WorksMathematics
Laplace approximation of the parameter posterior with embedded error for efficient uncertainty quantification of expensive model calibrations
Abstract
In practical engineering, there is a clear need for uncertainty quantification (UQ) of models used in design because forward UQ puts probabilities on performance predictions. However, without surrogate building, most forward engineering models are too expensive for traditional UQ methods like Markov chain Monte Carlo (MCMC) sampling. Our goal is to provide inexpensive, approxi- mate UQ that can be applied after an expensive traditional least squares fit, where the UQ handles both uncorrelated noise and parameter variance as in the embedded-error treatment developed by Sargsyan, Najm, and Ghanem (2015). After developing the basic theory to connect a Laplace-like approximation to Bayesian calibration, we compare the proposed method to black-box variational inference in terms of cost and fidelity. Then we demonstrate the method on real-world calibration of solid material models.
Introduction
Uncertainty quantification (UQ) provides confidence in simulations used in design and performance estimation. It can also uncover limitations of the calibration process, such as insufficient data and overly correlated parameters. However, the sampling-based methods, such as Markov chain Monte Carlo [1], are typically too computationally expensive (and temperamental in terms of convergence and hyperparameter settings) to deploy for complex material models embedded in the full simulations of experiments used for calibration. Inexact methods, such as variational inference [2], are more efficient but also can be prohibitively expensive for these scenarios, depending on the optimization technique [3].
Laplace approximation is a form of variational inference in the sense that a surrogate posterior is formed via an optimization procedure. It leads to Gaussian approximation of the observed outputs at the maximum a posteriori (MAP) point in parameter space. Since it only requires a Hessian at the optimum found by deterministic calibration, it can be extremely model sample/evaluation efficient, especially for a small number of parameters. It is a classical technique going back to Laplace’s method to approximate integrals [4, 5]. Here, we fuse this approach with the embedded error formulation [6] to accommodate both uncorrelated external noise and correlated internal variations that are evident in the experimental data [7]. The embedded error formulation allows the model parameters to be random variables, as opposed to the classical Kennedy-O’Hagan [8] external/additive error, which enables the formulation to capture correlated deviations from a mean response. This allows the separation of external, uncorrelated errors from internal, correlated errors. The first source of uncertainty is usually due to measurement resolution/precision limits while the second is an outcome of specimen (or environmental) variability. When there is significant material variability, the measurement noise is typically of smaller magnitude with modern testing equipment.
The inexact nature of the Laplace approximation is not a hindrance in our applications since the main goal is to cover likely responses in prediction and hence have credible estimates for performance evaluation. Although not a goal oriented [9, 10] technique, it has some commonalities with this more sophisticated class of methods.
In the next section we put the proposed method in the larger context of uncertainty quantification. Then, in § 3, we develop the theoretical and methodological basis for the proposed method while emphasizing the contribution of extending the Laplace method to embedded error and a straightforward linear algebra solution for the parameter covariance. In § 5 we explore the algorithmic properties of the method with simple examples before providing results for real-world material calibrations with diverse experimental data. Finally, we conclude with a summary and avenues for future work in § 7.
Related work
Uncertainty quantification (UQ) has a long history of development [1]. Although still the reference standard, Markov chain Monte Carlo (MCMC) sampling has a slow convergence, where is the number of samples. Newer techniques, such as Hamiltonian Monte Carlo (HMC) [11] and Langevin sampling [12, 13, 14], which employ gradients of the score and a connection to statistical physics, can be more sample efficient but still are prohibitive for expensive model evaluations, such as those from design-level finite element simulations. The cost of these methods has motivated optimization-based techniques, such as variational inference (VI) [2] where a surrogate posterior is determined by optimization. However, the more scalable mean field VI, which uses a diagonal covariance structure, is well known to under-predict the extent of posterior probabilities [15]. Hybrid methods, like Stein VI [16] and other ensemble/particle methods [17], try to obtain the best qualities of sampling and optimization techniques.
