Estimating theoretical uncertainties of the two-nucleon observables by using backpropagation
K. Topolnicki, R. Skibiński, J. Golak · 2024
Read Estimating theoretical uncertainties of the two-nucleon observables by using backpropagation by K. Topolnicki, R. Skibiński, J. Golak (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-06304.txt.
Introduction
Topolnicki, Skibiński, and Golak present a 2024 arXiv paper on how to size the error in work with pairs of nucleons. Their method uses backpropagation, a tool that traces how a shift in each input changes a model output. They note that it works well when a model has many inputs but only a few outputs to check.
The team tests the method on a bound deuteron and on nucleon-nucleon scattering. With PyTorch in Python, they trace how each model setting shapes the binding energy and some phase-shift values, and they carry its error through. They also check the results by sampling the settings in a direct way. The two ways agree very well, and the tools can show how the functions that define the nuclear potentials add to the total error.
Catalog record
| CREATOR | K. Topolnicki, R. Skibiński, J. Golak |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-06304
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Estimating theoretical uncertainties of the two-nucleon observables by using backpropagation K. Topolnicki, R. Skibiński, J. Golak Abstract We present a novel approach to calculating theoretical uncertainties in few-nucleon calculations, making use of automatic differentiation via backpropagation, which is particularly efficient when there are many input variables but only a few outputs. The methods described in this paper constitute tools that can be used to investigate the properties of scalar functions used to define nuclear potentials and quantify their contribution to the uncertainty of few nucleon calculations. We demonstrate these methods in deuteron bound state and nucleon - nucleon scattering calculations. Backpropagation, implemented in the Python pytorch library, is used to calculate the gradients with respect to model parameters and propagate errors from these parameters to the deuteron binding energy and selected phase-shift parameters. The uncertainty values obtained using this approach are validated by directly sampling from the potential parameters. We find very good agreement between two ways of estimating that uncertainty. pacs 03.65.GeBound states, quantum mechanics and 24.10.-iNuclear reactions, models of 1 Introduction The progress of research on the nuclear potential, and in particular the increasing precision of scattering experiments, has drawn researchers’ attention to estimating the uncertainty of the theoretical models used [ 1 ] . In the first stage, the influence of experimental uncertainties on the determination of free potential parameters was examined. An important step in that direction was the preparation of a self-consistent database for nucleon-nucleon scattering data by the group from Granada [ 2 ] and using it to determine the uncertainty of the parameters of the OPE-Gausian [ 3 ] and other potentials along with the corresponding correlation coefficients. Also the introduction of new generations of few-nucleon forces from the Chiral Effective Field Theory, see e.g. reviews [ 4 , 5 , 6 ] , raises the important question of determining the errors associated with few-nucleon quantum mechanical calculations. The concept of determining the correlations between potential parameters was taken over by the Bochum group, which delivered uncertainties in the potential parameters and the resulting uncertainties in nucleon-nucleon phase shifts for the chiral SMS potential [ 7 ] . Using statistical analysis methods in [ 8 , 9 ] it was examined how these uncertainties are transferred to observables in elastic nucleon-deuteron scattering and in the deuteron breakup reaction. These methods involve calculating observables for various sets of potential parameters and then examining the obtained probability distributions for individual observables. These distributions depend on both the reaction energy and kinematic variables, e.g. scattering angles. It turned out that the uncertainties of the observables related to the uncertainties of the potential parameters are relatively small and usually remain below 1% for the differential cross section. Similar study but based on approximated solution of three-nucleon problem has been presented in [ 10 ] , where also other types of uncertainties are discussed. In [ 11 ] some uncertainty related to 3NF parameters have been estimated for the nucleon-deuteron scattering by using the approximated solutions of the Faddeev equations. The rapidly increasing computational complexity with the number of nucleons makes investigations based on statistical methods very challenging for many-nucleon systems. Recently new technique based on reduced-basis method has been developed to create and train emulators for many-nucleon systems, see [ 12 , 13 ] . It should be remembered that among other types of theoretical uncertainties related to the chiral potential two play a dominant role. The first is the uncertainty related to the value of the cut-off parameter used in the potential regularization. The second source of uncertainty is the neglect of higher terms of chiral expansion in the nuclear potential, which leads to the so-called truncation errors. Estimating the uncertainties associated with different cut-off values involves repeating the calculations, in practice for a few values (four in case of the SMS chiral potential [ 7 ] ). It should be remembered that the regulator value also affects the values of the free potential parameters, so in fact we are dealing with several versions of the potential. The uncertainty associated with using different cut-off values is typically of the order of 1-3% for nucleon-deuteron scattering and for deuteron and 3 He photodisintegration [ 14 ] . The estimation of truncation errors was proposed in [ 15 , 16 ] , which treated observables as quantities subject to chiral expansion. This recipe was then extended to many-nucleon processes in [ 17 ] . The truncation errors remain typically below 10% depending strongly on the order of chiral expansion and reaction energy for nucleon-deuteron scattering and for 2 H or 3 He photodisintegration [ 14 ] . It was shown that this prescription is consistent with the estimation of truncation errors resulting from Bayesian analysis methods with a particular choice of priors [ 18 , 19 ] . The role of various uncertainties in the nuclear structure calculations as well as various aspects of uncertainty quantification are currently the subject of many efforts, see e.g. [ 20 , 21 , 22 ] and references therein. Despite the fact, that both the cut-off dependence and truncation errors of two- and many-nucleon observables turned out to be usually bigger then the above mentioned dependence on potential parameters, in this work we focus on the uncertainty of two-nucleon observables arising from uncertainty of potential parameters. We demonstrate that software libraries designed for machine learning can be used to investigate these uncertainties. Using these tools to propagate errors from model parameters to observables is, to the best of our knowledge, a novel alternative to using standard statistical methods. Few nucleon potentials are typically expressed as a linear combination of isospin - spin - momentum operators and scalar functions. The form of the operators is fixed and the scalar functions effectively define the force. In practice, the possibility to investigate the properties of scalar functions that define nuclear interactions and quantify their contribution to the uncertainty of few nucleon calculations using the tools described in this paper is valuable in the development and debugging of a numerical implementation of these potentials. We use the Python library pytorch [ 23 ] to implement the numerical calculation of the deuteron bound state energy and nucleon-nucleon scattering observables. This library allows for an efficient implementation of numerical calculations. The numerical operations within this library are vectorized making it is possible to automatically parallelize their execution. Most importantly, pytorch offers the possibility to use backpropagation [ 24 ] to calculate the gradients of the results obtained from computations. This will be used to estimate errors of the bound state energy and nucleon - nucleon phase shifts that result from uncertainties in the two-nucleon (2N) potential. The theoretical introduction to the calculations is formulated in the language of “tensors". This term is not used to refer to the mathematical construct but to multidimensional arrays. Numerical operations mapped over multidimensional arrays are relatively straightforward to parallelize over multiple CPU threads on graphics processing units making “tensors" a popular fundamental data type in many python libraries designed for machine learning and numerical calculations. Formulating the calculations in this language makes them easier to translate to other libraries, for example tensorflow [ 25 ] , or if automatic differentiation is not