Session D06: Minisymposium: Applications of Advanced Statistics and Machine Learning Methods in Nuclear Physics II

Applications of Advanced Statistics and Machine Learning Methods in Nuclear Physics

In this talk we will explore the role that machine learning and advanced statistics play in the theory-experiment/observation cycle of nuclear science. These multidisciplinary methods and techniques have become instrumental for several developments in the field over the last decade and are expected to become indispensable to fully capitalize on the investments in recent and upcoming computational, experimental, and observational facilities. We will discuss key highlights and examples including deep learning for data analysis and control, dimensionality reduction for model acceleration and discovery, and Bayesian machine learning for uncertainty quantification and experimental design. This overview will serve as an introduction for the cutting-edge talks that will follow in the mini symposium.   

Potentials of constrained Gaussian process modeling in Nuclear Physics

Gaussian processes have been broadly implemented in Nuclear Physics to approximate the underlying models as well as to understand hidden error structures in observations due to their flexibility in modeling and natural advantage in uncertainty quantification.  In this talk, we focus on two types of Gaussian process models under problem-orientated constraints and their implementation in Nuclear Physics problems. In the first part, we illustrate how to reduce uncertainties in constrained Gaussian process extrapolation via the proton radius puzzle problem, and we propose a general data-integrated Bayesian procedure to recover the source-dependent error structure in experimental data. In the second part, we discuss using a constrained Gaussian process to approximate the solution of a partial differential equation (PDE) where we treat the PDE as constraints over randomly drawn collocation points. A subtle trade-off between approximation accuracy and computational efficiency can be carried out by developing constraint-relaxed prior distributions in aforementioned two methods.   

Uncertainty Quantification for Phenomenological Optical Potentials

Optical-model potentials (OMPs), which model the interaction between a projectile and a nuclear target, remain an essential reaction theory ingredient. Though several esteemed phenomenological OMPs are available (including the Koning-Delaroche [KD] and Chapel Hill ’89 [CH89] OMPs), none are equipped with well-calibrated uncertainties. Without reliable UQ, experimentalists, reaction theorists, and evaluators have little way of knowing where to trust these models and where the models break down.

To address this gap, we revisited the classical KD and CH89 OMP analyses and identified two important statistical limitations that hampered the original analyses’ efforts at UQ. We then developed a generic OMP-UQ framework and assigned uncertainties to KD and CH89, yielding two new OMP ensembles, KDUQ and CHUQ. Against both the KD/CH89 training data and a large corpus of test data, our UQ versions show improved reliability and interpretability and can be dropped in anywhere KD and CH89 are already

used. To facilitate including OMP uncertainty as a standard practice in the field, we provide digital copies of our potentials and related tools for forward uncertainty propagation.   

Bayesian calibration and truncation error estimation of chiral effective theories

In the era of increasingly precise nuclear experiments, it has become imperative that high-quality models for nuclear interactions are created. However, we must extend our lens of high-quality to include rigorous quantification of uncertainty to compare model predictions to experimental observations properly. To this end, we employ models derived from Effective Field Theories (EFTs) derived from quantum chromodynamics, in particular pionless and chiral EFTs, as these have been highly useful in accurately calculating nuclear observables. Not only that, the perturbative structure of such EFTs provides a framework for an improbable interaction while simultaneously offering a robust and straightforward means of estimating model uncertainties by the truncation of the EFT model. Considering these ideas, we can employ a Bayesian framework to calibrate EFT models and estimate the truncation error. Furthermore, we identify the need to employ machine-learning techniques to emulate the calculation of nuclear observables to increase the deployability of Bayesian methods by reducing computational expenses.   

Mind the gaps: Applying Bayesian model mixing to the dense matter equation of state

The dense matter equation of state (EOS) has seen several breakthrough developments in recent years. Calculations in chiral effective field theory up to N3LO describe the low-density region, and calculations in perturbative QCD up to N2LO describe the deconfined quark matter that occurs at high densities. However, spanning the overall space in density that the EOS covers in a statistically consistent manner, with well-quantified uncertainties, remains a challenging problem. We present results for the EOS pressure and speed of sound squared obtained using Bayesian model mixing (BMM) of chiral effective field theory and perturbative QCD. Future work to improve the resulting EOS using additional information will be discussed.   

Bayesian analysis for constraining the equation of state of asymmetric nuclear matter

The equation of state of nuclear matter plays a crucial role in understanding the properties of neutron stars and heavy-ion collisions. However, due to the complexities of nuclear interactions, accurately constraining the EoS remains a challenging task. We present results of combining Bayesian analysis of the transport model simulations and the single particle spectra, and particle ratios from the Ca+Ni and Ca+Sn collisions at E/A=56 and 140 MeV performed at the National Superconducting Cyclotron Laboratory.    

Bayesian Refinement of Relativistic Mean Field Models

The last five years have seen remarkable progress in our quest to determine the equation of state of neutron rich matter. Recent advances across the theoretical, experimental, and observational landscape have been incorporated in a Bayesian framework to refine existing covariant energy density functionals previously calibrated by the properties of finite nuclei. In particular, constraints on the maximum neutron star mass from pulsar timing, on stellar radii from the NICER mission, on tidal deformabilities from the LIGO-Virgo collaboration, and on the dynamics of pure neutron matter as predicted from chiral effective field theories, have resulted in significant refinements to the models, particularly to those predicting a stiff symmetry energy. In addition to this, we examine new appraoches in reproducing simultaneously the neutron skin thickness of both 208Pb and 48Ca recently reported by the PREX/CREX collaboration.   

Deep Learning for Neutron Lifetime Measurement

The precise value of neutron lifetime τ_n to an uncertainty less than 1 s plays a critical role in the Standard Model of nuclear and particle physics, as well as cosmology. The UCNτ experiment at Los Alamos National Lab uses a magneto-gravitational trap to store ultracold neutrons (UCNs) and an in-situ neutron detector to count the number of UCNs that have not decayed after prescribed holding times. The lifetime, τ_n ,is extracted from blind, independent analyses of the experimental data by either pairing adjacent short and long holding time runs or by performing a global likelihood fit of all runs. While systematic uncertainties are accounted for as corrections to the estimated lifetime, the understanding of the underlying UCN distribution and evolution in phase space is desired. The UCNτ datasets show a time dependence in the neutron counts for a given holding period due to changes such as the quality of the solid deuterium crystal used to produce the UCN and the spallation neutron source intensity. We present results from using long short term memory (LSTM) neural networks for time-dependent experimental neutron lifetime data analysis, with a goal of better understanding the variation and evolution of the number of UCN initially loaded into the UCNtau trap. This work opens doors to physics-informed machine learning to enhance UCN lifetime experiments.

LA-UR-23-26634   

A Priori Dimensionality Reduction for Quantum Dynamics

This work presents a powerful application in dimensionality reduction of the lesser-used Jacobi-Davidson algorithm for the generalized eigenvalue decomposition. When combined with matrix-free implementations of relevant operators, this technique allows for the computation of the dynamics of an arbitrary quantum state to be done in time proportional to the size of the original Hilbert space.   

Machine Learning based Particle Track Reconstruction for JLab Hall-A GEM Trackers

The Super BigBite Spectrometer (SBS) program of experiments aimed at nucleon form factor measurements is currently underway at Jefferson Lab. The program has yet to start its most demanding experiment, GEp-V, which is expected to encounter very high particle rates and backgrounds, reaching unprecedented levels compared to any other experiment in the world. The SBS program is using Gas Electron Multiplier (GEM) detectors for particle track reconstruction. The conventional tracking algorithms are likely unable to efficiently reconstruct tracks with these high data rates in the GEp experiment and other future high luminosity experiments, such as those in the Jefferson Lab SoLID project. A possible solution would be to employ a machine learning (ML) based approach. ML algorithms will be utilized to generate hit points along the detector planes using signal data from the GEM electronics. These candidate hit points will then be used to generate particle tracks by employing additional algorithms. The second part of the project is currently underway, involving the utilization of Graph Neural Networks (GNNs). The existing traditional algorithms are being used to develop supervised ML models, which will then be extended to incorporate standalone unsupervised learning models. The progress of this novel approach will be presented.