betaFeedback

WorksPhysics and Astronomy

Uncertainty-quantified phenomenological optical potentials for single-nucleon scattering

Abstract

Optical-model potentials (OMPs) continue to play a key role in nuclear reaction calculations. However, the uncertainty of phenomenological OMPs in widespread use – inherent to any parametric model trained on data – has not been fully characterized, and its impact on downstream users of OMPs remains unclear. Here we assign well-calibrated uncertainties for two representative global OMPs, those of Koning-Delaroche and Chapel Hill '89, using Markov-Chain Monte Carlo for parameter inference. By comparing the canonical versions of these OMPs against the experimental data originally used to constrain them, we show how a lack of outlier rejection and a systematic underestimation of experimental uncertainties contributes to bias of, and overconfidence in, best-fit parameter values. Our updated, uncertainty-quantified versions of these OMPs address these issues and yield complete covariance information for potential parameters. Scattering predictions generated from our ensembles show improved performance both against the original training corpora of experimental data and against a new "test" corpus comprising many of the experimental single-nucleon scattering data collected over the last twenty years. Finally, we apply our uncertainty-quantified OMPs to two case studies of application-relevant cross sections. We conclude that, for many common applications of OMPs, including OMP uncertainty should become standard practice. To facilitate their immediate use, digital versions of our updated OMPs and related tools for forward uncertainty propagation are included as Supplementary Material.

Introduction

For more than fifty years, optical-model potentials (OMPs) have played an important role in nuclear scattering calculations by providing effective projectile-target interactions. Early successes in fitting basic phenomenological OMPs to elastic scattering data [1] motivated continuing theoretical improvements on several fronts, including construction of (semi-)microscopic OMPs via the local density approximation [2, 3, 4, 5], extension to deformed and actinide systems [6, 7], and formal connection with the single-particle Green’s function via application of relevant dispersion relations [8, 9, 10, 11, 12, 13, 14, 15]. The recent development of a global microscopic OMP [16] based on several χ\chiEFT nucleon-nucleon (NN) potentials opens a promising new avenue for making predictions of scattering on unstable nuclides with a minimum of phenomenology. For recent reviews of OMP topics, see [17, 18].

Despite these advances, a number of basic questions remain about the uncertainty and generality of OMPs. First are questions of interpolation and extrapolation: how far can OMPs be trusted to generate reliable scattering predictions where experimental data are not available, especially away from β\beta-stability? As new rare isotope beam facilities come online, reliable estimates of scattering on unstable targets will be needed to make sense of the wealth of new data that are anticipated. For meaningful comparison with these new data, OMP predictions should come equipped with well-calibrated uncertainty estimates, estimates that are typically absent from widely used phenomenological OMPs, such as the

Chapel-Hill 89’ OMP [19] (intended for 40≤A≤20940 \le A \le209 from 10 to 65 MeV) and Koning-Delaroche OMP [20] (intended for 24≤A≤20924 \le A \le209 from 0.001 to 200 MeV). In principle, a global microscopic OMP based on a χ\chiEFT-derived NN potential, such as [16], come “naturally” equipped with uncertainties from truncation in the chiral expansion and should be less prone to under- or overfitting problems that affect phenomenological potentials. To date, however, microscopic models do not achieve the accuracy of phenomenological OMPs in regions where experimental data do exist, especially for inelastic scattering observables, which may diminish their utility for nuclear data applications. Were it available, knowledge of OMP uncertainties would help evaluators rank the relative importance of OMPs among other sources of uncertainty that enter reaction models, such as nuclear level densities and γ\gamma-ray strength functions [21].

The second type of questions concern the functional form of potentials and their capacity to realistically describe the underlying physics. As a simple example, the Koning-Delaroche OMP includes an imaginary spin-orbit term, but the Chapel Hill ’89 OMP does not. Does inclusion of this term result in meaningful differences in scattering predictions, and if so, which experimental data actually constrain its parameters? The form of nonlocal terms [22, 23, 24], shape of the hole potential and relation to dispersive correctness [13], and the correct dependence of parameters on nuclear asymmetry [25] are important open topics that would benefit from a firmer understanding of uncertainty in extant OMPs.

To clarify these issues, several recent studies have investigated uncertainty-quantification (UQ) techniques for phenomenological OMPs, including direct comparisons of frequentist and Bayesian methods for model calibration [26, 27], introduction of Gaussian process priors for energy-dependent parameters [28], and introduction of an ad-hoc dedicated “model uncertainty” term in a dispersive OMP analysis [29]. The ambitious study of [21] confronts the mature reaction code TALYS [30] with virtually the entire EXFOR experimental reaction database [31] with the specific intent of generating uncertainties for evaluations. Such theoretical studies are being complemented by the creation of templates for experimentalists for capturing the many (often undercharacterized) sources of uncertainty in experimental measurement [32, 33, 34], designed specifically so that newly collected data will be maximally useful for theory and evaluation efforts going forward. Most recently, the work of [28] proposes a statistically sound, reproducible pipeline for nuclear data evaluations, including characterization of OMP uncertainties, demonstrating the potential to accelerate and standardize the challenging process of evaluation.