In all these approaches the primary concern, beyond the forward simulation cost, is size of the parameter space, whereby a large number of parameters to explore leads to the curse of dimensionality [18] due to the exponential rise in the number of samples needed to characterize a parameter or output covariance structure. This fundamental issue emphasizes the need for efficient UQ for highly parameterized, expensive models. The formulation of an appropriate likelihood tying observed data to model predictions is also a challenge. Kennedy and O’Hagan [8] proposed an additive, independent error model-observation discrepancy that has been widely adopted and is particularly well suited for representing (external) measurement noise. The Kennedy and O’Hagan approach leads to a straightforward likelihood formulation. In contrast, Sargsyan et al. [6, 19] proposed a correlated embedded error model whereby parameters are associated with random variables which induce a covariance structure on the pushed forward model responses.1 This approach suggests a number of possible likelihood formulations. These sources of discrepancies can be combined, but this increases the complexity of the likelihood and the number of hyperparameters that need to be selected/determined.
Calibration
In this section, we provide the theory necessary to understand the foundations of the proposed method. The classical Laplace approximation provides an analytical approximation to posterior probability using a quadratic approximation of the log posterior distribution. This Gaussian surrogate is typically centered at the maximum a posteriori (point) estimate of the parameters (which corresponds to the maximum likelihood fit for an uninformative prior). The precision matrix is given by the Hessian at the center, which is typically further approximated by a Gauss-Newton outer product of the model sensitivities (which is related to the Fisher information).
In the following developments, we start with least squares calibration (of an expensive model) given data, connect this to Bayesian calibration and, finally, to Laplace approximation. The use of an embedded error model is emphasized.
Data We consider experimental data that consists of input-output pairs , which are subject to external measurement noise. In many applications, such as material characterization via mechanical testing, measurement noise is high frequency and uncorrelated with the underlying response. On the other hand, the underlying response can have intrinsic/internal variation that systematically affects the data trends.
Least squares calibration In general, a model has parameters and produces output as a function of an input :
Given measured (experimental) data , the model can be calibrated using the usual least squares optimization
based on the (weighted) sum squares error (SSE) to obtain the best fit parameters . Here, the objective uses the residuals:
weighted by which quantifies the uncertainty in by identifying it with the inverse covariance (precision) matrix of the responses/outputs , as in standard generalized least squares calibration. Both and are of dimensions where is the length of ; hence experimental repeats are handled by making both and two dimensional with dimensions where is the number of (data) repeats. The inputs are the same dimensions as but identical/duplicated in the repeat dimension.
Bayesian calibration The observed data can be related to the model by adding (probabilistic) mechanisms for discrepancies
where we have output data that displays (relatively small) uncorrelated measurement error (noise) and correlated model inconsistencies (discrepancy) which we associate with , as mentioned. Both and are considered (independent) random variables. We associate with uncorrelated, irreducible measurement noise and with reducible, parameter uncertainty due to limited data. We make the common assumption that the measurement noise is mean-zero, independent, identically distributed (IID) and normal , where is, at this point, an unspecified hyperparameter. In the simplest case, we assume that both and are normally distributed. In this case, we need to determine and the mean and covariance so that
by the well-known sum of Gaussians rule. The mean parameters and covariance are subject to epistemic convergence; while the push-forward model response maintains a finite variance.2
This formulation naturally leads to Bayesian calibration [8, 1]. In fact, the least squares optimization (2) can be related to Bayesian calibration through Bayes rule
with posterior , likelihood , prior , and normalizing factor (evidence) . In the case of maximum likelihood estimation (MLE) (or maximum a posteriori estimation (MAP) with an uninformative prior )
A likelihood that corresponds to the MSE objective, :
provides the connection. Note, is a scalar, and units of likelihood, being similar to a probability density, are the inverse of those of the output . Here again, the expected precision appears and the likelihood is a multivariate normal in terms of the residuals , defined in Eq. (3), with parameters and . Hence, the log likelihood is
where the first two terms do not depend on the parameters and the last term is the weighted SSE . Note a double sum is implied in the last term, inner sum over and outer sum over , likewise for Eq. (8). Here we have introduced the notation
for an average over the repeats. This likelihood is consistent with the Kennedy and O’Hagan additive random discrepancy term [8].
Laplace approximation
In general, obtaining the moments that determine a Gaussian distribution in Eq. (8) analytically is difficult because the model is non-linear in its parameters . Laplace approximation is based on the connection between the SSE and the likelihood through a Taylor expansion of the log likelihood (9):
where is a perturbation around and the second term is zero at an optimum of , as defined in Eq. (2). The Hessian at consists of two terms:
the first is formed from the parameter sensitivities of the model , and the second term can be omitted if the model discrepancies are negligible or the model is linear in its parameters.3 Here are the (mean) parameter sensitivities at :
where is the input for the first repeat (which is equal to all other repeats). Now using the association of with from Eq. (10) we can form the approximation
where the Hessian acts as a precision and we have arbitrarily set .
Embedded Laplace Method
Here we enhance the classical Laplace approximation by adopting an embedded uncertainty [6] perspective where we treat the model parameters as random variables, e.g.
with mean , standard deviation , expressed as a Cholesky decomposition , and , and aim to determine the moments of that define these distributions. In this context, the parameters defining the distributions converge with increasing data, while distributions themselves maintain a finite width.
We can obtain the variance of the parameters through the linearization of the model at an optimum in Eq. (2), so that
If we assume the expectation of is approximately zero, then
via the variance rule of a Gaussian of a linear transformation, . 4 Note has the rank of , i.e. , and the dimensions of , i.e. , such that is and .
If we examine Eq. (17), it is clear that has the rank of and, hence, at most rank the number of parameters . For this work, we will assume is full rank due to the experimental loading fully exercising the model. So the solution for and splits into the span of and its complement. Since the complement only depends on , we use it to determine the aleatoric noise scale . By applying an eigenvalue decomposition of ,
the largest values determine and we can estimate with the average of the lowest singular values. Since the data can have more information content than the model applied to the same loading we omit the first singular values in this estimate. With determined, we have a least squares solution for
given is uncorrelated with , and .
In practice since we already have the eigensystem, we use the largest eigenvectors to form and and then solve
where, for numerical conditioning we use the (Moore-Penrose) pseudoinverse for .
Fit Subsequently, we can fit the parameter covariance using an optimization objective. In fact, we can determine by optimizing log posterior predictive (9) with the linearization of the model response
given . This optimization minimizes the SSE (the second term) regularized by the first term.
To start the optimization we use Eq. (20) with a sample covariance for as an initial guess. Also, for numerical conditioning, we decompose the covariance matrix into variances and correlation coefficients [20, 21, 22]:
with the diagonal of the covariance matrix and the correlation with off-diagonal components , which are reparameterized as: and .
Linear-in-parameters models The method is exact for linear-in-parameters models, and this case provides an analytical illustration. A general form for a model that is linear in its parameters is:
where the sensitivities are composed of basis-like functions ; for instance, for a model that is also linear in its inputs
where, here, we use for its average over repeats for clarity. In this case, Eq. (4) becomes:
where are samples of and are samples of , And the optimum . Hence,
and
if we assume: (a) and (b) so that .
Variational inference Like the embedded Laplace fitting technique, the VI technique minimizes the difference between the posterior and a surrogate through optimization of the parameters . The primary objective is the Kullback-Leibler divergence (KLdiv):
but the evidence lower bound (ELBO) is a more tractable optimization objective:
that combines a data consistency objective with a prior similarity regularizing term.
In the black box variational inference (BBVI) [23] formulation of mean field VI a Gaussian with a diagonal covariance (with elements ) and a uniform prior allow the reduction of the ELBO to:
The first term in Eq. (30) is the log likelihood from (9) (ignoring constant terms) and is model dependent, while the remainder only depends on the Gaussian prior, where is the entropy. For comparison, the entropy of full covariance Gaussian is:
Embedded variational inference To accommodate the embedded error formulation (4) within VI, we take the mean parameters from the MAP and determine the variance parameters through (subsequent) optimization. In particular we use the form:
where are the mean parameters and scale the contributions of independent Gaussian random variables . This is consistent with a mean field approximation and, hence, is in contrast with a Cholesky-like structure that would allow cross-correlations. For this formulation we chose a likelihood consistent with the Gaussian approximation to marginalized likelihood from Ref. [6]:
Error metrics To assess deterministic quantities we use the (usual) root mean square error (RMSE):
To assess probabilistic predictions, we use the Wasserstein-1 () distance to compare the associated distributions:
where the cumulative distribution functions (CDFs) result from empirical probability density functions (PDFs):
For two univariate Gaussians with the same mean, the Wasserstein-1 reduces to:
which provides the expected scaling of with the standard deviations of the PDFs being compared. For multivariate Gaussians, and , can be bounded:
which provides additional intuition.
Demonstrations
First we use simple calibrations to illustrate epistemic and aleatory behavior of the embedded Laplace (EL) and compare it with BBVI. Then, we apply the EL method to real-world calibration of complex material models to experimental data.
Illustrations
For the illustrations, we employ the two parameter models listed in Table 1 for easy interpretation and visualization. The linear and quadratic models are linear in their parameters and the third is nonlinear. For each data generating model, we take realizations of the parameters from where:
| linear | |
| quadratic | |
| nonlinear |
Table 1. Two parameter models
and is a scaling factor. The standard deviation of the external Gaussian noise was set to .
Aleatory/epistemic behavior
In this study, we fix the measurement noise and change parametric variance via . Figure 1a shows the relative error as measured by (as motivated by Eq. (37)). The trend is for the relative error to improve as the parametric variance becomes larger than the noise. Clearly, Figure 1b,Figure 1c show that the EL method provides predictions, as indicated by the 2 standard deviation bands, that cover the data, as depicted by the trajectory realizations, in both the large and small variance cases.
Figure 1. Illustration: (a) convergence in terms of a normalized distance and prediction comparisons for (b) high parametric variance, (c) low parametric variance, .
Comparison with VI
In this illustration, we compare BBVI to the proposed EL method. For the BBVI approach, we assume independent internal variance
in analog with mean field VI. Specifically
so that 4 parameters need to be determined.
For the EL method, we assume we can afford the cross-terms typically omitted in (mean field) VI. Figure 2 plots convergence of both the standard mean field VI and the embedded version versus the EL result. With standard VI the results collapses to the mean, as expected, while mean-field embedded version of VI maintains a finite variance but has errors due to the diagonal covariance assumption. On the other hand, the EL with a full covariance captures the data distribution accurately. With reference to Figure 3, clearly it takes thousands of epochs of gradient descent with a commensurate number of function evaluations. The optimization method (e.g. switching to a quasi-Newton method like BFGS [24] or learning rate (0.01 for this illustration) will change the convergence but not the conclusion that EL is substantially more model sample efficient. For this example, EL is both more accurate in terms of estimating the true variance and more efficient in terms of model evaluations than standard VI and a version that accommodates embedded parameter error.
Figure 2. Illustration: comparison of VI, embedded VI, EL push-forwards for a linear model with parametric variance (blue: data, red: prediction).
Figure 3. Illustration: convergence of embedded VI compared to EL in terms of the KL divergence between the underlying data variance and the predicted posterior. Note the difference in the KL divergence profiles for EL and the final embedded VI in Figure 3 are due to the assumed covariance structure, i.e. full for EL and diagonal for the embedded VI result.
Model discrepancy
In this last example we illustrate the behavior of the EL method with significant model discrepancy. For reference, Figure 4 shows the errors for the case where the data-generating and the fitting model are the same. The parameter means have lower errors than the covariance (a higher statistical moment), while errors for the linear algebra guess are on par with the calibrated . On the other hand, the push-forward response statistics show that the fitted covariance have marginally better error in some cases. Figure 5 shows corresponding results for the fitting model being not necessarily the same as the data generating model. In this case, if the fitting model is sufficiently similar to the data generating model, e.g. for a quadratic model fitted to nonlinear data, the errors, as measured by , are on par with the nonlinear model fitted to the same data; otherwise, the distance is substantially larger. Nevertheless, these results also show that the fitted can improve on the estimate but is not necessary for an engineering-level result.
Figure 4. Illustration: RMSE of the estimated mean and variance for no model discrepancy calibrations (a) parameter errors, (b) push-forward errors.
Figure 5. Illustration: distance for model discrepancy (a) quadratic data, (b) nonlinear data.
Applications
Our engineering-oriented applications involve experimental data and empirical material models. In particular, many high-consequence performance evaluations require predicting the likelihood of failure due to plastic deformation.
Plasticity is a history-dependent material response that has a recoverable elastic regime and an irreversible plastic regime that evolves with loading. The basic framework of these models is: (a) a stress rule
relating stress to strain and internal state , (b) a yield condition
that delineates elastic () from plastic () behavior, and (c) a flow rule
that governs the evolution of internal state when . Together the components of (a-c) elastoplasticity constitute a complex differential-algebraic equation (DAE) system [25, 26]. More details can be found in texts [27, 28] and the papers [29] these calibrations are based on. All the uncertain parameters are associated with the Voce-like [30, 31, 32, 33] hardening rule:
that controls the radius of the yield surface , and is dependent on the equivalent plastic strain (an evolving internal state variable) and its rate. Setting removes the rate dependence and reduces the model to the standard Voce form. Also, the relative sensitivity to is typically high due to it appearing inside the exponential function. We assume the elastic parameters are precisely determined by the baseline calibration and do not vary across the test specimens.
For these demonstrations the model sensitivities around the MLE fit are generated with finite differences. The parameter sensitivities are gradients which require function evaluations, where the exact coefficient depends on the scheme, likewise a full Hessian evaluation. We use the noise estimate to set the finite difference step size .
Synthetic validation
To test how well the method does with typical engineering nonlinearities and measurement noise, we calibrate an elastic plastic model and treat the parameters of Eq. (45) as uncertain. We use a tension experiment for calibration and a more involved test geometry, the -specimen shown in Figure 6, for validation. We add white noise to the calibration data but otherwise the set-up is self-consistent in the sense that the same material model is used in calibration and generating the data.
Figure 6. Aluminum × specimen: (left) undeformed, (right) deformed and colored by stress along the loading direction
We omit the rate dependence to minimize the amount of calibration data that we need to generate to fully calibrate the model. To generate the tension calibration data and the corresponding validation data we use a multivariate Gaussian parameter distribution:
and sample sizes for both datasets.5 To construct the calibration data, we added noise with a standard deviation 0.5 MPa to the data generated by the truth model.
Figure 7 shows that residual covariance is of rank and the singular values fall off steeply after that, which is consistent with our assumptions.
Figure 7. Synthetic validation residual spectrum, where the red line indicates the number of parameters
Figure 8 shows that the true and estimated parameter distributions resemble each other but are visibly different. Note the true and estimated means, particularly for , do not coincide due to imperfect least squares calibration. We conjecture this contributes significantly to the discrepancies in the parameter distributions. However, the push-forwards for both the calibration and training data cover the data well. In fact, distance shows the similarity of the predicted and true push-forward distributions is nearly constant up until the model is pushed into extrapolation.

Figure 8. Synthetic validation: (a) parameter covariance, (b) push-forward predictions.
Aluminum
In this example, we calibrate a Hill-48 orthotropic yield surface [34] with Voce hardening [30] to the aluminum data for a rolled plate. One complication is that the rolling process imparts a texture to the material microstructure that results in orthotropic plasticity behavior; the experiments were designed to characterize this orthotropic behavior. ASTM E8 uniaxial tension testing [35] in three directions relative to the material rolling direction and a shear top-hat testing in six orientations relative to the material rolling direction [36, 37]. Figure 9 shows the stress and deformation of the top-hat specimen.

Figure 9. Aluminum “top hat” specimen: (left) undeformed, (right) deformed and colored by shear stress
Figure 10 shows that residual covariance is of rank and the remaining singular values are consistent with the measurement noise level for these experiments.
Figure 10. Aluminum: residual spectrum
Figure 11a shows that there is some correlation between and . This anticorrelation is plausible given the hardening function Eq. (45) and the fact that this function largely resembles and determines the shape of the tensile response curves. The data and predicted trajectories are shown in Figure 11b. The predicted trajectories largely cover the data, with the exception of some outliers. The distances also indicate that the data and push-forward distributions are generally close with some exceptions near the beginning and end of the batch 1 tension data in particular.
Figure 11. Aluminum calibration: (a) parameter covariance, (b) push-forward predictions.
Stainless steel
In this example, we calibrate the rate-dependent plasticity model with a von Mises yield surface with hardening (45) to ductile failure data for 304L bar material [29], including , , , and . The experimental data set used for this calibration was designed to characterize the rate-dependence of viscoplastic behavior. A tensile bar was used for this dataset. Figure 12 shows the undeformed shape of the standard ASTM E8 [35] tension specimen that is held in grips at both ends and loaded in tension. The necking/localized narrowing inducing geometric softening and the resulting axial stress field are also shown in Figure 12. The calibration is complicated by the variance in the failure of the specimens as indicated by black dots at the end of stress-strain trajectories.

Figure 12. Steel tension specimen: (upper) undeformed, (lower) deformed and colored by axial stress
We varied the number of parameters that were treated as uncertain. Figure 13a shows a mixture of well-characterized and correlated parameters, and considerable redistribution as the size of the parameter space is changed. Nevertheless, the push-forward trajectories, shown in Figure 13b, clearly cover the experimental data, which implies that the systematic variations induced by the parameters are somewhat redundant in representing the tension data. The distances corroborate this observation.
Figure 13. Steel calibration: (a) parameter covariance, (b) push-forward predictions. Note contours on a log-scale
Conclusion
We proposed an extension to the classic Laplace approximation that handles both uncorrelated noise and correlated discrepancies in the mode of the embedded error formulation [6, 19]. This is particularly useful in calibrating material models where the material properties vary due to manufacturing inconsistencies and the like. Furthermore, we provided an estimate of the posterior distribution using straightforward linear algebra that is often as accurate as optimization. The method is both efficient, requiring relatively little computation after finding the MLE, and interpretable, providing a parameter covariance matrix and a noise estimate.
We plan to rectify some of the weaknesses of the method in future work. First, the reliance on repeats to assess variance is costly in terms of experiments. In fact, the repeated experiments currently allow us to separate correlated variance versus the uncorrelated noise. Second, it seems that the method is fairly robust to moderate model discrepancy, but this must have limits. We expect initializing the mean of the distribution at the MLE and then allowing it to adjust through optimization with the covariance should ameliorate some of the potential issues. Lastly, the reliance on a multivariate normal surrogate posterior leads to violation of physical constraints, such as positivity, which is a common problem; we will explore other remedies than simple truncation and reassignment of probability mass.
Acknowledgements
We thank Cosmin Safta (Sandia) for helpful discussions on the implementation of the BBVI method.
Sandia National Laboratories is a multi-mission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC (NTESS), a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration (DOE/NNSA) under contract DE-NA0003525. This written work is authored by an employee of NTESS. The employee, not NTESS, owns the right, title and interest in and to the written work and is responsible for its contents. Any subjective views or opinions that might be expressed in the written work do not necessarily represent the views of the U.S. Government. The publisher acknowledges that the U.S. Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce the published form of this written work or allow others to do so, for U.S. Government purposes. The DOE will provide public access to results of federally sponsored research in accordance with the DOE Public Access Plan.
Code and data availability
The code and examples related to these developments are available at the MATCAL GitHub site, https://github.com/sandialabs/matcal.
References
- [1]Roger Ghanem, David Higdon, Houman Owhadi, et al. Handbook of uncertainty quantification, volume 6. Springer New York, 2017.
- [2]David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational inference: A review for statisticians. Journal of the American statistical Association, 112(518):859–877, 2017.
- [3]Léon Bottou. Stochastic gradient descent tricks. In Neural networks: tricks of the trade: second edition, pages 421–436. Springer, 2012.DOI
- [4]Luke Tierney and Joseph B Kadane. Accurate approximations for posterior moments and marginal densities. Journal of the american statistical association, 81(393):82–86, 1986.DOI
- [5]David JC MacKay. Bayesian interpolation. Neural computation, 4(3):415–447, 1992.
- [6]K Sargsyan, HN Najm, and R Ghanem. On the statistical calibration of physical models. International Journal of Chemical Kinetics, 47(4):246–276, 2015.DOI
- [7]Vicente Romero, Amalia Black, George Orient, and Bonnie Antoun. Propagating stress-strain curve variability in multi-material problems: Temperature-dependent material tests to plasticity models to structural failure predictions. In Engineering Failure Analysis. IntechOpen, 2019.DOI
- [8]Marc C Kennedy and Anthony O’Hagan. Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(3):425–464, 2001.
- [9]Karthik Duraisamy and Praveen Chandrashekar. Goal-oriented uncertainty propagation using stochastic adjoints. Computers & fluids, 66:10–20, 2012.
- [10]Siddhant Wahal and George Biros. Bimc: the bayesian inverse monte carlo method for goal-oriented uncertainty quantification. part i. arXiv preprint arXiv:1911.00619, 2019.arxiv.org/abs/1911.00619
- [11]Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2(11):2, 2011.DOI
- [12]Gareth O Roberts and Richard L Tweedie. Exponential convergence of langevin distributions and their discrete approximations. Bernoulli, 2(4):341–363, 1996.DOI
- [13]Gareth O Roberts and Jeffrey S Rosenthal. Optimal scaling of discrete approximations to langevin diffusions. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 60(1):255–268, 1998.DOI
- [14]Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011.
- [15]Bo Wang and D Michael Titterington. Inadequacy of interval estimates corresponding to variational bayesian approximations. In International workshop on artificial intelligence and statistics, pages 373–380. PMLR, 2005.
- [16]Qiang Liu and Diliin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems, 29, 2016.arxiv.org/abs/1608.04471
- [17]Chang Liu, Jingwei Zhuo, Pengyu Cheng, Ruiyi Zhang, and Jun Zhu. Understanding and accelerating particle-based variational inference. In International Conference on Machine Learning, pages 4082–4092. PMLR, 2019.arxiv.org/abs/1807.01750
- [18]Richard Bellman. Dynamic programming, volume 39. Princeton university press, 1957.
- [19]Khachik Sargsyan, Xun Huan, and Habib N Najm. Embedded model error representation for bayesian model calibration. International Journal for Uncertainty Quantification, 9(4), 2019.DOI
- [20]Ronald A Fisher. Frequency distribution of the values of the correlation coefficient in samples from an indefinitely large population. Biometrika, 10(4):507–521, 1915.DOI
- [21]Ronald A Fisher. On the” probable error” of a coefficient of correlation deduced from a small sample. Metron, 1:3–32, 1921.
- [22]Ilya Archakov and Peter Reinhard Hansen. A new parametrization of correlation matrices. Econometrica, 89(4):1699–1715, 2021.arxiv.org/abs/2012.02395
- [23]Rajesh Ranganath, Sean Gerrish, and David Blei. Black box variational inference. In Artificial intelligence and statistics, pages 814–822. PMLR, 2014.
- [24]J Nocedal and SJ Wright. Numerical optimization. Springer Series in Operations Research and Financial Engineering, 2006.
- [25]CW Gear. Differential-algebraic equations. In computer aided analysis and optimization of mechanical system dynamics, pages 323–334. Springer, 1984.DOI
- [26]Uri M Ascher and Linda R Petzold. Computer methods for ordinary differential equations and differential-algebraic equations. SIAM, 1998.DOI
- [27]Jacob Lubliner. Plasticity theory. Courier Corporation, 2008.
- [28]Morton E Gurtin, Eliot Fried, and Lallit Anand. The mechanics and thermodynamics of continua. Cambridge university press, 2010.DOI
- [29]Kyle N Karlson, Alyssa J Skulborstad, Jonathan D Madison, Andrew T Polonsky, Helena Jin, Amanda Jones, Brett Sanborn, Charlotte LB Kramer, Bonnie R Antoun, and Wei-Yang Lu. Toward accurate prediction of partial-penetration laser weld performance informed by three-dimensional characterization–part ii: μct based finite element simulations. Tomography of Materials and Structures, 2:100007, 2023.DOI
- [30]Eric Voce. The relationship between stress and strain for homogeneous deformation. Journal of the Institute of Metals, 74:537–562, 1948.
- [31]E Voce. A practical strain hardening function. Metallurgia, 51:219–226, 1955.
- [32]H-W Swift. Plastic instability under plane stress. Journal of the Mechanics and Physics of Solids, 1(1):1–18, 1952.DOI
- [33]Gordon R Johnson. A constitutive model and data for metals subjected to large strains, high strain rates and high temperatures. In Proceedings of the 7th International Symposium on Ballistics, The Hague, Netherlands, 1983, 1983.
- [34]Rodney Hill. A theory of the yielding and plastic flow of anisotropic metals. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 193(1033):281–297, 1948.DOI
- [35]Astm e8/e8m-22: Standard test methods for tension testing of metallic materials. In ASTM standards. American Society for Testing and Materials International, 2016.
- [36]Edmundo Corona, Lisa Anne Deibler, Benjamin Reedlunn, Mathew Duffy Ingraham, and Shelley Williams. An experimental study of shear-dominated failure in the 2013 sandia fracture challenge specimen. Technical report, Sandia National Lab. (SNL-NM), Albuquerque, NM (United States), 2015.DOI
- [37]Edmundo Corona, Matthew Spletzer, Brian T Lester, and Carter J Fietek. Validation of material models for puncture of 7075-t651 aluminum plate. International Journal of Solids and Structures, 257:111893, 2022.DOI









