ‘– [𝑄 | 𝑧𝑖 ] 𝑆𝑖 ≑ , (3.4) V(𝑄) where zβˆΌπ‘– denotes the set of model parameters excluding 𝑧𝑖 . The Sobol’ main effect index can be interpreted as the fraction of the variance in 𝑄 that can be attributed to varying 𝑧𝑖 alone. Similarly, the total effect index is defined as   EzβˆΌπ‘– V𝑧𝑖 (𝑄 | zβˆΌπ‘– ) 𝑇𝑖 ≑ (3.5) V(𝑄) and can be interpreted as the fraction of the variance of 𝑄 that can be attributed to 𝑧𝑖 including variance from its interactions with any other inputs. 18 Grouped Sobol’ indices measure the importance of a group of parameters and their interactions to a model output. For a group of parameters zu where u βŠ‚ {1, 2, . . . , 𝑑} and its complement z∼u , the Sobol’ indices for the group zu are defined as    𝑔 Vzu Ez∼u [𝑄 | zu ] Ez∼u Vzu (𝑄 | z∼u ) 𝑆u ≑ , 𝑇u ≑ (3.6) V(𝑄) V(𝑄) Interaction effects measure the influence of interactions between inputs on a model output. They can be defined recursively from second-order interactions, which for inputs 𝑧𝑖 , 𝑧 𝑗 are defined as 𝑔 𝑆𝑖 𝑗 = 𝑆 {𝑖, 𝑗 } βˆ’ 𝑆𝑖 βˆ’ 𝑆 𝑗 . (3.7) Third-order interaction effects are defined for inputs 𝑧𝑖 , 𝑧 𝑗 , 𝑧 π‘˜ as 𝑔 𝑆𝑖 𝑗 π‘˜ = 𝑆 {𝑖, 𝑗,π‘˜ } βˆ’ 𝑆𝑖 𝑗 βˆ’ π‘†π‘–π‘˜ βˆ’ 𝑆 𝑗 π‘˜ βˆ’ 𝑆𝑖 βˆ’ 𝑆 𝑗 βˆ’ 𝑆 π‘˜ , (3.8) and higher-order interaction effects are defined recursively in a similar manner. The most common β€œpick-freeze” approach to computing Sobol’ indices is computationally in- tractable for practical PA problems due to its high computational cost. The approach requires 𝑁 (𝑑 + 2) model evaluations to compute main and total effects indices, where 𝑁 is the number of independent input samples and 𝑑 is the number of inputs as before [31]. Additionally, pick-freeze methods require structured samples, so only 2𝑁 model evaluations are independent. The other 𝑑𝑁 samples are not independent and therefore their utility for other downstream UQ analyses is limited. Two alternatives to the pick-freeze approach are a binning-based approach and a surrogate-based approach. We summarize these options herein. 3.2.1. Binning-based main effects indices To mitigate the computational cost associated with β€œpick-freeze” approaches, binning-based meth- ods to compute main effects indices have been developed which require only 𝑁 random input-output samples (note their cost does not scale with input dimension 𝑑) [3, 28, 19]. Broadly, these methods work by binning input-output samples based on input values, approximating the inner statistic of the output within each bin, then the outer statistic over the bins. Let the probability space for input 𝑧𝑖 be denoted Ω𝑖 , and let it be partitioned into 𝑀 bins 𝐴 π‘˜ such that 𝑀 Ø Ω𝑖 = π΄π‘˜ , 𝐴 π‘˜ ∩ 𝐴 π‘—β‰ π‘˜ = βˆ…, 𝐴 π‘˜ = [π‘Ž π‘˜βˆ’1 , π‘Ž π‘˜ ). (3.9) π‘˜=1 By the law of total variance E [V(𝑄|𝑧𝑖 )] 𝑆𝑖 = 1 βˆ’ . (3.10) V(𝑄) 19 Such binning-based methods typically compute the numerator of this fraction as 𝑀 βˆ‘οΈ 𝑛( 𝐴 π‘˜ ) E [V(𝑄|𝑧𝑖 )] β‰ˆ 𝐸𝑉 c𝑖 = 𝑠2 ( 𝐴 π‘˜ )𝑃( 𝐴 π‘˜ ), 𝑃( 𝐴 π‘˜ ) = , (3.11) π‘˜=1 𝑁 where 𝑠2 ( 𝐴 π‘˜ ) is the sample variance of 𝑄 in bin 𝐴 π‘˜ and 𝑛( 𝐴 π‘˜ ) is the number of samples in bin 𝐴 π‘˜ . The binning-based main effect estimator is thus computed as 𝐸𝑉 c𝑖 𝑆ˆ𝑖 = 1 βˆ’ 2 , (3.12) 𝑠 (𝑄) where the denominator is computed using all 𝑁 samples of 𝑄. For continuous random variables an equiprobable partition is most commonly applied, so that the number of samples per bin is approximately equal. For discrete or categorical random variables, such as DFN realization, one bin is assigned per distinct value the variable takes. 3.2.2. PCE-based indices In [36] it is shown that Sobol’ indices can be computed using the coefficients of a PCE. Given the expansion in Equation (3.1), the total variance represented by the PCE can be computed as βˆ‘οΈ V(𝑄) = 𝛼i2 , (3.13) i∈I𝑝 \{0} that is, by computing the sum of squares of all coefficients, neglecting 𝛼0 which is the mean of 𝑄. The numerator of the PCE-based main effect (first-order) Sobol’ index for input 𝑧 𝑗 can be computed as βˆ‘οΈ 𝛼i2 , I𝑝 = i ∈ I𝑝 : 𝑖 𝑗 > 0, 𝑖 π‘˜ = 0 βˆ€ π‘˜ β‰  𝑗 , 𝑗  𝑆𝑖 V(𝑄) = (3.14) 𝑗 i∈I𝑝 that is, by computing the sum of squares of all coefficients for which 𝑖 𝑗 is nonzero, and for which all other index elements are 0 (to neglect interaction terms). The numerator for the PCE-based total effect Sobol’ index for input 𝑧 𝑗 can be computed as βˆ‘οΈ 𝛼i2 , I𝑝𝐽 = i : 𝑖 𝑗 > 0 ,  𝑇 𝑖 V(𝑄) = (3.15) i∈I𝑝 ∈I𝑝𝐽 that is, by computing the sum of squares of all coefficients for which 𝑖 𝑗 is nonzero (thus incorporating interaction terms). The numerator for interaction effects for inputs indexed by u βŠ‚ {1, 2, . . . , 𝑑} can be computed as βˆ‘οΈ 𝛼i2 , I𝑝u = i ∈ I𝑝 : 𝑖 𝑗 > 0 βˆ€ 𝑗 ∈ u and 𝑖 π‘˜ = 0 βˆ€ π‘˜ βˆ‰ u ,  𝑆u V(𝑄) = (3.16) i∈I𝑝u 20 that is, by computing the sum of squares of all coefficients where all inputs indexed by u are active and all complementary inputs are inactive. The numerator for group first-order indices for inputs indexed by u βŠ‚ {1, 2, . . . , 𝑑} can be computed as βˆ‘οΈ u,𝑔 𝛼i2 , I𝑝 = i ∈ I𝑝 : 𝑖 π‘˜ = 0 βˆ€ π‘˜ βˆ‰ u , 𝑔  𝑆u V(𝑄) = (3.17) u,𝑔 i∈I𝑝 that is, the sum of squares of all coefficients for which at least one input in u is active and all complementary inputs are inactive. Finally, the numerator for the group total-order indices can be computed as βˆ‘οΈ 𝛼u2 , I𝑝u,𝑇 = i ∈ I𝑝 : i ∩ u β‰  βˆ… ,  𝑇u V(𝑄) = (3.18) i∈I𝑝u,𝑇 that is, the sum of squares of all coefficients for which at last one input in u is active (allowing for interactions with complementary inputs). 21 22 4. METHODS Here we discuss the methods developed and applied to address the challenge of spatial heterogeneity. First, we discuss the development of scalar proxy variables to capture important characteristics of heterogeneous spatial uncertainties. The proxy variables will not in general perfectly capture the influence of these heterogeneous uncertainties, so there will not be a perfect functional mapping from proxy variable to model outputβ€”there will be β€œnoise” in the output as a function of the proxy variables. Additionally, the proxy variables are derived from random realizations of our heterogeneous uncertainties, so surrogate construction methods that must specify the location of input samples cannot be applied. We thus discuss our approach to PCE construction that addresses this lack of control of input sample locations and the noise in model outputs. 