betaFeedback

WorksMathematics

Optimally Emulating Smooth Functions for Bayesian Analyses

Abstract

For Bayesian analyses aimed at constraining the parameter space of computationally expensive models by comparison with heterogeneous data sets, model emulators play a pivotal role. Model emulators enable searches through high-dimensional parameter spaces, searches which might otherwise require millions of full model runs. Here, an emulation strategy is presented based on the assumption that the model output can be well described by a polynomial expansion with all terms being independent of one another and with higher order terms being given decreasing weights according to a convergence parameter. It is found that if the expansion's form is constrained to being invariant to rotation and translation of the parameter space that the polynomial form becomes equivalent to that of a Gaussian process emulator using a correlation kernel with a Gaussian form. In the limit that the function is smooth, quantified by a convergence parameter, ΛΛ, it is found that rather few training points are required to reproduce the full model and provide a good estimate of its uncertainty. An accuracy metric is defined based on an average of the expected uncertainty over the prior before training data are actually gathered. Analytic expressions for the expected accuracy are provided based solely on an expected value of ΛΛ and the location of the training points. This enables a tenable optimized placement of training points, which is shown to be significantly superior to random placement or that from Latin hyper-cube sampling. Examples of optimized placement are provided, and the accuracy is studied as a function of various variables, such as the number of model parameters, the number of training points, and the convergence parameter.

Introduction

In the quest to provide rigorous quantitative constraints of high-dimension model-parameter spaces by comparing sophisticated computationally expensive numerical models to experimental results, one must sample the model-parameter space by running the full-model at large numbers of points. For example, sampling a 12-parameter model with only three points in each direction requires 312≈500,0003^{12} \approx500{,}000 runs. Thus, Markov Chain Monte Carlo (MCMC) searches, weighted by the likelihood that the model reproduces the experiment, often involve millions of full-model evaluations. If a single run of the full-model at one point in parameter space requires days of CPU time, the analysis becomes untenable. Model emulation provides a solution to this problem. If one can train an emulator to reproduce the full-model by training it on only a few hundred full-model runs or fewer, and if the emulator can provide a reasonable estimate of its uncertainty, the emulator can take the place of the full model in the MCMC search, and thus enable the analysis.

This paper is motivated by the theory and practice of heavy-ion physics, a field where emulators have been applied to various large-scale analyses, often involving more than a dozen model parameters, with models that often require days of CPU time. Contextual references in heavy-ion physics abound: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43]. For example, in a global analysis of 14 model parameters [34], Gaussian-processes (GP) [44, 45] were used as emulators of the output from hybrid hydrodynamic/Boltzmann approaches with accuracies near one percent, after being trained on 1200 points. One might ask whether one can do as well, or better, with fewer training points. The field of heavy-ion physics is similar to many other physical or biological fields of application, in that the relation between the model parameters and model output tends to be roughly linear in a physical parameter range. For example, in high-energy Lead-Lead (Pb-Pb) collisions at the Large Hadron Collider (LHC), the mean momentum of outgoing protons rises monotonically with the stiffness of the equation of state. That dependence should not be perfectly linear, but one would expect that, within a reasonable prior parameter range, the dependence is well described by linear or quadratic functions, and that higher-order terms should become increasingly negligible. This expectation also applies to most any other observable with respect to nearly any other model parameter. For lower-energy collisions, where phase transitions are perhaps more relevant, one then imagines a dependence that has significant structure, e.g. a clear threshold w.r.t. some parameter or set of parameter hypersurfaces, with roughly linear behaviors in the regions determined by these surfaces. At the higher LHC energies, these complications would not occur, and the methods in this paper would apply directly to the entire physical parameter range.

As a quick sanity check, putting aside complete UQ considerations, imagining a scalar model, with NparsN_{\mathrm{pars}} parameters, if it were nearly perfectly linear, a first-order polynomial (linear approximation) would provide an excellent fit, and would presumably result in acceptable UQ bands. One could then run the full model only Npars+1N_{\mathrm{pars}} + 1 times, at points fairly regularly far apart, to constrain the linear approximation’s NparsN_{\mathrm{pars}} slopes and the intercept. For a 12-parameter model, this would require only 13 full model runs. Imagining that the model was instead nearly perfectly quadratic, one would need only Npars(Npars+1)/2N_{\mathrm{pars}}(N_{\mathrm{pars}} + 1)/2 additional full-model runs to constrain all the quadratic contributions. For the 12-parameter case, this would lead to 91 full model runs, with an expectation that this way of contraining the model would still lead to acceptable UQ bands. This is fewer than the number of full-model runs often used in recent analyses in heavy-ion physics, by orders of magnitude, for modeling scenarios where there is arguably little benefit, within physical parameter ranges, of a cubic approximation over a quadratic one.

Another way to say this, which is prudent and parsimonious, both in computational and modeling terms, is that in any domain of application, such as heavy-ion physics, where most models feature rather smooth responses to their parameters, is that one should reconsider polynomial fits, and investigate the degree to which they might, or might not, improve upon the GP approaches cited above. The first goal of this paper is to exploit the expectation that a polynomial fit should be adequate systematically, by presenting a way to construct a polynomial expansion where the importance of various terms in the expansion decreases with increasing order. In the physics literature, this expectation of adequacy (of a polynomial approximation) is often labeled as “convergence” of the approximation, where the asymptotic notion in this convergence is with respect to the degree of the polynomial as it increases to infinity, even though practitioners talk about convergence even if the degree is small, as long as the error in the approximation fulfills the practitioners’ “expectations”, which in this case should be understood as their stated or implicit tolerances. To be able to declare that approximations are “converging” in the sense that the errors are tolerable, it is best to quantify those expectations in the most transparent terms possible, which is often done in terms of a “convergence” parameter, Λ\Lambda. Once one has agreed on such a quantitative framework, one can work to define and find an optimum set of expansion coefficients, along with a quantification of their uncertainty. Herein, the emulator is built on a polynomial expansion where the polynomial’s variables are the NparsN_{\mathrm{pars}} values θi/Λ\theta_i/\Lambda, where the model parameters are θi\theta_i; i=1,…,Nparsi = 1,\ldots,N_{\mathrm{pars}}, and are all scaled so that they vary of order unity within the prior. Thus, higher-order terms contribute as negative integer powers of the convergence parameter Λ\Lambda.

Emulators defined by choosing polynomial fits based on least-square minimization algorithms have long been viewed as good classes of candidates. More modern approaches improve on such fits by providing estimates that exactly reproduce the training values when evaluated at the training points, and provide an estimate of the emulator’s uncertainty as a function of the vector of parameters θ\theta. These approaches which reproduce training values are typically based on GP assumptions, where the predictions are all based on an assumed form of the covariance between values YY and Y′Y' of the model evaluated at two parameter-vector points θ\theta and θ′\theta'. That correlation kernel might be a function of θ\theta and θ′\theta', but a particularly simple and popular form is a “Gaussian” covariance kernel, Cov⁡0(Y1,Y2∣σA,Λ)=σA2e−∣θ−θ′∣2/(2Λ2)\operatorname{Cov}_0(Y_1,Y_2\mid\sigma_A,\Lambda) = \sigma_A^2 e^{-\lvert\theta-\theta'\rvert^2/(2\Lambda^2)}. Here, the parameters σA\sigma_A and Λ\Lambda are referred to as hyper-parameters, to differentiate them from model parameters. They are used only to tune the emulator, and their tuning can be optimized based on training data. Here, the subscript 0 denotes the fact that the covariance is taken with respect to the prior distribution. It has been shown that any polynomial expansion, or any expansion over some set of basis functions, where the expansion’s terms have random coefficients, can be optimized to reproduce training points. These can then be equated to GP forms once one defines the appropriate correlation kernel [44]. Previous approaches have typically been constructed from sets of orthogonal basis functions, e.g. [46], but this requirement is not necessary and is not enforced herein. The sum over polynomial terms with random coefficients is sometimes referred to as a “polynomial chaos” [47]. Once trained, one can optimize all the random coefficients for the polynomial expansion. One can then calculate the emulator values at any parameter vector θ\theta using those optimal coefficients. However, there are many more coefficients than training constraints, since one has the option to truncate the full Taylor series at an arbitrarily high degree. Fortunately, once one has written an equivalent form for the correlations, the emulator can much more efficiently be expressed in terms of the correlation kernel and one can avoid performing sums over all monomial terms. This is sometimes referred to as the kernel trick [44].

Narrative summary of this paper’s contributions

In this work’s first part, we consider emulators as polynomial expansions with random coefficients, and we constrain the emulators’ form by requiring that the emulating behavior be invariant by rotation and by translation. Under this requirement, we find that the equivalent correlation kernel must then have a simple Gaussian form. Thus, if one believes that the model is equally sensitive to all parameters and that there is no reason to expand the polynomials about one point rather than around some other point, then the equivalent GP correlation kernel must be Gaussian, and described via the form involving the two hyper-parameters σA\sigma_A and Λ\Lambda mentioned above. The polynomial description then becomes equivalent to a GP description with a correlation kernel with a simple Gaussian form. If the user believes that the sensitivity w.r.t. some model parameter should be less than that of another parameter, rather than applying different convergence parameters, one can scale the parameters so that the expected sensitivity is isotropic in the scaled parameter basis. Thus, the assumption of the model having a rotation-invariant behavior can reasonably apply to many problems after appropriate rescaling. The assumption of translation invariance is one of spatial stationarity across parameter space, which is also a strong modeling preference for most problems. In particular, it significantly reduces the statistically grounded UQ on any parameter, avoiding the risk of overfitting a model for UQ itself.

On the strength of the translation and rotation invariance assumption of the model’s parameter-space sensitivity, the single scalar convergence parameter, Λ\Lambda, sets the relative mutual importance of terms of different orders in the polynomial expansion. We find that Λ\Lambda is precisely the same as the correlation length in the GP interpretation with a Gaussian correlation kernel. For the polynomial expansion to “converge”, i.e. for its truncation to yield a tolerable error, within a few orders, at a minimum, one must have the convergence parameter Λ\Lambda larger than the characteristic scale of the prior on parameter space, and this would typically not be enough for the truncation error to be tolerable. If the model parameters are scaled to vary roughly between −1-1 and +1+1, then Λ\Lambda should be significantly larger than unity if the expansion is to be well defined by a few terms, or equivalently, to be constrained via calibration by a few training points. After translating to the equivalent GP paradigm, this is equivalently stated as having the GP correlation length, which coincides with Λ\Lambda, being significantly larger than the size of the prior.

Conversely, if one assumes a Gaussian correlation kernel in a GP context, the relations derived here identify the form of the corresponding polynomial expansion. Thus, one can use this form to generate random functions that are consistent with the Gaussian kernel, albeit with some cutoff in the order of the expansion’s truncation. This should be useful for testing emulation software, as it should allow one to build models that correspond to the desired correlation kernel and should, by construction, have uncertainties that are well understood and quantified.

Our work’s second thrust is to understand, for expansions which are rapidly converging (where only a small number of terms are needed to get tolerable truncation errors), how one should optimally choose the points at which the full model is run to train the emulator. We present a metric that represents the expected accuracy (expected truncation error) of the emulator as a function of the convergence parameter, Λ\Lambda, and as a function of the positions of the training points. For a fixed Λ\Lambda, optimized arrangements of training points are then generated for various cases involving either Gaussian or uniform priors on the parameters. For test cases, this procedure is found to be significantly superior to either random placement of the training points throughout the prior, or to Latin hyper-cube sampling [48]. In this setting of rapidly converging expansions, one might expect that the number of training points might be chosen to match that needed for linear, quadratic or cubic fits. For a 6-parameter example, that would require 7, 28 or 84 full-model runs respectively. Here, we show that the marginal improvement to the emulation decreases at each of the two corresponding thresholds, assuming that the training points are optimally chosen. In practice, this analysis would provide a user who has a clear understanding of their computational costs, and who has a clear quantification of their aversion to error risk, which can simply take the form of a worst-case scenario cost of tolerance, to run an elementary cost-benefit analysis.

By accomplishing the goals above, one should be able to gain a better understanding of how many model runs might be necessary to design an emulator to the desired accuracy, where the term “desired” can be built into a cost-benefit analysis of the aforementioned ilk. This decision framework will still depend on some pre-estimate or other choice of the convergence parameter Λ\Lambda, but it turns out, from our testing, that the actual placement of the training points appears rather insensitive to that estimate for Λ\Lambda, even though the accuracy is strongly sensitive to the estimate. For users who knows ahead of time that some of the model parameters are more likely to drive the functional behavior than others, we provide a seamless method to exploit that prior knowledge. This can assist with overcoming the challenges of high dimensionality. We find herein that such flexibility turns out to be critical in determining how confidently one might ultimately emulate models.

In the next section, we present the mathematical form of such an emulator, and a derivation of the optimum set of coefficients, given the training data. Closed expressions for the emulator uncertainty are also derived. The emulator’s form also accounts for models with noisy training, i.e. those for which point-by-point uncertainty is acknowledged and taken into account, i.e. where that uncertainty is propagated in a fashion consistent with the modeling. The physical arguments for choosing the specific emulator form are discussed. The equivalence with a GP emulator is established carefully. The subsequent section presents a tractable method to optimize the location of training points in the NparsN_{\mathrm{pars}}-dimensional parameter space. For the case of Gaussian priors with Npars+1N_{\mathrm{pars}}+1 training points, the placement follows a straightforward NparsN_{\mathrm{pars}}-dimensional simplex, but for most cases the optimized solution does not have a simple symmetry. Instead, our method finds the optimal points numerically. In Section IV we test our emulator procedure by applying it to some models with random polynomial expansions or randomized trigonometric or exponential forms. This allows us to gain a solid quantified grasp for how well one might, in practice, not only emulate a function, but also understand how well one might trust the fidelity of our emulator’s quantitatively estimated uncertainty. This faithfulness is studied as a function of the dimensionality of the model’s parameter space and of the convergence parameter Λ\Lambda.

Theoretical Foundation of Emulating Convergent Functions

The form and basic properties of our proposed class of emulators is described in detail in this section. The form is that of a polynomial expansion in terms of the model parameters, i.e. expanded as a multivariate power series with respect to the parameter vector, where the coefficient for each term in the polynomial is given an independent centered Gaussian (mean-zero normal) weight. Since centered Gaussian variables have a scale property with respect to their standard deviation, it is convenient and elementary to standardize the Gaussian coefficients by pulling out their scales as explicit deterministic emulation weights. Using these weights, our form, while general, is easily constrained to assume that terms of higher order should contribute less. For example, in the absence of training data, quadratic terms should be expected to contribute less than linear terms, and cubic terms should contribute less than quadratic terms. What separates the form presented here compared to previous studies is that we also constrain the expansion to behave invariantly by rotation and by translation in parameter space. Thus, moving the origin about which the polynomial expansion is constructed, or rotating the coordinate system about any such origin, should not affect the emulator.

In fact, in the next subsection we show that this natural constraint implies a precise and unique form for the deterministic emulator weights, well illustrating the intuition that weights should decrease as terms’ order increases. We shall see that the form of the resulting polynomial expansion depends only on two emulator parameters, σA\sigma_A, which sets the strength of the fluctuation of the pre-trained emulator at any point, and Λ\Lambda, which drives the convergence of the polynomial expansion, as it is a power series expansion in terms of powers of the scaled parameter vector Λ−1θ⃗\Lambda^{-1}\vec{\theta}. A reader might pause at this point to recognize that we will be dealing with NparsN_{\mathrm{pars}} physical model parameters assembled in the vector θ⃗\vec{\theta} and with two scalar emulator parameters σA\sigma_A and Λ\Lambda.

It is well known that expansions, including polynomial expansions, which are linear in their random coefficients, lead to emulators whose behavior can easily be expressed in terms of the pre-trained covariance kernel, C0(θ⃗1,θ⃗2)C_0(\vec{\theta}_1,\vec{\theta}_2). Both the emulator and its error are easily expressed in terms of C0C_0, which allows one to avoid calculations where one must sum over the actual terms of the polynomial expansion. This result is known as the kernel trick [44]. As we said, this is not new. The new constructive element we introduce here, is that our assumption that all the random terms in the polynomial expansion are independent of each other, and our constraint of rotational and translational invariance, together imply a unique form for the expansion, and thus a unique form for the kernel. That unique form turns out to be that of an ordinary Gaussian kernel (i.e. a simple quadratic exponential decay) with Λ\Lambda playing the role of characteristic length, in parameter space, of that quadratic exponential correlation-decay. By standardizing all parameters to be of the same magnitude, one then only requires a single characteristic length, i.e. just one scalar Λ\Lambda.

Throughout this section the model parameters, θ⃗\vec{\theta}, are assumed to have been translated, rotated and scaled so that their expected sensitivities to a given step size ΔΘ\Delta\Theta are similar. For example, if a model parameter xix_i has an initial prior, ai<xi<bia_i < x_i < b_i, one defines a new translated and scaled variable

θi=2θmax⁡,i(bi−ai)(xi−(ai+bi)/2).(1)\theta_i = \frac{2\theta_{\max,i}}{(b_i-a_i)}\left(x_i-(a_i+b_i)/2\right). \tag*{(1)}

where now the prior on θi\theta_i is uniform in the symmetric range ±θmax⁡,i\pm\theta_{\max,i}. The flexibility with choosing the scale θmax⁡,i\theta_{\max,i} is use so that, for each ii, one would not expect the observable, y(θ⃗)y(\vec{\theta}), to be any more sensitive to steps of size ΔΘ\Delta\Theta in the direction of θi\theta_i than to steps of the same size in the direction of some other parameter θj\theta_j. Thus, if the sensitivity to θ1\theta_1 was expected to be higher than the sensitivity to θ2\theta_2, one would reduce the ratio of θmax⁡,2\theta_{\max,2} to θmax⁡,1\theta_{\max,1} until the expected sensitivities were similar. All this can be attempted mathematically when partial derivatives of the model for y(θ⃗)y(\vec{\theta}) can be estimated with some confidence, or it can be done empirically thanks to a limited number of full-model numerical evaluation, not emulation. Further, on the question of fixing issues with rotational dependence, to consider a simple example, if one expected y(θ⃗)y(\vec{\theta}) to be more sensitive to a given direction in parameter space, e.g., θ1+θ2\theta_1+\theta_2, compared to some orthogonal direction, θ1−θ2\theta_1-\theta_2, one could rotate the coordinate system by 45 degrees, then scale the two dimensions accordingly. More generally, one can define a matrix, SijS_{ij}, describing the expected or empirical variability of yy with respect to ΔθiΔθj\Delta\theta_i\Delta\theta_j, then redefine a coordinate system where the new parameters were eigenvector of SS, then scale the new parameters so that the eigenvalues of SS were similar. One would then perform the model emulation in this new basis where the functions was not expected to vary more in any specific direction for the same step size, although the range of the prior could vary per directions. For the case where one had no information indicating that the model were more likely to vary across the prior in one direction than in another, all the values of θmax⁡,i\theta_{\max,i} would all be set to equal 1.

As we mentioned, when using a uniform prior on each parameter, we shall assume that the parameter θi\theta_i is distributed according to the following probability density:

Pi(θi)={12θmax⁡,i,−θmax⁡,i<θi<θmax⁡,i,0,otherwise(2)P_i(\theta_i) = \begin{cases} \frac{1}{2\theta_{\max,i}}, & -\theta_{\max,i}<\theta_i<\theta_{\max,i},\\ 0, & \text{otherwise} \end{cases} \tag*{(2)}

and as we just said, if all parameters are considered to be of equal importance, we let θmax⁡,i=1\theta_{\max,i}=1 for every ii. If the model function y(θ⃗)y(\vec{\theta}) is expected to vary mostly along a specific iith parameter’s direction, since everthing is invariant to a single scaling factor, we need only scale that parameter with θmax⁡,i=1\theta_{\max,i}=1, and scale other parameters with appropriately chosen smaller values of θmax⁡,j\theta_{\max,j}, j≠ij\ne i.

If the parameters have Gaussian priors, one can scale and translate each of them so that the new marginal priors are Gaussian of the form

Pi(θi)=1(2π/βi)1/2e−βiθi2/2.(3)P_i(\theta_i)=\frac{1}{(2\pi/\beta_i)^{1/2}}e^{-\beta_i\theta_i^2/2}. \tag*{(3)}

where the βi\beta_i’s are chosen to place the parameter sensitivities on presumed equal footing. But the Gaussian case is easier to treat formally than the uniform case, because a multivariate Gaussian distribution has an intrinsic property of rotational invariance that the multivariate uniform distribution lacks. Therefore, if one insists on a specific multivariate Gaussian prior distribution on the entire vector θ⃗\vec{\theta}, with mean μ⃗\vec{\mu} and covariance matrix SS, one then simply shifts the entire vector θ⃗\vec{\theta} by subtracting μ\mu, and one rotates and scales it simultaneously by multiplying it by a Cholesky square root LL of SS, i.e. a lower-triangular matrix LL such that S=LLTS=LL^{\mathsf{T}}. The new resulting multivariate prior is then the product of all the normal densities in (3) with each βi=1\beta_i=1. These βi\beta_i’s can then also be adjusted to make up for uneven sensitivities, similarly to what we suggested with the θmax⁡,i\theta_{\max,i} scales for uniform priors.

Again in the multivariate normal case, if we now believe that all the parameters are put on equal footing, but we want this multivariate prior to have the same dispersion properties as in the uniform case, one might want to match variances. The variance of the uniform distribution on [−1,1][-1,1] being equal to 1/31/3, the value of all the βi\beta_i’s need only be set to equal 3 as well for the variances to match in the Gaussian and uniform priors cases. As we mentioned in the uniform prior case, if some given parameter (# ii) is expected to drive most of the variability that parameter is assigned βi=3\beta_i=3 and all other parameters are assigned lower values for their βj\beta_j’s. Certainly, matching variances is not the only way to match dispersions. If instead one wanted a density in (3) with a βi\beta_i which matches the support of the uniform distribution [−1,1][-1,1] at a confidence/credibility level of 95%, then one will want to choose not βi=3\beta_i=3 but rather closer to βi=4\beta_i=4 (strictly speaking, it would be closer to βi=3.84\beta_i=3.84.) This is because the normal distribution has higher moments that increase slightly faster than those of the uniform distribution. In the sequel, we will stick with the variance-matching choice of βi=3\beta_i=3 for our expository purposes.

References

  1. [1]T. S. Domingues, R. Krupczak, J. Noronha, T. N. da Silva, J. F. Paquet and M. Luzum, Phys. Rev. C 110, no.6, 064904 (2024).
  2. [2]R. Ehlers et al. [JETSCAPE], [arXiv:2408.08247 [hep-ph]].arxiv.org/abs/2408.08247
  3. [3]R. Ehlers et al. [JETSCAPE], EPJ Web Conf. 296, 15009 (2024).
  4. [4]J. F. Paquet, J. Phys. G 51, no.10, 103001 (2024).
  5. [5]W. Fan et al. [JETSCAPE], Phys. Rev. C 109, no.6, 064903 (2024).
  6. [6]M. R. Heffernan, C. Gale, S. Jeon and J. F. Paquet, Phys. Rev. Lett. 132, no.25, 252301 (2024).
  7. [7]M. R. Heffernan, C. Gale, S. Jeon and J. F. Paquet, Phys. Rev. C 109, no.6, 065207 (2024).
  8. [8]B. Weiss, J. F. Paquet and S. A. Bass, J. Phys. G 50, no.6, 065104 (2023).
  9. [9]Y. Ji, H. S. Yuchi, D. Soeder, J. F. Paquet, S. A. Bass, V. R. Joseph, C. F. J. Wu and S. Mak, [arXiv:2209.13748 [stat.ME]].arxiv.org/abs/2209.13748
  10. [10]M. Heffernan, C. Gale, S. Jeon and J. F. Paquet, Acta Phys. Polon. Supp. 16, no.1, 1-A151 (2023).
  11. [11]D. Liyanage, Y. Ji, D. Everett, M. Heffernan, U. Heinz, S. Mak and J. F. Paquet, Phys. Rev. C 105, no.3, 034910 (2022).
  12. [12]J. Mulligan et al. [JETSCAPE], [arXiv:2106.11348 [nucl-th]].arxiv.org/abs/2106.11348
  13. [13]S. Cao et al. [JETSCAPE], Phys. Rev. C 104, no.2, 024905 (2021).
  14. [14]J. E. Bernhard, J. S. Moreland and S. A. Bass, Nature Phys. 15, no.11, 1113-1117 (2019).
  15. [15]J. S. Moreland, J. E. Bernhard and S. A. Bass, Phys. Rev. C 101, no.2, 024911 (2020).
  16. [16]J. S. Moreland, J. E. Bernhard and S. A. Bass, Nucl. Phys. A 982, 503-506 (2019).
  17. [17]Y. Xu, J. E. Bernhard, S. A. Bass, M. Nahrgang and S. Cao, Phys. Rev. C 97, no.1, 014907 (2018).
  18. [18]W. Ke, J. S. Moreland, J. E. Bernhard and S. A. Bass, Nucl. Part. Phys. Proc. 289-290, 483-486 (2017).
  19. [19]Y. Xu, S. Cao, M. Nahrgang, J. E. Bernhard and S. A. Bass, Nucl. Part. Phys. Proc. 289-290, 257-260 (2017).
  20. [20]J. Auvinen, J. E. Bernhard, S. A. Bass and I. Karpenko, Phys. Rev. C 97, no.4, 044905 (2018).
  21. [21]Y. Xu, M. Nahrgang, J. E. Bernhard, S. Cao and S. A. Bass, Nucl. Phys. A 967, 668-671 (2017).
  22. [22]J. Auvinen, I. Karpenko, J. E. Bernhard and S. A. Bass, Nucl. Phys. A 967, 784-787 (2017).
  23. [23]J. E. Bernhard, J. S. Moreland and S. A. Bass, Nucl. Phys. A 967, 293-296 (2017).
  24. [24]W. Ke, J. S. Moreland, J. E. Bernhard and S. A. Bass, Phys. Rev. C 96, no.4, 044912 (2017).
  25. [25]J. Auvinen, J. E. Bernhard and S. A. Bass, Acta Phys. Polon. Supp. 10, 455 (2017).
  26. [26]J. E. Bernhard, P. W. Marcy, C. E. Coleman-Smith, S. Huzurbazar, R. L. Wolpert and S. A. Bass, Phys. Rev. C 91, no.5, 054910 (2015).
  27. [27]S. A. Jahan, H. Roch and C. Shen, Phys. Rev. C 110, no.5, 054905 (2024).
  28. [28]A. Mankolli et al. [JETSCAPE], EPJ Web Conf. 296, 05010 (2024).
  29. [29]C. Shen, B. Schenke and W. Zhao, EPJ Web Conf. 296, 14001 (2024).
  30. [30]H. Mäntysaari, B. Schenke, C. Shen and W. Zhao, Acta Phys. Polon. Supp. 16, no.1, 1-A33 (2023).
  31. [31]H. Mäntysaari, B. Schenke, C. Shen and W. Zhao, Phys. Lett. B 833, 137348 (2022).
  32. [32]D. Everett et al. [JETSCAPE], Phys. Rev. C 103, no.5, 054904 (2021).
  33. [33]J. F. Paquet et al. [JETSCAPE], Nucl. Phys. A 1005, 121749 (2021).
  34. [34]S. Pratt, E. Sangaline, P. Sorensen and H. Wang, Phys. Rev. Lett. 114, 202301 (2015).
  35. [35]E. Sangaline and S. Pratt, Phys. Rev. C 93, no.2, 024908 (2016).
  36. [36]J. Novak, K. Novak, S. Pratt, J. Vredevoogd, C. Coleman-Smith and R. Wolpert, Phys. Rev. C 89, no.3, 034917 (2014).
  37. [37]B. A. Li and W. J. Xie, [arXiv:2501.02579 [nucl-th]].DOI
  38. [38]M. Alqahtani, R. S. Bhalerao, G. Giacalone, A. Kirchner and J. Y. Ollitrault, Phys. Rev. C 110, no.6, 064906 (2024).
  39. [39]J. M. Wang, X. G. Deng, W. J. Xie, B. A. Li and Y. G. Ma, [arXiv:2406.07051 [nucl-th]].DOI
  40. [40]B. A. Li and W. J. Xie, Nucl. Phys. A 1039, 122726 (2023).
  41. [41]A. V. Giannini et al. [ExTrEMe], Phys. Rev. C 107, no.4, 044907 (2023).
  42. [42]J. Jia and S. Mohapatra, Phys. Rev. C 88, no.1, 014907 (2013).
  43. [43]P. Morfouace et al., Phys. Lett. B 799, 135045 (2019).
  44. [44]C.E. Rasmussen and C.K.I. Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006, ISBN 026218253X.DOI
  45. [45]Williams, C. K. I., & Rasmussen, C. E. (1996). Gaussian Processes for Regression in Advances in Neural Information Processing Systems 8. In MIT Press.
  46. [46]T. Tang, S. Mak and D. Dunson, Hierarchical Shrinkage Gaussian Processes: Applications to Computer Code Emulation and Dynamical System Recovery SIAM/ASA J. on Uncertainty Quantification 12, no. 4, 10.1137 (2024).arxiv.org/abs/2302.00755
  47. [47]Xiu, D. (2010). Generalized Polynomial Chaos in Numerical Methods for Stochastic Computations. In Princeton University Press.
  48. [48]M.D. McKay, R.J. Beckman, R.J. and W.J. Conover. Technometrics. 21 (2). American Statistical Association: 239–245.

Paper details

vlnc:W_23may2d3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents