Session JA: Bayesian Uncertainty Quantification: BayUQ

The Bayesian Analysis of Nuclear Dynamics Framework

The atomic nucleus is a complex system and simulation models provide essential insights into many nuclear physics (NP) phenomena. However, simulation models are imperfect representations of reality and they almost always include uncertainties associated with model assumptions, inputs, and experiments. Uncertainty quantification aims to systematically account for relevant uncertainties for more accurate and precise predictions. Quantifying the uncertainty in a model is crucial to establishing trust in its predictions, and it requires cutting edge research in machine and statistical learning that can handle the cost of running high-fidelity NP models. The goal of the Bayesian Analysis of Nuclear Dynamics (BAND) Framework is to facilitate the uncertainty quantification in NP models. The framework integrates four Bayesian statistical tools: 1) model emulation, 2) model calibration, 3) model mixing, and 4) experimental design. The project aims to transform computational and theoretical research in those four areas into tools and techniques for making reliable predictions of complex NP systems with well-quantified uncertainties. These techniques are delivered as publicly-available open-source software tools along with guidelines for users. The goal is for these software tools to play a transformative role in facilitating and producing a full assessment of the uncertainty in NP predictions. In this talk, we introduce the BAND framework and overview the four main statistical tools. Finally, we illustrate some of BAND's uncertainty quantification tools and techniques with real NP examples.   

Bayesian Tools for a Better Optical Model

Optical potentials are pervasive in the description of nuclear reactions and effectively take into account the many-body complexity of the projectile-target system.  They are often determined phenomenologically, predominately through fits to elastic scattering data.  The parametrizations of these potentials are constrained using reaction data on stable targets but then extrapolated to significantly more exotic (e.g. neutron-rich) systems.  The fitting procedure and extrapolation can lead to significant uncertainties in the resulting reaction observables that were typically only quantified by calculating the same theory using two different parametrizations.  Recently, we have developed a Bayesian optimization procedure that has been shown to give more realistic uncertainties, compared to standard chi-squared minimization and covariance propagation.  Modern statistical tools additionally provide the ability to compare the information content of observables and provide the means to explore which experiments would be most useful for giving insights and constraining theoretical models.  In this talk, we discuss three such tools:  principal component analysis, sensitivity analysis, and Bayesian evidence.  We first apply these tools to a toy model to demonstrate their effectiveness and then use them to investigate the information content of two reaction observables.

LA-UR-22-25990   

Bayesian analysis with information field approach and the inference of the temperature-dependent jet transport parameter

Bayesian inference has been widely applied to extract physical parameters with correlated uncertainty quantification. Many quantities of interest in high-energy nuclear physics are functions, such as the temperature- and energy-dependent transport coefficients and parton distribution functions. For functional inference, the choice of the prior distribution is subtle but also critical to obtaining reliable results. Existing studies performing functional inference heavily rely on explicit parametrization, which can impose unwanted long-range correlations in the input parameter space of the function, limiting the ability to incorporate datasets that are supposed to provide independent constraints in different input regions. 

We propose to use the information field method to solve the above problem, where the unconstrained function is treated as a random field with a typical correlation length that removes long-range correlations. As proof of the principle of this method, we performed a global Bayesian analysis on the temperature-dependent jet transport parameter using the NLO parton model calculations with Higher-Twist modified fragmentation. This study has included single hadron, dihadron, and gamma-hadron nuclear modification factors at both RHIC and LHC energies. We illustrate how the value of jet transport parameters from low to high temperatures is progressively constrained by incrementally including datasets from peripheral to central collisions from lower to higher beam energies. Furthermore, the information field approach allows straightforward sensitivity analysis to guide future measurements to target temperature regions with insufficient constraints.