4.1. Proxy variables for heterogeneous spatial uncertainties In this chapter we describe an approach to account for spatial uncertainties that are not inherently defined parametrically through the development of scalar β€œproxy variables”. These proxy variables aim to capture important characteristics of a random realization of such a spatial uncertainty. The approach taken to derive these proxy variables will depend on the application problem and the nature of the spatial uncertainty. In our case, we wish to characterize key properties of a discrete fracture network realization that might influence repository performance QoIs, namely the 129 I concentration in the glacial aquifer region. However, in general other quantities might be of interestβ€”for example, summary statistics such as correlation length, mean, or variance of the random realization. The proxy variables we pose here are application-problem specific, but the methods to assess the quality or adequacy of the proxy variables are application-problem agnostic. As noted in Chapter 2, flow pathways through fractured crystalline rock are predominantly along the fractures; this has motivated investigations to better understand the network topology of DFNs and their effect on flow and transport. Recent works have investigated representing DFNs as networks or graphs (nodes connected by edges), where typically fractures are nodes and intersections between fractures are edges. Such a graph representation of a DFN preserves the network topology of the DFNs while enabling an intuitive and computationally tractable means of extracting information about the network, e.g., the average number of intersections per fracture, the shortest path between two points in the network, etc. [2]. Several studies have used graph representations to characterize and study the structure of 2D and 3D DFNs, as well as DFN structure’s effect on flow and transport properties [2, 13, 42, 14, 7, 32, 8]. Additionally, recent work has shown promise in using the graphs themselves as models for flow and transport through DFNs [43, 17]. Given this success, we pursued a graph representation of DFNs here. 23 Glacial region Fracture graph 6 5 3 # fractures: 4 # intersections: 6 4 2 # intersections with repository: 1 Intersection graph Average intersections per fracture: 2.5 1 2 5 6 … 1 3 Repository region 4 Figure 4-1. An illustration of mapping DFNs to graphs, then extracting topological properties from the graphs. It should be noted that the performance assessment model incorporates other coupled physical processes with flow and transport, so the application of DFN graph representations here is more complex than previous studies. The DFN interacts with a spatially-resolved repository, and the DFN is converted to a meshed ECPM for repository simulations. These additional complexities and couplings manifest in the performance QoIs, which may make it challenging to identify the topological features of DFN realizations that are most influential. An illustration of the process by which topological properties are extracted from DFNs is shown in Figure 4-1. We used the dfnGraph utility in dfnWorks to generate two types of graph representations, which are described in detail in [43]: β€œfracture graphs” and β€œintersection graphs”. In a fracture graph, each fracture in a DFN is a node, and each intersection between two fractures is an edge between their nodes. Fracture graphs are useful for computing metrics related to the network topology, e.g., average degree, which corresponds to the average intersections per fracture. In an intersection graph, each intersection is a node and each edge is a fracture. Edges are endowed with geometric and hydrological properties of the fractures they represent, making it possible to compute metrics related to the flow and transport properties of the DFN, for example, to compute a shortest travel time from one or more starting nodes to one or more ending nodes in the DFN [43, 17]. Computation of these metrics is implemented in the dfnFlow utility of dfnWorks. There are challenges associated with incorporating proxy variables in UQ analyses as low-dimensional representations of more complex, higher-dimensional sources of uncertainty. First, the proxy vari- ables will likely not perfectly capture the effects of the high-dimensional uncertainty; second, the proxy variables may be statistically dependent if not chosen carefully. Second, dependence between variables violates common assumptions in surrogate model construction and global sensitivity anal- ysis. Since GSA is of critical interest for PAs, we aim to develop a set of proxy variables which are independent and capture as much of the effect of the DFNs on repository performance QoIs as possible. To do so, we performed a preliminary correlation analysis between candidate proxy variables and repository performance QoIs based on our 1000 input-output samples. Because of the nested sampling structure of the problem, the same DFN realization was used for 40 inner- loop random samples. To perform correlation analysis with proxy variables for a given DFN, performance QoIs were averaged over these 40 inner-loop samples. 24 4.1.1. Proxy variable selection Candidate proxy variables were determined based on physics intuition of the properties of a DFN that might influence repository performance. These candidate variables were: β€’ number of fractures in the DFN β€’ number of intersections in the DFN β€’ number of intersections with the repository region (NIwR) β€’ average degree of the fracture graph (avgDegree, average number of intersections per fracture) β€’ density of the fracture graph (global connectivity of the network) β€’ log relative shortest travel time between repository and aquifer regions (logRSTT, the orders- of-magnitude deviation of the shortest travel time between repository and aquifer relative to the median). In addition to the graph metrics described herein, proxy variables derived from the ECPMs as- sociated with each DFN were computed, namely the means of continuum properties (tortuosity, porosity, and log-base-10-transformed permeability) computed over all cells in the mesh. Correlation coefficients can serve as a preliminary screening tool to downselect proxy variables. Correlations between candidate proxy variables and repository performance QoIs are reported in Figure 4-2. For this problem we observe that NIwR, avgDegree, and logRSTT exhibit significant correlations (magnitude β‰₯ 0.32 β†’ 𝑅 2 βͺ† 0.1) with the most QoIs. Additionally, density, number of intersections, and mean log10 permeability have a significant correlation coefficient for at least one QoI. Beyond correlations between proxy variables and QoIs, it is important to consider correlations between the variables. Variance-based sensitivity analysis methods like Sobol’ indices are under- pinned by assumptions of independence between input variables. Without this independence, it becomes challenging to interpret the indices. We thus additionally screen for correlations between the proxy variables, removing any variables that are significantly correlated with others. The corre- lation coefficients between the proxy variables that are correlated with at least one QoI are presented in Figure 4-3. In this case, the density, number of intersections, and mean log10 permeability proxy variables exhibit significant correlation with the proxy variables that correlated with the most QoIs: average degree, NIwR, and logRSTT. To avoid correlations between proxy variables while maintaining a set of variables which are influential for the majority of Qois, we thus downselect to the most influential proxy variables which are relatively mutually independent: avgDegree, NIwR, and logRSTT. The magnitude of correlations between these proxy variables is no greater than 0.1. We deem this sufficient to treat the proxy variables as independent for the purposes of VBSA. We downselected to a single set of proxy variables for PCE-based VBSA of all QoIs for the sake of simplicity in our analyses, but different proxy variables could have been chosen for different QoIs. For example, the mass flow rate ratio between π‘…π‘œπ‘π‘˜ π΄π‘ž and π‘…π‘œπ‘π‘˜ 𝐸 𝑏 (rock-to-aquifer and rock- to-east-boundary) was most strongly correlated with the number of intersections proxy variable. Future analyses could consider selecting proxy variables for each QoI. No proxy variable exhibits particularly strong correlations with any QoI. As shown in Figure 4-2, the highest magnitude is 0.47, corresponding to an 𝑅 2 β‰ˆ 0.22. That is, no individual proxy variable 25 Average N N Mean log Mean Mean Degree Density NIwR Intersections Fractures logRSTT Permeability Porosity Tortuosity Peak 129 I Concentration in -0.07 0.05 0.47 -0.17 -0.10 0.14 0.05 -0.08 0.02 Aquifer [M] MTT at Peak -0.47 -0.33 -0.41 0.08 0.25 0.12 0.07 -0.05 0.19 129 I [yr] Time of Peak -0.44 -0.23 -0.25 -0.07 0.13 0.04 0.14 -0.06 0.13 129 I [yr] x Location of -0.03 0.05 -0.47 -0.11 -0.07 -0.20 0.34 -0.24 0.18 Peak 129 I [M] y Location of 0.12 0.18 -0.20 -0.17 -0.17 -0.15 0.11 -0.01 -0.16 Peak 129 I [M] MdRT of -0.21 -0.08 -0.37 -0.07 0.03 0.36 0.07 -0.06 0.22 Repository [yr] FMF from Repository 0.19 0.06 0.39 0.09 -0.01 -0.35 -0.02 -0.02 -0.06 at 3 kyr [yrβˆ’1 ] FMF from Repository 0.32 0.27 0.32 -0.14 -0.24 -0.24 -0.05 0.09 -0.27 at 1 Myr [yrβˆ’1 ] AqEb RockEb MFR Ratio 0.01 0.10 0.04 -0.17 -0.13 0.13 -0.06 -0.02 0.13 at 1 MYr [-] RockAq RockEb MFR Ratio -0.02 0.22 -0.08 -0.39 -0.28 0.20 0.14 -0.11 0.13 at 1 Myr [-] Figure 4-2. Pearson correlation coefficients between repository performance QoIs (rows) vs. candidate proxy variables (columns). Proxy variables with at least one correlation coefficient greater than .32 (an 𝑅 2 β‰ˆ 0.1) are: average degree, density, number of intersections with the repository, number of intersections, logRSTT, and mean log permeability of the ECPM. explains more than about 20% of variation in the QoIs, under an assumption of linear dependence of QoIs on proxy variables. Note that the nested sampling structure employed in this problem means only 25 samples were available to compute these correlation coefficients, so there is a chance that coefficients could increase or decrease with more samples. Other methods to identify and downselect proxy variables may reveal stronger dependencies. Other measures of dependence might be more insightful, e.g., partial rank correlation coefficients, which measure the monotonic dependence between inputs and outputs, controlling for interaction effects with other inputs. Additional analyses could also be performed to expand the set of candidate 26 N Mean log Density NIwR Intersections logRSTT Permeability Average 0.85 0.05 -0.43 0.09 0.08 Degree Density 0.15 -0.84 -0.04 0.19 NIwR -0.22 -0.09 0.08 N 0.18 -0.26 Intersections logRSTT -0.24 Figure 4-3. Pearson correlation coefficients between candidate proxy variables that are correlated with at least one repository performance QoI. Density and number of intersections are significantly correlated with average degree, and mean log10 permeability is correlated with logRSTT. variables, for example using physics-based, data-driven methods to identify β€œfingerprints” driving performance QoIs [44]. However, such methods can require significant training data, which in our case means repository model evaluations. Given the high computational cost of our model, we limited the computational effort associated with proxy variable derivation and selection and accepted any resulting lack of predictivity in mapping from proxy variables to performance QoIs. This lack of predictivity is addressed in our approach to PCE construction, discussed below. 4.1.2. Incorporating proxy variables in PCE construction As discussed in Section 3.1, PCEs are stochastic expansions with basis functions defined to be orthonormal with respect to the probability densities of the uncertain inputs to the expansion. Thus, to incorporate proxy variables into a PCE surrogate, we must endow the proxy variables with analytical probability density functions (PDFs). To do so, we generated 1080 DFN realizations, computed their corresponding proxy variables, then fit analytical PDFs to the proxy variable samples. This is feasible since generating DFN realizations is not a significant computational cost relative to running the repository model. The appropriate PDF for each proxy variable was selected by visual inspection of their histograms, then PDF hyperparameters were fit computationally. Fitted normal PDFs for the proxy variables are shown in Figure 4-4. In this case visual inspection was sufficient to determine the appropriate PDF, but this may not hold in general. In the future, it may be necessary to develop a framework that automatically tests a range of transformations, scalings, and density types to find optimal analytical distributions for the proxy variables. 27 Histograms PDF Density fit 3.4 3.5 3.6 58.8 82.8 106.9 βˆ’2.2 0.0 2.3 avgDegree NIwR logRST T Figure 4-4. Density fits to the proxy variables, compared to histograms with 10, 20, and 30 bins to show the robustness of the empirical distribution. 4.2. Regression PCE approach Our PCE surrogate construction approach takes into account the fact that (1) proxy variables cannot be specified by a surrogate construction algorithm but must be derived from a DFN realization and (2) the β€œnoise” in repository performance QoIs as a function of proxy variables due to their imperfect/incomplete representation of spatial heterogeneity from the DFNs. To address (1), we employ a regression PCE approach as discussed in Section 3.1, which solves for the coefficients of the PCE given input-output samples. When the collocation ratio exceeds 1, standard regression is used to fit PCE coefficients; below 1, LASSO is used. To address (2), we investigate the use of K-fold cross-validation to select the PCE order that minimizes cross validation error, mitigating the chance of overfitting to β€œnoise” in the outputs. 10-fold cross-validation is employed when doing so would not reduce the collocation ratio below 1, else one-at-a-time cross validation is performed. 4.3. Accuracy checks We employ multiple diagnostics to assess the quality of our results. To assess the goodness of fit of our surrogate, we compute the 𝑅 2 of the surrogate over build points and generate scatterplots comparing build points to PCE predictions at those points. Additionally, we corroborate our PCE- based Sobol’ indices by comparing them to the given-data main effects estimators. As a measure of efficacy of the proxy variables to represent the influence of spatially heterogeneous uncertainty from DFNs, we compare the empirical Sobol’ index for DFN realization to the grouped PCE-based Sobol’ index for the group of proxy variables. Finally, and perhaps most importantly, we cross- reference our Sobol’ indices with input-output scatterplots to verify whether we observe similar trends. 28 5. RESULTS The full set of uncertain parameters considered is presented in Table 5-1. All UQ studies were carried out using Dakota, an optimization and UQ software, which has implemented the most popular multimodel sampling-based methods and single- and multimodel PCEs [1]. All PCEs were constructed using regression as described in Section 4.2. We considered PCEs constructed with and without cross validation with a maximum total order of no more than 3. Parameter Description Distribution name log U 10βˆ’15 , 10βˆ’13   kGlacial Glacial aquifer permeability [m2 ] log U 10βˆ’19 , 10βˆ’16   Material permDRZ DRZ permeability [m2 ] properties log U 10βˆ’20 , 10βˆ’17   permBuffer Buffer permeability [m2 ] pBuffer Buffer porosity [-] U [0.3, 0.5] Waste- IRF Instant release fraction, fraction of U [0.038, 0.156] related radionuclide inventory in a waste properties package released instantly upon waste package breach [-] meanWPRate Mean of the truncated log normal U [βˆ’5.5, βˆ’4.5] distribution on base normalized gen- eral corrosion rate (R) [log yrβˆ’1 ] stdWPRate Standard deviation of the truncated U [0.15, 0.4] log normal distribution [log yrβˆ’1 ] N 3.48, (0.0376) 2  avgDegree Average degree (average intersec- DFN tions per fracture) of the DFN [-] proxy N 82.8, (12.0) 2  NIwR Number of intersections with the variables repository region [-] N 0.0227, (1.13) 2  logRSTT Log10 relative shortest travel time be- tween repository and aquifer [-] Table 5-1. Descriptions and distributions for the parameters included in the UQ studies. Uniform (and log-uniform) distributions with lower and upper bounds 𝑙𝑏 and 𝑒𝑏 are denoted U [𝑙𝑏, 𝑒𝑏] (and log U [𝑙𝑏, 𝑒𝑏]). Normal distributions with mean πœ‡ and standard deviation 𝜎 are denoted N (πœ‡, 𝜎 2 ). We first compute the first-order indices directly using the 1000 input-output samples and the binning-based method defined in Section 3.2.1. Sensitivity to DFN uncertainty is computed with 29 respect to DFN realization directly as a baseline for accuracy of Sobol’ indices computed using the PCE with proxy variables. Statistical sampling error may be significant with only 1000 samples. We thus use 100 bootstrap samples, which resample the original sample set with replacement, to estimate the uncertainty in the computed indices [4]. These indices are presented in Figure 5-1 for a subset of repository performance QoIs that exhibit unique sensitivities (for example, we do not present sensitivities for the π‘₯ location of peak 129 I concentration in the aquifer because they are similar to the 𝑦 location). The 95% confidence intervals for the bootstrap samples are presented in orange; these intervals overlap with the empirical indices, and the bounds are relatively tight and non-overlapping, indicating there is low enough statistical error to confidently rank the inputs in order of importance to output QoIs. Peak 129 I in MTT at Peak Peak 129 I MdRT of Aquifer [M] 129 I [yr] y Location [m] Repository [yr] permDRZ kGlacial IRF pBuffer DFN 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 FMF from Repo FMF from Repo AqEb RockAq RockEb MFR RockEb MFR at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] Ratio at 1 MYr Ratio at 1 Myr permDRZ kGlacial IRF pBuffer DFN 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 Figure 5-1. Empirical first-order Sobol’ indices for repository performance QoIs computed using 1000 samples (blue bars) and the 95% confidence interval for 100 bootstrap samples (orange). DFN realization is a predominant source of uncertainty for many of the repository performance QoIs. It is the sole source of uncertainty for all the QoIs related to peak 129 I concentration in the aquifer, so inclusion of proxy variables for surrogate model based GSA will be critical for these QoIs. It is interesting to note that the fractional mass flow rate from the repository exhibits different sensitivities at different time points in the simulation. In early time, the buffer permeability is significant, while in later time DFN realization dominates. We now turn our attention to PCE surrogates constructed with proxy variables, assessing the accuracy of QoI predictions and associated Sobol’ indices. We consider three cases: a 2𝑛𝑑 -order PCE without cross-validation, a 2𝑛𝑑 -order PCE with cross-validation, and a 3π‘Ÿ 𝑑 -order PCE with cross-validation. We consider all quality checks previously discussed in Section 4.2. To simplify 30 discussion we focus on two important and illustrative repository performance QoIs: the peak 129 I concentration in the aquifer and the median residence time in the repository. Results for all QoIs are presented in A. PCE accuracy checks for our three cases are presented in Figure 5-2. Scatterplots show the spread in PCE predictions relative to the build points upon which they were constructed. Each scatterplot also reports the 𝑅 2 value for these points, which is the fraction of total variance in the testing dataset that is captured in the surrogate model. A perfect surrogate fit would result in all points falling along the 𝑦 = π‘₯ line and an 𝑅 2 = 1. First-order Sobol’ indices were computed using the build points using the binned method discussed in Section 3.2.1, treating DFN realization as a categorical random variable. The grouped DFN index for the PCE based Sobol’ index was computed for the group of proxy variables as described in Section 4.2. The grouped index is thus the sum of the first-order indices for avgDegree, NIwR, and logRSTT, as well as the interaction effects between them. 31 Peak 129 I in Aquifer [M] Order 2 PCE Order 2 PCE Order 3 PCE no cross-validation with cross-validation with cross-validation 3.5e-8 PCE samples -3e-9 R2 = 0.57 R2 = 0.57 R2 = 0.90 8e-11 4.2e-8 8e-11 4.2e-8 8e-11 4.2e-8 Model samples Model samples Model samples permDRZ kGlacial IRF pBuffer DFN 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 Main effect index Main effect index Main effect index MdRT of Repository [yr] 2.9e5 PCE samples 7e4 R2 = 0.83 R2 = 0.83 R2 = 0.94 7e4 2.9e5 7e4 2.9e5 7e4 2.9e5 Model samples Model samples Model samples permDRZ kGlacial IRF pBuffer DFN 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 Main effect index Main effect index Main effect index Figure 5-2. A comparison of results for 2𝑛𝑑 -order PCEs without and with cross-validation and a 3π‘Ÿ 𝑑 -order PCE with cross-validation. PCE results are presented columnwise. Scatterplots depict the PCEs evaluated at the build points vs. the build points. Bar charts depict the PCE-based Sobol’ indices (green) vs. those computed using the build points (blue) and the 95% confidence interval for 100 bootstrap samples (orange). 32 If we were to only judge the quality of the PCE surrogates based on scatterplots and 𝑅 2 values, we would mistakenly deduce that the 3π‘Ÿ 𝑑 -order PCE was most accurate. The scatterplots are much more concentrated and linear for the 3π‘Ÿ 𝑑 -order PCE, and the 𝑅 2 values are significantly higher. The Sobol’ main effects indices agree well with the binning-based indices for peak 129 I concentrations in all PCE cases. However, the 3π‘Ÿ 𝑑 -order PCE-based indices disagree significantly with the binning- based indices for the median residence time QoIβ€”the disagreement is so severe that the indices misrank the importance of the uncertain inputs for this QoI. As shown in Figure A-6, it misranks uncertain input importance for other QoIs, including the FMF from the repository at 3 kyr and the ratio of mass flow rates to the east boundary. We hypothesize that this disagreement comes from the 3π‘Ÿ 𝑑 -order PCE β€œfitting to noise” in this case, since (as discussed in Section 4.1.1) the proxy variables do not perfectly capture the influence of spatial heterogeneity in the DFNs on the repository performance QoIs, but further investigations would be required to test this hypothesis. While we observe this effect for a 3π‘Ÿ 𝑑 -order PCE in this case, we emphasize that the best choice of surrogate model order, with or without cross-validation, will be problem dependent. However, our approach to assessing the accuracy of the surrogate can be applied in general. Note that all the PCEs allow for negative 129 I concentrations, which is nonphysical. As shown in Figure 5-3, while the PCE allows for negative values, most samples are positive, and the density of the 129 I QoI agrees relatively well with the samples from the PFLOTRAN repository model. To eliminate this effect, one could construct the PCE on log-transformed concentration samples. However, because we would be introducing a nonlinear transformation of the output in doing this, the Sobol’ indices computed using the coefficients of the PCE would not be representative of the sensitivities for the untransformed concentrations. It would be possible to compute sample-based indices by querying the PCE and exponentiating its prediction to obtain untransformed 129 I samples, but exponentiation may further pronounce inaccuracies in the PCE surrogate. Since the PCE and PFLOTRAN-based probability densities agree relatively well, and since the corresponding Sobol’ indices agree, we do not investigate methods to impose positivity on the surrogate model herein. Model samples Density PCE samples 0.0 2.00e-8 4.000e-8 129 Peak I in Aquifer [M] Figure 5-3. A histogram comparing the 2𝑛𝑑 order, no cross-validation PCE predictions of peak 129 I concentration in the aquifer, vs. the build points. We now compare our sensitivity results to input-output scatterplots to assess the veracity of our Sobol’ index results. We focus here on the peak 129 I concentration on the aquifer as one of the most important QoIs, and because it provides a good illustration of the importance of this step in the analysis. Recall from Figure 5-2 that the surrogate and empirical first-order Sobol’ indices identify DFN as the sole individual source of uncertainty in the peak 129 I QoI. However, the empirical 33 index is less than 1, indicating some fraction of variance can be attributed to interactions between inputs. The importance of DFN is reflected in the input-output scatterplot of DFN realization vs. peak 129 I (note that we have sorted the DFN realizations according to the variance in the QoI to more clearly reveal the dependence), where the increase in variance over DFN realizations indicates an interaction effect. Similarly, the significant reduction in variance from smaller to larger kGlacial values indicates an interaction effect and significant influence from kGlacial. These visual cues thus corroborate the first-order empirical Sobol’ indices. Conversely, all the surrogates here attribute nearly 100% of the variance to DFN. Additional improvements to the surrogate model could target capturing these interaction effects more accurately. Peak 129 I in Aquifer [M] 4e-8 2e-8 0 1 25 0 2 75 100 3.45 3.50 3.55 DFN logRSTT NIwR AvgDegree 4e-8 2e-8 0 0.3 0.4 0.5 βˆ’5.5 βˆ’5.0 βˆ’4.5 0.2 0.4 0.05 0.10 0.15 pBuffer meanWPRate stdWPRate IRF 4e-8 2e-8 0 0 1e-13 0 1e-16 0 1e-17 kGlacial permDRZ permBuffer Figure 5-4. Input-output scatterplots for the peak 129 I concentration in the aquifer. The DFN indices have been sorted according to QoI variance from smallest to largest to make the dependence more clear. The scatterplots of proxy variables vs. peak 129 I concentration in the 2𝑛𝑑 to 4π‘‘β„Ž plots on the top row of Figure 5-4 indicate a weak dependence on these proxy variables. Further investigation would be needed to discern whether correlations with these proxy variables were driven by the most extreme DFN realizations that resulted in the highest variance of the QoI; additionally, further investigations considering interaction scatterplots of pairs of proxy variables vs. QoI would help identify whether interactions between proxy variables are influential and consistent with those identified by the PCE surrogates. 34 These results demonstrate that typical surrogate quality metrics like 𝑅 2 and methods to control for overfitting like cross validation are not adequate when incorporating proxy variables that approx- imate the influence of unparameterized sources of uncertainty, like the DFNs in this application problem. Comparing to sensitivity indices computed directly using input-output samples from our simulation model was critical to assessing the quality of our surrogate models. Finally, comparing to input-output scatterplots is a simple and highly important analysis step to corroborate sensitivity indices, whether sample- or surrogate-based. By combined assessment of scatterplots, 𝑅 2 , and comparison of sensitivity indices in Figure 5-2, the 2𝑛𝑑 order PCE with cross-validation appears to achieve the best accuracy overall. However, with an 𝑅 2 of only 0.57 for the peak 129 I QoI, it is clear that the fit is suboptimal. The surrogate could be further improved by seeking proxy variables that more accurately capture the influence of spatial heterogeneity on repository performance QoIs. 35 36 6. CONCLUSIONS Unparametrized spatially heterogeneous uncertainties often arise in geophysical application prob- lems, e.g., probabilistic post-closure performance assessment for deep geologic repositories for nuclear waste. These spatially heterogeneous uncertainties are often a significant, even dominant, source of uncertainty in model predictions. However, it is not straightforward to incorporate them into uncertainty analyses, especially sensitivity analyses, which require uncertainty sources to be parametrized. It is thus desirable to develop proxy variables that capture the mechanisms by which these uncertainties influence model outputs, so that they can be incorporated into uncertainty analyses, especially sensitivity analysis. In this work we present methods by which such proxy variables can be developed, selected, and incorporated into surrogate models for global sensitivity analysis. We also demonstrate how surrogates constructed using these proxy variables can be assessed for accuracy. In applying these methods to our nuclear waste repository application problem, understanding and separating the effect of spatial heterogeneity on repository QoIs has proven challenging, especially with the limited number of DFNs and parameter samples that can be run for realistic problem sizes. Identifying predictive proxy variables that adequately capture the influence of the spatially heterogeneous uncertainties is nontrivial. Additionally, we demonstrate that typical methods to assess surrogate accuracy and control overfitting (such as scatterplots, 𝑅 2 values, and cross-validation) are not sufficient when incorporating proxy variables that capture some, but not all, of the impacts spatially heterogeneous uncertainties have on outputs. We found that additional comparison of surrogate-based sensitivity indices to binning-based indices computed from input- output samples directly was important to ensure the surrogate was appropriately capturing model dependencies. Additionally, we found that comparison of binning- and surrogate-based indices to input-output scatterplots is a critical analysis step to corroborate the dependencies identified by sensitivity indices. There is still significant work required to develop general methods for extracting accurate proxy variables from unparametrized spatially heterogeneous uncertainties. Data-driven methods like variational autoencoders, which map high-dimensional uncertain inputs to low-dimensional latent embeddings, pose an opportunity to identify more nuanced proxy variables than those developed herein. Finally, further work could be done to develop a more nuanced approach to downselect proxy variables for individual QoIs, in contrast to the uniform downselect performed here. 37 38 APPENDIX A. Expanded numerical results Numerical results for all repository performance QoIs and PCE configurations are presented herein. In Figures A-1, A-3 and A-5 scatterplots depict build points (from PFLOTRAN model evaluations) vs. the PCEs evaluated at the build points. 𝑅 2 values computed at the build points are reported per QoI. In Figures A-2, A-4 and A-6 we compare the PCE-based first-order Sobol’ indices (green) to the binning-based indices (blue) and 100 bootstrap samples (orange). The binning-based indices for DFN were computed by interpreting DFN realization as a categorical random variable. The PCE-based indices for DFN were computed as the grouped first-order index of the proxy variables avgDegree, NIwR, and logRSTT (recall that the grouped index is the sum of main effect indices and all interaction effects between the variables in the group). Numerical trends are as follows. 𝑅 2 values for most QoIs were somewhat low for the 2𝑛𝑑 order PCEs. They were significantly higher overall for the 3π‘Ÿ 𝑑 order PCE, but disagreement of PCE- based Sobol’ indices in comparison to the binning-based indices indicate the 3π‘Ÿ 𝑑 order PCE is overfitting in order to achieve this increased 𝑅 2 value. The 2𝑛𝑑 order PCE exhibited the best balance of agreement with binning-based Sobol’ indices and 𝑅 2 values, but it is by no means a perfect surrogate. Further development is needed to develop more accurate proxy variables, which would in turn improve surrogate accuracy. 39 2nd order PCE, no cross-validation Peak 129 I in MTT at Peak Peak 129 I Aquifer [M] 129 I [yr] y Location [m] 2.4e-8 8.8e5 1.7e3 -2e-9 R2 = 0.57 2e5 R2 = 0.62 2e2 R2 = 0.51 8e-11 4.2e-8 1e4 8.4e5 7e1 2.0e3 MdRT of FMF from Repo FMF from Repo Repository [yr] at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] 2.8e5 2.7e-5 1.6e-6 PCE samples 6e4 R2 = 0.83 1e-5 R2 = 0.86 6e-7 R2 = 0.45 7e4 2.9e5 1e-5 2.9e-5 6e-7 1.7e-6 AqEb RockAq RockEb MFR RockEb MFR Ratio at 1 MYr Ratio at 1 Myr 1.4 1.1 -2e-1 R2 = 0.81 -6e-1 R2 = 0.23 8e-3 2.9 -7e-1 3.5 Model samples Figure A-1. Scatterplots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. 40 2nd order PCE, no cross-validation Peak 129 I in MTT at Peak Peak 129 I MdRT of Aquifer [M] 129 I [yr] y Location [m] Repository [yr] permDRZ kGlacial IRF pBuffer DFN FMF from Repo FMF from Repo AqEb RockAq RockEb MFR RockEb MFR at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] Ratio at 1 MYr Ratio at 1 Myr permDRZ kGlacial IRF pBuffer DFN Main effects indices Figure A-2. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). 41 2nd order PCE with cross-validation Peak 129 I in MTT at Peak Peak 129 I Aquifer [M] 129 I [yr] y Location [m] 2.4e-8 8.6e5 1.5e3 -2e-9 R2 = 0.57 2e5 R2 = 0.61 3e2 R2 = 0.50 8e-11 4.2e-8 1e4 8.4e5 7e1 2.0e3 MdRT of FMF from Repo FMF from Repo Repository [yr] at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] 2.7e5 2.7e-5 1.5e-6 PCE samples 6e4 R2 = 0.83 1e-5 R2 = 0.86 6e-7 R2 = 0.43 7e4 2.9e5 1e-5 2.9e-5 6e-7 1.7e-6 AqEb RockAq RockEb MFR RockEb MFR Ratio at 1 MYr Ratio at 1 Myr 1.3 7.5e-1 -3e-2 R2 = 0.78 -5e-1 R2 = 0.21 8e-3 2.9 -7e-1 3.5 Model samples Figure A-3. Scatter plots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. 42 2nd order PCE with cross-validation Peak 129 I in MTT at Peak Peak 129 I MdRT of Aquifer [M] 129 I [yr] y Location [m] Repository [yr] permDRZ kGlacial IRF pBuffer DFN FMF from Repo FMF from Repo AqEb RockAq RockEb MFR RockEb MFR at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] Ratio at 1 MYr Ratio at 1 Myr permDRZ kGlacial IRF pBuffer DFN Main effects indices Figure A-4. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). 43 3rd order PCE with cross-validation Peak 129 I in MTT at Peak Peak 129 I Aquifer [M] 129 I [yr] y Location [m] 3.5e-8 8.5e5 1.8e3 -3e-9 R2 = 0.90 8e4 R2 = 0.78 2e2 R2 = 0.70 8e-11 4.2e-8 1e4 8.4e5 7e1 2.0e3 MdRT of FMF from Repo FMF from Repo Repository [yr] at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] 2.9e5 2.8e-5 1.7e-6 PCE samples 7e4 R2 = 0.94 1e-5 R2 = 0.96 5e-7 R2 = 0.75 7e4 2.9e5 1e-5 2.9e-5 6e-7 1.7e-6 AqEb RockAq RockEb MFR RockEb MFR Ratio at 1 MYr Ratio at 1 Myr 2.0 3.0 -9e-2 R2 = 0.90 -8e-1 R2 = 0.86 8e-3 2.9 -7e-1 3.5 Model samples Figure A-5. Scatter plots of build points from PFLOTRAN simulations vs. PCE predictions at the build points. The 𝑅 2 value for each output is reported in the bottom right of its subplot. 44 3rd order PCE with cross-validation Peak 129 I in MTT at Peak Peak 129 I MdRT of Aquifer [M] 129 I [yr] y Location [m] Repository [yr] permDRZ kGlacial IRF pBuffer DFN FMF from Repo FMF from Repo AqEb RockAq RockEb MFR RockEb MFR at 3 kyr [yrβˆ’1 ] at 1 Myr [yrβˆ’1 ] Ratio at 1 MYr Ratio at 1 Myr permDRZ kGlacial IRF pBuffer DFN Main effects indices Figure A-6. Comparison of empirical first-order Sobol’ indices (blue) and bootstrap sample 95% confidence intervals (orange) vs. PCE-based indices (green). 45 46 References [1] Brian M Adams, William Bohnhoff, Keith Dalbey, Mohamed Ebeida, John Eddy, Michael Eldred, Russell Hooper, Patricia Hough, Kenneth Hu, and John Jakeman. Dakota A Mul- tilevel Parallel Object-Oriented Framework for Design Optimization Parameter Estimation Uncertainty Quantification and Sensitivity Analysis: Version 6.14 User’s Manual. Technical report, Sandia National Lab.(SNL-NM), Albuquerque, NM (United States), 2021. [2] Christian Andresen, Alex Hansen, Romain Le Goc, Philippe Davy, and Sigmund Hope. Topology of fracture networks. Frontiers in Physics, 1:7, 2013. [3] Emanuele Borgonovo, Gordon B. Hazen, and Elmar Plischke. A Common Rationale for Global Sensitivity Measures and Their Estimation. Risk Analysis, 36(10):1871–1895, October 2016. Publisher: John Wiley & Sons, Ltd. [4] Michael R. Chernick. Bootstrap Methods: A Guide for Practitioner and Researchers. Wiley Series in Probability and Statistics. John Wiley & Sons, Inc., 2nd edition edition, 2008. [5] Sven Follin, Lee Hartley, Ingvar RhΓ©n, Peter Jackson, Steven Joyce, David Roberts, and Ben Swift. A methodology to constrain the parameters of a hydrogeological discrete fracture network model for sparsely fractured crystalline rock, exemplified by data from the pro- posed high-level nuclear waste repository site at Forsmark, Sweden. Hydrogeology Journal, 22(2):313–331, March 2014. [6] G. A. Freeze, M. D. Voegele, P. Vaughn, J. Prouty, and W. M. Nuttand E. L. Hardin. Generic Deep Geologic Disposal Safety Case. Technical Report SAND2013-0974P, Sandia National Laboratories, Albuquerque, New Mexico, 2013. [7] G. Aldrich, J. D. Hyman, S. Karra, C. W. Gable, N. Makedonska, H. Viswanathan, J. Woodring, and B. Hamann. Analysis and Visualization of Discrete Fracture Networks Us- ing a Flow Topology Graph. IEEE Transactions on Visualization and Computer Graphics, 23(8):1896–1909, August 2017. [8] HO Ghaffari, MHB Nasseri, and RP Young. Fluid flow complexity in fracture networks: Analysis with graph theory and LBM. arXiv preprint arXiv:1107.4918, 2011. [9] G.E. Hammond, P.C. Lichtner, R.T. Mills, and C. Lu. Toward petascale computing in geo- sciences: application to the hanford 300 area. In Journal of Physics, Conference Series 125 012051. SciDAC 2008: Scientific Discovery through Advanced Computing, 2008. [10] Glenn E Hammond, Peter C Lichtner, and RT Mills. Evaluating the performance of parallel subsurface simulators: An illustrative example with PFLOTRAN. Water resources research, 50(1):208–228, 2014. 47 [11] Lee Hartley and Steven Joyce. Approaches and algorithms for groundwater flow modeling in support of site investigations and safety assessment of the Forsmark site, Sweden. Journal of Hydrology, 500:200–216, September 2013. [12] B.A Howard, M.B Crawford, D.A Galson, and M.G Marietta. Regulatory basis for the Waste Isolation Pilot Plant performance assessment. Reliability Engineering and System Safety, 69(1):109–127, 2000. [13] J. D. Hyman and J. JimΓ©nez-MartΓ­nez. Dispersion and Mixing in Three-Dimensional Discrete Fracture Networks: Nonlinear Interplay Between Structural and Hydraulic Heterogeneity. Water Resources Research, 54(5):3243–3258, May 2018. Publisher: John Wiley & Sons, Ltd. [14] Jeffrey D. Hyman, Aric Hagberg, Gowri Srinivasan, Jamaludin Mohd-Yusof, and Hari Viswanathan. Predictions of first passage times in sparse discrete fracture networks using graph-based reductions. Physical Review E, 96(1):013304, July 2017. Publisher: American Physical Society. [15] Jeffrey D. Hyman, Satish Karra, Nataliia Makedonska, Carl W. Gable, Scott L. Painter, and Hari S. Viswanathan. dfnWorks: A discrete fracture network framework for modeling subsurface flow and transport. Computers & Geosciences, 84:10–19, November 2015. [16] J.D. Jakeman, M.S. Eldred, and K. Sargsyan. Enhancing β„“1-minimization estimates of poly- nomial chaos expansions using basis selection. Journal of Computational Physics, 289:18–34, May 2015. [17] S. Karra, D. O’Malley, J. D. Hyman, H. S. Viswanathan, and G. Srinivasan. Modeling flow and transport in fracture networks using graphs. Physical Review E, 97(3):033304, March 2018. Publisher: American Physical Society. [18] Tara Catherine LaForce, Eduardo Basurto, Lisa Amelia Bigler, Kyung Won Chang, Mo- hamed Salah Ebeida, Forest Taylor Good, Franciszek Jozef Hasiuk, Richard Scott Jayne, Jordan Leo Jensen, Rosemary Claire Leone, Paul Mariner, and Jeff Homer Sharpe. GDSA Repository Systems Analysis Investigations in FY 2024. Technical report, United States, 2. [19] Chenzhao Li and Sankaran Mahadevan. An efficient modularized sample-based method to estimate the first-order Sobol’ index. Reliability Engineering & System Safety, 153:110–121, September 2016. [20] P.C. Lichtner and G.E. Hammond. Quick reference guide: PLFOTRAN 2.0 (la-cc-09-047) multiphase-multicomponent-multiscale massively parallel reactive transport code. Technical Report LA-UR-06-7048, Los Alamos National Laboratory, Los Alamos, New Mexico, 2012. [21] Peter C Lichtner, Glenn E Hammond, Chuan Lu, Satish Karra, Gautam Bisht, Benjamin Andre, Richard Mills, and Jitendra Kumar. PFLOTRAN user manual: A massively parallel reactive flow and transport model for describing surface and subsurface processes. Technical Report LA-UR-15-20403, Los Alamos National Laboratory (LANL), Los Alamos, NM; Sandia National Laboratories (SNL), Albuquerque, NM, 2015. 48 [22] Paul Mariner, Michael Nole, Eduardo Basurto, Timothy M Berg, Kyung Won Chang, Bert Debusschere, Aubrey Celia Eckert, Mohamed Salah Ebeida, Michael B Gross, and Glenn Hammond. Advances in GDSA Framework Development and Process Model Integration. Technical report, Sandia National Lab.(SNL-NM), Albuquerque, NM (United States); San- dia . . . , 2020. [23] Paul E. Mariner, Emily R. Stein, Jennifer M. Frederick, S. David Sevougian, Glenn Edward Hammond, and D. G. Fascitelli. Advances in Geologic Disposal System Modeling and Application to Crystalline Rock. Technical report, United States, September 2016. [24] P.E. Mariner, E.R. Stein, J.M. Frederick, S.D. Sevougian, G.E. Hammond, and D.G. Fascitelli. Advances in geologic disposal system modeling and application to crystalline rock. Tech- nical Report FCRD-UFD-2016-000440/SAND2016-9610 R, Sandia National Laboratories, Albuquerque, New Mexico, 2016. [25] NEA. Establishing and Communicating Confidence in the Safety of Deep Geologic Disposal: Approaches and Arguments. Technical report, Organisation for Economic Co-operation and Development, Paris, France, 2002. [26] NEA. Post-closure Safety Case for Geologic Repositories: Nature and Purpose. Technical Report NEA No. 3679, Organisation for Economic Co-operation and Development, Paris, France, 2004. [27] NEA. International Experiences in Safety Cases for Geological Repositories (INTESC). Technical Report NEA No. 6251, Organisation for Economic Co-operation and Development, Paris, France, 2009. [28] Elmar Plischke, Emanuele Borgonovo, and Curtis L. Smith. Global sensitivity measures from given data. European Journal of Operational Research, 226(3):536–550, May 2013. [29] ClΓ©mentine Prieur and Stefano Tarantola. Variance-Based Sensitivity Analysis: Theory and Estimation Algorithms. In Roger Ghanem, David Higdon, and Houman Owhadi, editors, Handbook of Uncertainty Quantification, pages 1217–1239. Springer International Publish- ing, Cham, 2017. [30] Rob P. Rechard, Thomas A. Cotton, and Michael D. Voegele. Site selection and regulatory basis for the Yucca Mountain disposal system for spent nuclear fuel and high-level radioactive waste. Reliability Engineering and System Safety, 122:7–31, 2014. Special Issue: Performance Assessment for the Proposed High-Level Radioactive Waste Repository at Yucca Mountain, Nevada. [31] Andrea Saltelli, Paola Annoni, Ivano Azzini, Francesca Campolongo, Marco Ratto, and Stefano Tarantola. Variance based sensitivity analysis of model output. Design and estimator for the total sensitivity index. Computer Physics Communications, 181(2):259–270, February 2010. [32] Elizabeth Santiago, Jorge X. Velasco-HernΓ‘ndez, and Manuel Romero-Salcedo. A method- ology for the characterization of flow conductivity through the identification of communities in samples of fractured rocks. Methods and Applications of Artificial and Computational Intelligence, 41(3):811–820, February 2014. 49 [33] Jan-Olof Selroos, Douglas D. Walker, Anders StrΓΆm, BjΓΆrn Gylling, and Sven Follin. Com- parison of alternative modelling approaches for groundwater flow in fractured rock. Journal of Hydrology, 257(1):174–188, February 2002. [34] Emily Stein, Jennifer M Frederick, Glenn Edward Hammond, Kristopher L Kuhlman, Paul Mariner, and S David Sevougian. Modeling Coupled Reactive Flow Processes in Fractured Crystalline Rock. In Proceedings of the 16th International High-Level Radioactive Waste Management Conference, Charlotte, North Carolina, 2017. [35] Steven Joyce, Lee Hartley, David Applegate, Jaap Hoek, and Peter Jackson. Multi-scale groundwater flow modeling during temperate climate conditions for the safety assessment of the proposed high-level nuclear waste repository site at Forsmark, Sweden. Hydrogeology Journal, 22(6):1233–1249, September 2014. [36] Bruno Sudret. Global sensitivity analysis using polynomial chaos expansions. Bayesian Networks in Dependability, 93(7):964–979, July 2008. [37] Laura Painton Swiler, Eduardo Basurto, Dusty M. Brooks, Tara LaForce, Rosemary Leone, Paul E. Mariner, Teresa Portone, Caitlin Condon, Josh Hargraves, and Tristan Hay. Uncertainty and Sensitivity Analysis Methods and Applications in the GDSA Framework (FY2024). Technical Report SAND2024-11075R, United States, August 2024. [38] Laura Painton Swiler, Eduardo Basurto, Dusty Marie Brooks, Aubrey Celia Eckert, Paul Mariner, Teresa Portone, and Emily Stein. Advances in Uncertainty and Sensitivity Analysis Methods and Applications in GDSA Framework. Technical Report SAND2020-10802R, Sandia National Lab.(SNL-NM), Albuquerque, NM (United States), 2020. [39] L.P. Swiler, E. Basurto, D.M. Brooks, A.C. Eckert, R. Leone, P.E. Mariner, T. Portone, and M.L. Smith. Uncertainty and sensitivity analysis methods and applications in the GDSA framework (fy2022). Technical Report M3SF-23SN01030408, SAND2022-11220 R, Sandia National Laboratories, Albuquerque, New Mexico, 2022. [40] L.P. Swiler, E. Basurto, D.M. Brooks, A.C. Eckert, R. Leone, P.E. Mariner, T. Portone, M.L. Smith, and E.R. Stein. Uncertainty and sensitivity analysis methods and applications in the GDSA framework (fy2021). Technical Report M3SF-21SN010304042, SAND2021-9903 R, Sandia National Laboratories, Albuquerque, New Mexico, 2021. [41] L.P. Swiler, E. Basurto, D.M. Brooks, A.C. Eckert, P.E. Mariner, T. Portone, and E.R. Stein. Advances in uncertainty and sensitivity analysis methods and applications in GDSA framework. Technical Report M3SF-20SN010304032, SAND2020-10802 R, Sandia National Laboratories, Albuquerque, New Mexico, 2020. [42] Manuel Valera, Zhengyang Guo, Priscilla Kelly, Sean Matz, Vito Adrian Cantu, Allon G. Percus, Jeffrey D. Hyman, Gowri Srinivasan, and Hari S. Viswanathan. Machine learning for graph-based representations of three-dimensional discrete fracture networks. Computational Geosciences, 22(3):695–710, June 2018. 50 [43] H. S. Viswanathan, J. D. Hyman, S. Karra, D. O’Malley, S. Srinivasan, A. Hagberg, and G. Srinivasan. Advancing Graph-Based Algorithms for Predicting Flow and Transport in Fractured Rock. Water Resources Research, 54(9):6085–6099, September 2018. Publisher: John Wiley & Sons, Ltd. [44] Elise Walker, Nathaniel Trask, Carianne Martinez, Kookjin Lee, Jonas A. Actor, Sourav Saha, Troy Shilt, Daniel Vizoso, Remi Dingreville, and Brad L. Boyce. Unsupervised physics- informed disentanglement of multimodal data. Foundations of Data Science, 7(1):418–445, March 2025. 51 DISTRIBUTION Emailβ€”Internal Name Org. Sandia Email Address Technical Library 1911 sanddocs@sandia.gov Hardcopyβ€”Internal Number of Name Org. Mailstop Copies Hardcopyβ€”External Number of Company Name and Name(s) Copies Company Mailing Address 52 53 Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.