Despite these methodological improvements over the last decade, many OMP users do not yet consider the OMP contribution to the uncertainty budget of their applications, either because it is assumed to be negligible or because tools to do so are difficult to use. Those that do (e.g. [35, 36]) typically estimate uncertainty qualitatively by manually varying a handful of parameters thought to be important and by comparing predictions from a handful non-UQ OMPs against each other by eye. Even when OMP parameter uncertainty estimates are available (e.g., [30]), they are more often based on hard-earned evaluator intuition rather than on detailed tests of empirical performance. In the ideal case, each OMP would ship with complete covariance information for potential parameters, be tested against trusted, easily-accessed data libraries, be based on reproducible statistical practices and stated assumptions, and make it easy for any downstream user to forward-propagate OMP uncertainty into their research application. Robust OMP UQ of this type would be a building block for larger UQ efforts such as the evaluation efforts mentioned earlier [21, 28] or improved experimental analysis pipelines. Motivating and demonstrating such a framework for phenomenological OMP UQ is the main goal of the present work.

To demonstrate our approach, we apply it to both the Koning-Delaroche global OMP (KD) [20] and the Chapel-Hill ’89 OMP (CH89) [19], yielding two new uncertainty-quantified OMP ensembles we designate KDUQ and CHUQ, respectively. To train these OMPs, we recompiled the same training data corpora as used in the original treatments (we refer to our recompilations as the CHUQ corpus and KDUQ corpus). The resulting UQ OMPs can be directly inserted into existing user codes to incorporate the parametric uncertainty of these OMPs. By applying our approach to multiple OMPs and comparing with microscopic and semi-microscopic alternatives, we can develop insight into how the next generation of uncertainty-equipped potentials can be gainfully constructed. In particular, we will emphasize the importance of two key steps in fitting phenomenological OMPs — managing outliers and experimental uncertainty underestimation — that are paramount for empirical UQ, both in OMPs and otherwise.

Our findings are organized in the following sections. Section II introduces the generic parameter inference problem and its application to OMP fitting, including challenges faced in the original CH89 and KD analyses. Section III proposes a new likelihood function and inference strategy based on Markov-Chain Monte Carlo (MCMC) that we argue is better suited for generating realistic OMP uncertainties. Section IV applies this strategy to retrain the KD and CH89 OMP forms against their original training data, yielding updated, uncertainty-quantified OMPs: KDUQ and CHUQ. Section V illustrates the impact of KDUQ and CHUQ both on Hauser-Feshbach calculations for two radiative capture test cases and on the reliability of OMP extrapolation along neutron-proton asymmetry. Section VI summarizes our findings. Following the main text, further technical details appear in the Appendix and three sections of Supplemental Material [37], including explicit definitions of the OMPs and scattering formulae, all experimental data used for training and testing, and our recommended KDUQ and CHUQ parameter values for future use. We hope that by providing thorough documentation, readers will be able to reproduce or extend our results without guesswork.

Challenges in OMP parameter inference

In this section, we first present a generic parameter inference problem, illustrating some common challenges with a pedagogical example. We then turn to the original KD and CH89 analyses, showing that certain assumptions, while necessary for making these canonical analyses tractable, can result in overconfidence in the fitted parameters.

Generic parameter inference

The goal of a parameter inference problem is to determine optimal parameters for a given functional form, where “optimal” usually means best matching a corpus of training data. In the specific case of OMP optimization, the OMP constitutes a model MM with unknown, possibly correlated potential parameters θ\boldsymbol{\theta}, and the task is to determine an optimal set of parameters θopt\boldsymbol{\theta}_{\mathrm{opt}} that minimizes the residuals between experimental scattering data and scattering-code predictions made using MM. (In these and all following definitions, we use bold typeface to denote a vector or matrix quantity.) A natural starting point for the probability density function of θ\boldsymbol{\theta} is to use a kk-dimensional normal distribution:

p(θ)=1(2π)k∣Σ∣e−12rTΣ−1rp(\boldsymbol{\theta}) = \frac{1}{\sqrt{(2\pi)^k\lvert\boldsymbol{\Sigma}\rvert}}e^{-\frac{1}{2}\mathbf{r}^{\mathsf{T}}\boldsymbol{\Sigma}^{-1}\mathbf{r}}
r=θ−θopt.(1)\mathbf{r} = \boldsymbol{\theta} - \boldsymbol{\theta}_{\mathrm{opt}}. \tag*{(1)}

Here, Σ\boldsymbol{\Sigma} is the k×kk \times k covariance matrix associated with θ\theta. If θopt\theta_{opt} and Σ\Sigma were known, the inference problem would be solved (at least up to the assumption of multivariate normality), with Σ\Sigma holding the variance information enabling downstream uncertainty propagation. Because we do not have direct measurements of θ\theta, only experimental scattering measurements, we cannot use Eq. (1) directly to train the model. Instead, we need to construct a likelihood function that connects the probability of observing a given experimental value given a candidate parameter vector. For OMP optimization, this involves mapping a candidate θ\theta to predicted scattering observables via a scattering code, evaluated at the relevant experimental indices (e.g., scattering energy, angle, target). This mapping is highly nonlinear in θ\theta as it involves, among other things, solving for the scattering matrix. Because the covariance matrix between experimental measurements is rarely known (discussed in detail in [21]), connecting OMP parameters with experimental data via selection of a likelihood function requires making certain assumptions about the scattering data. The overwhelming majority of past OMP analyses (including CH89 and KD) use a maximum likelihood approach based on some version of weighted-least-squares for their likelihood function:

L(y∣x,δy,θ)=e−12∑iri2δyi2(2)L(\mathbf{y}|\mathbf{x},\boldsymbol{\delta}_y,\theta)=e^{-\frac{1}{2}\sum_i\frac{r_i^2}{\delta_{y_i}^2}} \tag*{(2)}
ri=yi−M(θ,xi).r_i=y_i-M(\theta,x_i).

In this expression, for the ithi^{th} training data point, xix_i are the experimental conditions (such as energy, angle, etc.), yiy_i is the observed value, such as the cross section, and δyi\delta_{y_i} is the reported uncertainty of the observed value. Thus (x,y,δy)(\mathbf{x},\mathbf{y},\boldsymbol{\delta}_y) denotes the entire training corpus. Experimental data also often include an estimate of uncertainty in the experimental conditions δx\delta_x but these are usually omitted from the OMP analysis as they are more difficult to incorporate using standard optimization approaches. The predicted values, M(θ,x)M(\theta,\mathbf{x}), are an output of the scattering code evaluated at each x\mathbf{x} and using the OMP realization M(θ)M(\theta) for the projectile-target interaction.

If several conditions apply, including model linearity in the parameters, experimental uncertainties characterized by a known, positive-definite covariance matrix, and measurement samples being drawn from the same underlying distribution, the weighted least-squares estimator (Eq. (2)) guarantees an analytic solution that minimizes bias in θopt\theta_{opt} [38]. Unfortunately for OMP analysts, each of these conditions is violated in traditional OMP optimization analyses that are concerned primarily with θ\theta, and these violations are especially problematic for the present UQ task (Σ\Sigma estimation). Most impactful is the weighted-least-squares assumption that experimentally reported uncertainties are independent and complete (that is, that the vector of individual data point uncertainties δy\boldsymbol{\delta}_y fully represents the true, unknown data covariance matrix). In effect, this assumption assigns more independent information to residuals than they actually have, making the inference problem erroneously overdetermined and causing bias in θopt\theta_{opt} and underestimation of θopt\theta_{opt} uncertainty. Even if the full experimental data covariance were known, the OMP, by definition, is a projection of the true projectile-target interaction onto a reduced space of simple potential forms. As such we should expect it to suffer at least somewhat from “model defects” that, if unaccounted for during inference, may lead to overconfidence in an incorrect θopt\theta_{opt}, as demonstrated for a simple physical model in [39]. Further, model nonlinearity in θ\theta means that the likelihood function surface is not guaranteed to be convex, which can stymie simple optimization approaches such as gradient descent but which may be tractable with other optimization algorithms such as, e.g., simulated annealing.

A toy model

To illustrate how outliers and uncertainty underestimation impact parameter inference, we present a toy problem using a simple linear model. Imagine we wish to describe some generic phenomenon, T(x)T(x), that occurs on a domain x∈[−1,1]x \in[-1,1]. The true T(x)T(x) is:

T(x)=2.5P0(x)+2.0P1(x)+1.5P2(x)+1.0P3(x),(3)T(x)=2.5P_0(x)+2.0P_1(x)+1.5P_2(x)+1.0P_3(x), \tag*{(3)}

where PnP_n is the nthn^{th} Legendre polynomial. Suppose we know the functional form of T(x)T(x) but not the values of the coefficients, which we would like to learn through inference against data. So we collect yy observations at experimental conditions xx, using a device subject to measurement uncertainty. Aware of this uncertainty, we estimate measurement imprecision for each data point as δy\delta_y. We then define a model, MM, and compare model predictions M(x,θ)M(x,\theta) to the measured data, where θ\theta are the nn unknown coefficients that we want to learn. Because our model is linear in θ\theta and our data measurements are independent and uncorrelated, Eq. (2) provides the best unbiased estimator of the true coefficients, denoted θtrue\theta_{true}. We can find an optimum set of parameter values θopt\theta_{opt} analytically using maximum likelihood estimation or numerically using, e.g., gradient descent until we reach some threshold for convergence. The covariance matrix at θopt\theta_{opt} is the inverse of the Hessian matrix H(θopt)\mathcal{H}(\theta_{opt}), which can be easily assessed numerically.

So far, we have described a simple, generic inference problem and its solution. We now consider four possible scenarios for solving this problem, each involving a different possible distribution for yy and δy\delta_y. These differing distributions are plotted in panels (a) to (d) of Figure 1, and defined according to:

Four data-fitting scenarios and normalized residual histograms

Figure 1. The four data-fitting scenarios for the toy model discussed in the text are compared. Panel (a) shows a fit to data with accurate uncertainty estimates and no outliers. Panel (b) shows a fit to data with underestimated uncertainties and no outliers. Panel (c) shows a fit to data with accurate uncertainty estimates but with outliers. Panel (d) shows a fit using data with both underestimated uncertainties and with outliers. The simulated data used for fitting are shown as black bars, the “true” underlying function used to generate the data is shown as a gray line, and the fit to the “experimental” data is shown in red. In panels (e) to (h), the normalized residuals for data in the corresponding plots are histogrammed. A normal distribution with μ=0\mu= 0 and σ2=1\sigma^{2} = 1 (gray dashed line) is shown for reference.

yi∼N(T(xi),0.32)δyi=0.32T(xi)(4)\begin{aligned} y_i &\sim\mathcal{N}(T(x_i),0.3^2)\\ \delta_{y_i} &= 0.32T(x_i) \tag*{(4)} \end{aligned}
yi∼N(T(xi),0.322)δyi=0.10T(xi)(5)\begin{aligned} y_i &\sim\mathcal{N}(T(x_i),0.32^2)\\ \delta_{y_i} &= 0.10T(x_i) \tag*{(5)} \end{aligned}
yi∼N(T(xi),0.322)+αδyi=0.32T(xi)(6)\begin{aligned} y_i &\sim\mathcal{N}(T(x_i), 0.32^{2}) + \alpha\\ \delta_{y_i} &= 0.32T(x_i) \tag*{(6)} \end{aligned}

for each ii, where

α{∼N(3,0.62),10% probability,=0,90% probability.\alpha\begin{cases} \sim\mathcal{N}(3,0.6^{2}), & 10\% \text{ probability},\\ =0, & 90\% \text{ probability}. \end{cases}
yi∼N(T(xi),0.322)+αδyi=0.10T(xi),(7)\begin{aligned} y_i &\sim\mathcal{N}(T(x_i), 0.32^{2}) + \alpha\\ \delta_{y_i} &= 0.10T(x_i), \tag*{(7)} \end{aligned}

In this notation, ∼N(μ,σ2)\sim\mathcal{N}(\mu,\sigma^{2}) refers to sampling from a normal distribution of mean μ\mu and variance σ2\sigma^{2}.

We begin with the first scenario, shown in panel (a) of Figure 1. This is the best-case scenario, given the assumptions appropriate for weighted-least-squares: our measuring device suffers from zero bias and the true mean measurement uncertainties δy\delta_y are known (Eq. (4)). For example, our measuring device exhibits independent statistical and systematic uncertainties of 10% and 30%, yielding a total 32% total uncertainty via addition in quadrature. Because both the measured data and their uncertainties are faithful to the true underlying distribution, our estimated θopt\boldsymbol{\theta}_{opt} match θtrue\boldsymbol{\theta}_{true}, up to the estimated uncertainty of θopt\boldsymbol{\theta}_{opt}. Panel (e) shows that the distribution of standardized residuals between our model’s predictions and the corresponding experimental data are distributed according to a normal distribution with unit variance.

Panel (b) of Figure 1 shows the outcome of the second scenario: our measuring device performs identically as in the first scenario, but now our estimates of δy\delta_y are too small (Eq. (5)). This could arise if, for instance, both statistical and systematic uncertainty contribute to the overall uncertainty of our measuring device, but we have only recognized and reported the statistical uncertainty. Because the minimum of our weighted-least-squares likelihood function is not affected by overall rescalings of δy\delta_y, we recover the same θopt\boldsymbol{\theta}_{opt} as in the first scenario. However, our uncertainty estimates of θopt\boldsymbol{\theta}_{opt} have shrunk by a factor of three — the same factor by which we underestimated the measurement uncertainty – because the Hessian H(θopt)\mathcal{H}(\boldsymbol{\theta}_{opt}) scales proportionally with δy\delta_y. Panel (f) shows that while the standardized residuals remain normally distributed with a mean of zero, they are more dispersed than the reference distribution. Thus, underestimation of experimental uncertainties directly causes underestimation of parametric uncertainties. This is a generic feature of parameter inference and, as we will show in the following section, affects most previous OMP analyses.

Panel (c) of Figure 1 presents a third scenario: as in the first scenario, we have accurately estimated the experimental uncertainty δ\delta, but now our experimental device occasionally returns anomalous measurements (so-called “outliers”). The simulated data y\mathbf{y} have been drawn according to Eq. (6): each measurement has a 10% chance of being shifted upward by α\alpha, which is an artificial “outlier factor”. This is meant to represent a more realistic situation in which some fraction of experimental data are inconsistent with the model, either because of model defects or because of problems during experimental data collection. The outliers “pull” on the likelihood function, causing our recovered θopt\boldsymbol{\theta}_{opt} to differ from those of the previous scenarios, but, because our δy\delta_y are the same as in the first scenario, our uncertainty estimates of θopt\boldsymbol{\theta}_{opt} do not change. The parameter bias appears in panel (g) as asymmetry in the standardized residuals with respect to the reference distribution, even as the variance of the residuals is the same as in the first scenario. We note that even if our measuring device returned no outliers, if our underlying model was incorrect (i.e., model defect), certain data would appear to be outliers, and we would obtain a similar result.

Finally, panel (d) of Figure 1 combines the second and third scenarios: y\mathbf{y} contains occasional outliers and δy\delta_y are overconfident (Eq. (7)). Accordingly, our estimated θopt\boldsymbol{\theta}_{opt} is biased and our uncertainty estimates of θopt\boldsymbol{\theta}_{opt} are overconfident about the biased estimates. Both the bias and the dispersion of the normalized residuals are visible in panel (h). This scenario is the best analog to the OMP optimization task. For us to obtain well-calibrated uncertainties that span the experimental data, our loss function and optimization strategy must address both challenges: namely, underestimation of experimental (co)variances, and fundamental discrepancies between the model and data either due to model defects or problems with experimental data collection (which we do not attempt to disentangle).

Challenges for CH89 and KD

The difficulties of using weighted-least-squares estimators are well-known to OMP designers, including those of CH89 and KD. A common symptom is that initial fits to experimental data are often grossly unsatisfactory, clearly missing “the physics” present in the scattering data, leading to manual parameter adjustment. The authors of CH89 comment that, early in their analysis, there were often “significant contributions from the data that the model is not able to describe” even when training to a single scattering data set. They tested several alternative loss functions but found that in “reduc[ing] the emphasis of outlying points” they “lost sensitivity to even the good data”. After testing various functions, their compromise was to introduce a weight factor to their likelihood function for each data set ss, equal to the minimum loss for that data set obtained in a fit to only that data set, i.e.,

L(y∣x,δy,θ)=∑sLs(θ∣xs,ys)min⁡(Ls(θ∣xs,ys))(8)L(\mathbf{y}\mid\mathbf{x},\delta_y,\boldsymbol{\theta}) = \sum_s \frac{L_s(\boldsymbol{\theta}\mid\mathbf{x}_s,\mathbf{y}_s)}{\min\left(L_s(\boldsymbol{\theta}\mid\mathbf{x}_s,\mathbf{y}_s)\right)} \tag*{(8)}

where LsL_s is the contribution from data set ss to the overall weighted-least-squares fit as in Eq. (2). By deemphasizing data sets that were poor matches to the form of their OMP, they achieved a better visual fit to their training data. However, this solution also introduces problems: the introduced weights are not easily interpreted nor do they preserve the normalization of the likelihood function, which is important for estimating Σ\boldsymbol{\Sigma}. However, because finding θopt\boldsymbol{\theta}_{opt} is insensitive to overall rescalings of LL, most past authors have been willing to sacrifice the possibility of accurately estimating Σ\boldsymbol{\Sigma} in order to improve their single “best-fit” parameter vector.

Koning and Delaroche identified this issue in their global OMP characterization as well and also provided extensive quantitative evidence that traditional OMPs are incapable of reproducing the bulk of experimental data within the range of reported experimental uncertainties. In Table 12 of their OMP analysis [20], they present sums of uncertainty-weighted square residuals per degree of freedom (a reduced-χ2\chi^2 metric) for several prominent OMPs against a variety of experimental data sets. In their analysis, a value near unity was taken as an indication of good model-data agreement. For the widely used global OMPs they considered, they found values of χ2/N\chi^2/N ranging from 6.3 to 11.2 for differential elastic scattering cross sections and from 2.3 to 9.2 for neutron total cross sections. Using their new potential (KD), they found values of χ2/N\chi^2/N ranging from 4.5 to 7.4 for differential elastic scattering cross sections and from 1.2 to 6.7 for neutron total cross sections, depending on the experimental data corpus tested against. They echoed the comments of the CH89 authors, noting that “the optimization procedure is very sensitive” to underestimations in reported experimental uncertainties such that “even a slightly incorrect error estimation can easily vitiate an automated fitting procedure”. In [21], Koning further analyzed model-experiment discrepancies across the EXFOR database and combined several proposed remedies into an “evaluated” χ2\chi^2 expression meant to overcome the issues of using naïve weighted least squares.

To better understand these discrepancies between the trained model and training data, we began by reproducing the original CH89 and KD analyses. Figure 2 summarizes the performance of the standard CH89 and KD potentials against the experimental data used to train them, as reconstructed in the present work. For each experimental datum, the normalized residual for that datum (r/δyr/\delta_y) was tabulated, then all residuals histogrammed according to data type. In addition, in panel (b), two dotted curves show the performance of CH89 when the CH89 parameters are resampled according to the parameter covariance matrix presented in the original publication. If the assumptions underpinning weighed-least-squares were fulfilled, each line should follow the gray dashed line (a normal distribution with unit variance), indicating that the CH89 and KD predictions match the mean of the experimental data used to train them, and that the training data are dispersed about the predictions in keeping with their reported uncertainties. In reality, all types of scattering data show a variance several times larger than unity, an indication either of underestimation of experimentally reported uncertainties or of significant model deficiencies, or both. The means of the distributions are offset to varying degree, indicating that the canonical θopt\theta_{\mathrm{opt}} for these OMPs retain some bias with respect to the underlying experimental data. Table I lists the mean, standard deviation, and skewness of these observed distributions for each data type used to train the CH89 and KD OMPs. This confirms the issues identified by past authors: clearly, these OMPs do not span the variance of their training data, and for some data types, predictions show systematic bias with respect to experiment.

References

  1. [1]F. D. Becchetti and G. W. Greenlees, Phys. Rev. 182, 1190 (1969), URL https://link.aps.org/doi/10.1103/PhysRev.182.1190.DOI
  2. [2]J.-P. Jeukenne, A. Lejeune, and C. Mahaux, Phys. Rev. C 16, 80 (1977), URL https://link.aps.org/doi/10.1103/PhysRevC.16.80.DOI
  3. [3]J. Jeukenne, A. Lejeune, and C. Mahaux, Physics Reports 25, 83 (1976), ISSN 0370-1573, URL https://www.sciencedirect.com/science/article/pii/03701537690017X.
  4. [4]E. Bauge, J. P. Delaroche, and M. Girod, Phys. Rev. C 58, 1118 (1998), URL https://link.aps.org/doi/10.1103/PhysRevC.58.1118.DOI
  5. [5]E. Bauge, J. P. Delaroche, and M. Girod, Phys. Rev. C 63, 024607 (2001), URL https://link.aps.org/doi/10.1103/PhysRevC.63.024607.DOI
  6. [6]G. P. A. Nobre, A. Palumbo, M. Herman, D. Brown, S. Hoblit, and F. S. Dietrich, Phys. Rev. C 91, 024618 (2015), URL https://link.aps.org/doi/10.1103/PhysRevC.91.024618.DOI
  7. [7]E. S. Soukhovitskii, R. Capote, J. M. Quesada, S. Chiba, and D. S. Martynov, Phys. Rev. C 94, 064605 (2016), URL https://link.aps.org/doi/10.1103/PhysRevC.94.064605.DOI
  8. [8]C. Mahaux and R. Sartor, Adv. Nucl. Phys. 20, 1 (1991).
  9. [9]J. M. Quesada, R. Capote, A. Molina, M. Lozano, and J. Raynal, Phys. Rev. C 67, 067601 (2003), URL https://link.aps.org/doi/10.1103/PhysRevC.67.067601.DOI
  10. [10]B. Morillon and P. Romain, Phys. Rev. C 70, 014601 (2004), URL https://link.aps.org/doi/10.1103/PhysRevC.70.014601.
  11. [11]B. Morillon and P. Romain, Phys. Rev. C 76, 044601 (2007), URL https://link.aps.org/doi/10.1103/PhysRevC.76.044601.
  12. [12]J. M. Mueller, R. J. Charity, R. Shane, L. G. Sobotka, S. J. Waldecker, W. H. Dickhoff, A. S. Crowell, J. H. Esterline, B. Fallin, C. R. Howell, et al., Phys. Rev. C 83, 064605 (2011), URL https://link.aps.org/doi/10.1103/PhysRevC.83.064605.DOI
  13. [13]M. H. Mahzoon, R. J. Charity, W. H. Dickhoff, H. Dussan, and S. J. Waldecker, Phys. Rev. Lett. 112, 162503 (2014), URL https://link.aps.org/doi/10.1103/PhysRevLett.112.162503.DOI
  14. [14]M. Atkinson and W. Dickhoff, Physics Letters B 798, 135027 (2019), ISSN 0370-2693, URL https://www.sciencedirect.com/science/article/pii/S037026931930749X.
  15. [15]X. Zhao, W. Sun, R. Capote, E. S. Soukhovitskii, D. S. Martynov, and J. M. Quesada, Phys. Rev. C 101, 064618 (2020), URL https://link.aps.org/doi/10.1103/PhysRevC.101.064618.DOI
  16. [16]T. R. Whitehead, Y. Lim, and J. W. Holt, Phys. Rev. Lett. 127, 182502 (2021), URL https://link.aps.org/doi/10.1103/PhysRevLett.127.182502.DOI
  17. [17]W. H. Dickhoff, R. J. Charity, and M. H. Mahzoon, Journal of Physics G: Nuclear and Particle Physics 44, 033001 (2017), URL https://doi.org/10.1088/1361-6471/44/3/033001.
  18. [18]W. H. Dickhoff and R. J. Charity, Prog. Part. Nucl. Phys. (2018).
  19. [19]R. Varner, W. Thompson, T. McAbee, E. Ludwig, and T. Clegg, Physics Reports 201, 57 (1991), ISSN 0370-1573, URL https://www.sciencedirect.com/science/article/pii/0370175391900390.
  20. [20]A. Koning and J. Delaroche, Nuclear Physics A 713, 231 (2003), ISSN 0375-9474, URL https://www.sciencedirect.com/science/article/pii/S0375947402013210.
  21. [21]A. Koning, Eur. Phys. J. A 51, 184 (2015).
  22. [22]H. F. Arellano and G. Blanchon, Phys. Rev. C 98, 054616 (2018), URL https://link.aps.org/doi/10.1103/PhysRevC.98.054616.DOI
  23. [23]H. Arellano and G. Blanchon, Physics Letters B 789, 256 (2019), ISSN 0370-2693, URL https://www.sciencedirect.com/science/article/pii/S0370269318309274.
  24. [24]H. Arellano and G. Blanchon, Eur. Phys. J. A 57, 27 (2021), URL https://doi.org/10.1140/epja/s10050-020-00328-0.
  25. [25]S. Goriely and J.-P. Delaroche, Physics Letters B 653, 178 (2007), ISSN 0370-2693, URL https://www.sciencedirect.com/science/article/pii/S0370269307009069.
  26. [26]G. B. King, A. E. Lovell, L. Neufcourt, and F. M. Nunes, Phys. Rev. Lett. 122, 232502 (2019), URL https://link.aps.org/doi/10.1103/PhysRevLett.122.232502.DOI
  27. [27]A. E. Lovell, F. M. Nunes, M. Catacora-Rios, and G. B. King, Journal of Physics G: Nuclear and Particle Physics 48, 014001 (2020), URL https://doi.org/10.1088/1361-6471/abba72.
  28. [28]G. Schnabel, H. Sjöstrand, J. Hansson, D. Rochman, A. Koning, and R. Capote, Nuclear Data Sheets 173, 239 (2021), ISSN 0090-3752, Special Issue on Nuclear Reaction Data, URL https://www.sciencedirect.com/science/article/pii/S009037522100017X.
  29. [29]C. D. Pruitt, R. J. Charity, L. G. Sobotka, J. M. Elson, D. E. M. Hoff, K. W. Brown, M. C. Atkinson, W. H. Dickhoff, H. Y. Lee, M. Devlin, et al., Phys. Rev. C 102, 034601 (2020), URL https://link.aps.org/doi/10.1103/PhysRevC.102.034601.DOI
  30. [30]A. Koning and D. Rochman, Nuclear Data Sheets 113, 2841 (2012), ISSN 0090-3752, Special Issue on Nuclear Reaction Data, URL https://www.sciencedirect.com/science/article/pii/S0090375212000889.
  31. [31]Experimental Nuclear Reaction Data (EXFOR), online. Accessed 2022-07-06., URL https://www-nds.iaea.org/exfor/exfor.htm.
  32. [32]D. Smith and N. Otuka, Nuclear Data Sheets 113, 3006 (2012), ISSN 0090-3752, Special Issue on Nuclear Reaction Data, URL https://www.sciencedirect.com/science/article/pii/S0090375212000907.
  33. [33]Neudecker, Denise, Hejnal, Brooke, Tovesson, Fredrik, White, Morgan C., Smith, Donald L., Vaughan, Diane, and Capote, R., EPJ Nuclear Sci. Technol. 4, 21 (2018), URL https://doi.org/10.1051/epjn/2018026.
  34. [34]D. Neudecker, D. Smith, F. Tovesson, R. Capote, M. White, N. Bowden, L. Snyder, A. Carlson, R. Casperson, V. Pronyaev, et al., Nuclear Data Sheets 163, 228 (2020), ISSN 0090-3752, URL https://www.sciencedirect.com/science/article/pii/S0090375219300729.
  35. [35]S. Harissopulos, E. Vagena, P. Dimitriou, M. Axiotis, S. Galanopoulos, V. Foteinou, and A. Lagoyannis, Phys. Rev. C 104, 025804 (2021), URL https://link.aps.org/doi/10.1103/PhysRevC.104.025804.DOI
  36. [36]G. Gyürky, E. Somorjai, Z. Fü­löp, S. Harissopulos, P. Demetriou, and T. Rauscher, Phys. Rev. C 64, 065803 (2001), URL https://link.aps.org/doi/10.1103/PhysRevC.64.065803.DOI
  37. [37]See Supplemental Material at [URL will be inserted by publisher] for optical potential parameter values (Supplement A), training and test data used in our analysis (Supplement B), and figures comparing potential performance against these data (Supplement C).
  38. [38]A. C. Aitken, Proceedings of the Royal Society of Edinburgh 55, 42b (1936).
  39. [39]J. Brynjarsdóttir and A. O’Hagan, Inverse Problems 30 (2014).
  40. [40]R. N. Pérez, J. E. Amaro, and E. R. Arriola, Phys. Rev. C 88, 064002 (2013), URL https://link.aps.org/doi/10.1103/PhysRevC.88.064002.DOI
  41. [41]F. Gross and A. Stadler, Phys. Rev. C 78, 014005 (2008), URL https://link.aps.org/doi/10.1103/PhysRevC.78.014005.DOI
  42. [42]R. Eastgate, W. Thompson, and R. Hardekopf, Computer Physics Communications 5, 69 (1973), ISSN 0010-4655, URL https://www.sciencedirect.com/science/article/pii/0010465573900131.
  43. [43]F. James and M. Roos, Computer Physics Communications 10, 343 (1975), ISSN 0010-4655, URL https://www.sciencedirect.com/science/article/pii/0010465575900399.
  44. [44]W. E. Ormand, Yahfc, URL https://github.com/LLNL/Yet-Another-Hauser-Feshbach-Code.DOI
  45. [45]D. Foreman-Mackey, D. W. Hogg, D. Lang, and J. Goodman, Publications of the Astronomical Society of the Pacific 125, 306 (2013), URL https://doi.org/10.1086/670067.
  46. [46]P. Descouvemont and D. Baye, Rep. Prog. Phys. 73, 036301 (2010), URL https://iopscience.iop.org/article/10.1088/0034-4885/73/3/036301.
  47. [47]D. Baye, Physics Reports 565, 1 (2015), ISSN 0370-1573, the Lagrange-mesh method, URL https://www.sciencedirect.com/science/article/pii/S0370157314004086.
  48. [48]Galassi, M. et al., GNU Scientific Library Manual (3rd Ed.), URL https://www.gnu.org/software/gsl.
  49. [49]I. J. Thompson, Computer Physics Reports 7, 167 (1988), ISSN 0167-7977, URL https://www.sciencedirect.com/science/article/pii/0167797788900056.
  50. [50]I. J. Thompson, Fresco, URL https://github.com/LLNL/Frescox.
  51. [51]O. Tarasov and D. Bazin, Nucl. Instrum. Meth. B 376, 185 (2016), ISSN 0168-583X, URL https://www.sciencedirect.com/science/article/pii/S0168583X1600224X.
  52. [52]O. Tarasov, Lise++, URL http://lise.nscl.msu.edu/lise.html.
  53. [53]A. Ingemarsson, Physica Scripta 9, 156 (1974), URL https://doi.org/10.1088/0031-8949/9/3/004.
  54. [54]M. C. Atkinson, M. H. A. Hassan, M. A. Keim, B. A. Bordelon, C. D. Pruitt, R. J. Charity, and W. H. Dickhoff, Phys. Rev. C 101, 044303 (2020), URL https://link.aps.org/doi/10.1103/PhysRevC.101.044303.DOI
  55. [55]C. D. Pruitt, R. J. Charity, L. G. Sobotka, M. C. Atkinson, and W. H. Dickhoff, Phys. Rev. Lett. 125, 102501 (2020), URL https://link.aps.org/doi/10.1103/PhysRevLett.125.102501. L. Kurth Kerr, et al., Nuclear Physics A 653, 341 (1999), ISSN 0375-9474, URL https://www.sciencedirect.com/science/article/pii/S0375947499002365.DOI
  56. [56]B. Morillon and P. Romain, Phys. Rev. C 74, 014601 (2006), URL https://link.aps.org/doi/10.1103/PhysRevC.74.014601.
  57. [57]R. Capote, S. Chiba, E. S. Soukhovitskii, J. M. Quesada, and E. Bauge, Journal of Nuclear Science and Technology 45, 333 (2008).
  58. [58]R. Shane, R. J. Charity, J. M. Elson, L. G. Sobotka, M. Devlin, N. Fotiades, and J. M. O‘Donnell, Nucl. Instrum. Meth. 614, 468 (2010), URL http://dx.doi.org/10.1016/j.nima.2010.01.005.
  59. [59]F. S. Dietrich, J. D. Anderson, R. W. Bauer, S. M. Grimes, R. W. Finlay, W. P. Abfalterer, F. B. Bateman, R. C. Haight, G. L. Morgan, E. Bauge, et al., Phys. Rev. C 67, 044606 (2003), URL https://link.aps.org/doi/10.1103/PhysRevC.67.044606.DOI
  60. [60]P. Danielewicz, P. Singh, and J. Lee, Nuclear Physics A 958, 147 (2017), ISSN 0375-9474, URL https://www.sciencedirect.com/science/article/pii/S0375947416302895.
  61. [61]R. Surman, M. Mumpower, R. Sinclair, K. L. Jones, W. R. Hix, and G. C. McLaughlin, AIP Advances 4, 041008 (2014), https://doi.org/10.1063/1.4867191, URL https://doi.org/10.1063/1.4867191.
  62. [62]M. Thiel, C. Sfienti, J. Piekarewicz, C. J. Horowitz, and M. Vanderhaeghen, Journal of Physics G: Nuclear and Particle Physics 46, 093003 (2019), URL https://doi.org/10.1088/1361-6471/ab2c6d.
  63. [63]H. S. Camarda, T. W. Phillips, and R. M. White, Phys. Rev. C 29, 2106 (1984), URL https://link.aps.org/doi/10.1103/PhysRevC.29.2106.DOI
  64. [64]R. E. Shamu, E. M. Bernstein, J. J. Ramirez, and C. LaGrange, Phys. Rev. C 22, 1857 (1980), URL https://link.aps.org/doi/10.1103/PhysRevC.22.1857.DOI
  65. [65]G. C. Jon, H. Orihara, T. Niizeki, M. Oura, K. Ishii, A. Terakawa, M. Hosaka, K. Itoh, C. C. Yun, Y. Fujii, et al., Phys. Rev. C 56, 900 (1997), URL https://link.aps.org/doi/10.1103/PhysRevC.56.900.DOI
  66. [66]R. W. Finlay, W. P. Abfalterer, G. Fink, E. Monteil, T. Adami, P. W. Lisowski, G. L. Morgan, and R. C. Haight, Phys. Rev. C 47, 237 (1993), URL http://dx.doi.org/10.1103/PhysRevC.47.237.
  67. [67]W. P. Abfalterer, F. B. Bateman, F. S. Dietrich, R. W. Finlay, R. C. Haight, and G. L. Morgan, Phys. Rev. C 63, 044608 (2001), URL http://dx.doi.org/10.1103/PhysRevC.63.044608.
  68. [68]S. Galanopoulos, P. Demetriou, M. Kokkoris, S. Harissopulos, R. Kunz, M. Fey, J. W. Hammer, G. Gyürky, Z. Fülep, E. Somorjai, et al., Phys. Rev. C 67, 015801 (2003), URL https://link.aps.org/doi/10.1103/PhysRevC.67.015801.DOI
  69. [69]E. Vagena, M. Axiotis, and P. Dimitriou, Phys. Rev. C 103, 045806 (2021), URL https://link.aps.org/doi/10.1103/PhysRevC.103.045806.DOI
  70. [70]Y. Xu, S. Goriely, A. J. Koning, and S. Hilaire, Phys. Rev. C 90, 024604 (2014), URL https://link.aps.org/doi/10.1103/PhysRevC.90.024604.DOI
  71. [71]R. F. Carlson, Atom. Data Nucl. Data Tables 63, 93 (1996), URL http://www.sciencedirect.com/science/article/pii/S0092640X96900108.
  72. [72]A. Ingemarsson, J. Nyberg, P. Renberg, O. Sundberg, R. Carlson, A. Auce, R. Johansson, G. Tibell, B. Clark,

Paper details

vlnc:W_rapm97sq
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.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents