Hybrid Weight Window Techniques for Time-Dependent

Monte Carlo Neutronics · 2025

Read Hybrid Weight Window Techniques for Time-Dependent by Monte Carlo Neutronics (2025) 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-2501-06144.txt.

Introduction

Caleb Shaw and Dmitriy Anistratov automate a noise-cutting tool for time-changing neutron simulations, in a paper from two thousand twenty-five. Random-walk simulations waste effort where few particles go, so weight windows steer effort to where it counts. Their windows rest on a cheap solved model fed by second-moment terms from the random runs.

A filter cleans noise in those terms before the cheap model is solved. Two time-step schemes are offered for the cheap solve. On a standard test case the hybrid windows spread particles evenly in space. Wave fronts and low-flow zones then came out sharper. Readers in reactor math get a self-tuning way to save compute.

Catalog record

CREATORMonte Carlo Neutronics
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-06144

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hybrid Weight Window Techniques for Time-Dependent Monte Carlo Neutronics Abstract Efficient variance reduction of Monte Carlo simulations is desirable to avoid wasting computational resources. This paper presents an automated weight window algorithm for solving time-dependent particle transport problems. The weight window centers are defined by a hybrid forward solution of the discretized low-order second moment (LOSM) problem. The second-moment (SM) functionals defining the closure for the LOSM equations are computed by Monte Carlo solution. A filtering algorithm is applied to reduce noise in the SM functionals. The LOSM equations are discretized with first- and second-order time integration methods. We present numerical results of the AZURV1 benchmark. The hybrid weight windows lead to a uniform distribution of Monte Carlo particles in space. This causes a more accurate resolution of wave fronts and regions with relatively low flux. 1 Introduction Monte Carlo methods are used for solving time-dependent neutron transport problems due to their many advantages, such as large time steps, continuous energy physics and lack of discretization error. However, these methods converge relatively slowly with increasing number of particles. Analog Monte Carlo distributes particles inefficiently such that regions with sources end up oversampled, leaving other regions undersampled. This drives the development of variance reduction techniques such as implicit capture and weight windows [ 1 , 2 ] . Both of these techniques alter the weight of the Monte Carlo particles. The distribution of the Monte Carlo particles in phase-space can then be modified to improve efficiency. In this paper, we focus on the definition of weight windows using an auxiliary forward solution. This technique can obtain a uniform particle distribution and variance. It has also been shown that weight window centers can be efficiently defined by the numerical solution of hybrid moment equations such as the quasi-diffusion (aka variable Eddington factor) low-order equations [ 3 , 4 , 5 , 6 , 7 ] . We propose applying a different low-order problem that is used to formulate the second moment (SM) method [ 8 , 9 ] . The differential operator of the low-order SM (LOSM) equations is self-adjoint. This feature has advantages for solving multi-dimensional particle transport problems. To approximate LOSM equations in time we apply two time-integration scheme, Backward Euler, and Crank-Nicolson. A second-order finite volume scheme is used for spatial discretization of the LOSM equations. To reduce effects of noise on the hybrid LOSM problem, we apply a noise filtering technique to coefficients of SM closure terms and initial conditions at a timestep. The proposed automated weight windows defined by the hybrid solution of the LOSM equations improves the cell wise figure of merit. The developed automatic weight window algorithms have been implemented in the open source python-based dynamic neutron transport code MC/DC [ 10 ] . The remainder of the paper is organized as follows. In Section 2 we introduce the methodology used with a focus on the second moment method and the weight window definition. Section 3 presents the numerical results for the chosen test problem where we discuss the performance of the algorithm. Section 4 concludes with a summary of our findings and potential directions for future work. 2 Methodology 2.1 Second Moment Method The time dependent behavior of the ensemble of particles is described by the neutron transport equation. Here it is in 1D 1-group form with isotropic scattering and isotropic source: \frac{1}{v}\frac{\partial\psi}{\partial t}(x,\mu,t)+\mu\frac{\partial\psi}{\partial x}(x,\mu,t)+\Sigma_{t}(x,t)\psi(x,\mu,t)=\frac{1}{2}\big((\Sigma_{s}(x,t)+\nu_{f}\Sigma_{f}(x,t))\phi(x,t)+q(x,t)\big), (1) x\in[0,X],\text{ }\mu\in[-1,1],\text{ }t\geq 0, and boundary conditions (BCs) \psi(0,\mu,t)=\psi_{L}(\mu,t),\text{ for }\mu>0,\quad\psi(X,\mu,t)=\psi_{R}(\mu,t),\text{ for }\mu<0, and initial conditions (ICs) \psi(x,\mu,0)=\psi_{0}(x,\mu), where \psi is the particle angular flux, x is the particle spatial position, \mu is the directional cosine of particle motion, t is time, v is the particle speed, \Sigma_{t} , \Sigma_{s} , \Sigma_{f} are total, scattering, and fission cross sections, respectively, \nu_{f} is the expected number of neutrons born in fission, q is the external source of particles. The low-order equations are derived by integrating the transport equation (Eq. 1 ) with weights 1 and \mu over -1\leq\mu\leq 1 . The LOSM equations for the scalar flux, \phi(x,t)=\int_{-1}^{1}\psi(x,\mu,t)d\mu , and current, J(x,t)=\int_{-1}^{1}\mu\psi(x,\mu,t)d\mu , are given by [ 8 ] \frac{1}{v}\frac{\partial\phi}{\partial t}(x,t)+\frac{\partial J}{\partial x}(x,t)+\left(\Sigma_{t}(x,t)-\Sigma_{s}(x,t)-\nu_{f}\Sigma_{f}(x,t)\right)\phi(x,t)=q(x,t), (2) \frac{1}{v}\frac{\partial J}{\partial t}(x,t)+\frac{1}{3}\frac{\partial\phi}{\partial x}(x,t)+\Sigma_{t}(x,t)J(x,t)=\frac{\partial F}{\partial x}(x,t), (3) with the associated BCs J(0,t)=-\frac{1}{2}\phi(0,t)+2J_{L}(t)+P_{L}(t),\quad J(X,t)=\frac{1}{2}\phi(0,t)+2J_{R}(t)-P_{R}(t), (4) and ICs \phi(x,0)=\phi_{0}(x),\quad J(x,0)=J_{0}(x). The terms with the second angular moment of the high-order solution are closed with the function defined by F(x,t)=\int_{-1}^{1}\left(\frac{1}{3}-\mu^{2}\right)\psi(x,\mu,t)d\mu\,. (5) The BCs (Eq. 4 ) are formulated with P_{L}(t)=\int_{-1}^{1}\left(\frac{1}{2}-|\mu|\right)\psi(0,\mu,t)d\mu,\quad P_{R}(t)=\int_{-1}^{1}\left(\frac{1}{2}-|\mu|\right)\psi(X,\mu,t)d\mu\,. The LOSM equations are approximated in time using two different methods. We apply the Backward Euler (BE) time-integration scheme. Its stability properties are not dependent on the timestep size. The LOSM equations discretized with the BE scheme are given by \frac{1}{v\Delta t^{n}}\left(\phi^{n}-\phi^{n-1}\right)+\frac{dJ^{n}}{dx}+\left(\Sigma_{t}^{n}-\Sigma_{s}^{n}-\nu_{f}^{n}\Sigma_{f}^{n}\right)\phi^{n}=q^{n}, (6) \frac{1}{v\Delta t^{n}}\left(J^{n}-J^{n-1}\right)+\frac{1}{3}\frac{d\phi^{n}}{dx}+\Sigma_{t}^{n}J^{n}=\frac{dF^{n}}{dx}, (7) where n is the index of the time layer and e.g. \phi^{n}(x)=\phi^{n}(x,t^{n}) . Another temporal scheme for the LOSM equation is derived using a second-order Crank-Nicolson (CN) method. This scheme reduces effects of the numerical diffusion compared to the BE method and improves the solution accuracy for the same value of the timestep. The LOSM equations approximated by means of the CN scheme have the following form: \frac{1}{v\Delta t^{n}}\left(\phi^{n}-\phi^{n-1}\right)+\frac{1}{2}\left[\frac{dJ^{n}}{dx}+(\Sigma_{t}^{n}-\Sigma_{s}^{n}-\nu_{f}^{n}\Sigma_{f}^{n})\phi^{n}\right]=\\ \frac{1}{2}(q^{n}+q^{n-1})-\frac{1}{2}\left[\frac{dJ^{n-1}}{dx}+(\Sigma_{t}^{n-1}-\Sigma_{s}^{n-1}-\nu_{f}^{n-1}\Sigma_{f}^{n-1})\phi^{n-1}\right], (8) \frac{1}{v\Delta t^{n}}\left(J^{n}-J^{n-1}\right)+\frac{1}{2}\left[\frac{1}{3}\frac{d\phi^{n}}{dx}+\Sigma_{t}^{n}J^{n}\right]=\frac{1}{2}\left[\frac{dF^{n}}{dx}+\frac{dF^{n-1}}{dx}\right]-\frac{1}{2}\left[\frac{1}{3}\frac{d\phi^{n-1}}{dx}+\Sigma_{t}^{n-1}J^{n-1}\right]. (9) To spatially discretize the LOSM equations we use a second order finite volume scheme. We define a spatial mesh \{x_{i-1/2}\}_{i=0}^{I} . The semi-discretized balance equation (Eqs. ( 6 ) and ( 8 )) is integrated over the i^{th} spatial [x_{i-1/2},x_{i+1/2}] . The semi-discretized first-moment equation (Eqs. ( 7 ) and ( 9 )) is integrated over halves of spatial cells. The resulting discretized equations are defined for cell averaged fluxes, \phi_{i}^{n}~=~\frac{1}{\Delta x_{i}}\int_{x_{i-1/2}}^{x_{i+1/2}}\phi^{n}(x)dx , and cell-edge currents, J_{i\pm 1/2}^{n}=J^{n}(x_{i\pm 1/2}) , and are given by J_{i+1/2}^{n}-J_{i-1/2}^{n}+(\hat{\Sigma}_{t,i}^{n}-\Sigma_{s,i}^{n}-\nu_{f,i}^{n}\Sigma_{f,i}^{n})\Delta x_{i}\phi_{i}^{n}=Q_{0,i}^{n}\Delta x_{i}\,, (10) \frac{1}{3}\Big(\phi_{i}^{n}-\phi_{i-1}^{n}\Big)+\hat{\Sigma}_{t,i-1/2}^{n}\Delta x_{i-1/2}J_{i-1/2}^{n}=Q_{1,i-1/2}^{n}\,, (11) \hat{\Sigma}_{t,i}^{n}=\begin{cases}\Sigma_{t,i}^{n}+\frac{1}{v\Delta t^{n}},\quad\text{for BE},\\ \Sigma_{t,i}^{n}+\frac{2}{v\Delta t^{n}},\quad\text{for CN},\\ \end{cases}\quad\hat{\Sigma}_{t,i-1/2}^{n}=\frac{\hat{\Sigma}_{t,i-1}^{n}\Delta x_{i-1}+\hat{\Sigma}_{t,i}^{n}\Delta x_{i}}{\Delta x_{i-1}+\Delta x_{i}}\,, \Delta x_{i}=x_{i+1/2}-x_{i-1/2}\,,\quad\Delta x_{i-1/2}=\frac{1}{2}(\Delta x_{i+1}+\Delta x_{i})\,,\quad i=1,...,N_{x}\,, and the right-hand sides are defined by \displaystyle Q {}_{0,i}^{n}=\begin{cases}q_{i}^{n}+\frac{\phi_{i}^{n-1}}{v\Delta t^{n}},\quad\text{for BE,}\\ q_{i}^{n}+q_{i}^{n-1}+\frac{1}{\Delta x_{i}}(J_{i-1/2}^{n-1}-J^{n-1}_{i+1/2})+\Big(\Sigma_{s,i}^{n-1}+\nu_{f,i}^{n-1}\Sigma_{f,i}^{n-1}+\frac{2}{v\Delta t^{n}}-\Sigma_{t,i}^{n-1}\Big)\phi_{i}^{n-1},\quad\text{for CN},\end{cases} \displaystyle Q {}_{1,i-1/2}^{n}= \displaystyle\begin{cases}\frac{J_{i-1/2}^{n-1}}{v\Delta t^{n}},\quad\text{for BE,}\\ \frac{1}{\Delta x_{i-1/2}}\!\Big(\!F_{i}^{n-1}\!-\!F_{i-1}^{n-1}\!+\!\frac{1}{3}(\phi_{i-1}^{n-1}\!-\!\phi_{i}^{n-1})\!\Big)\!+\!\Big(\!\frac{2}{v\Delta t^{n}}\!-\!\frac{\Sigma_{t,i-1}^{n-1}\Delta x_{i-1}+\Sigma^{n-1}_{t,i}\Delta x_{i}}{\Delta x_{i-1}+\Delta x_{i}}\!\Big)J_{i-1/2}^{n-1},\ \text{for CN},\end{cases} where F_{i}^{n}=\frac{1}{\Delta x_{i}}\int_{x_{i-1/2}}^{x_{i+1/2}}F^{n}(x)dx\,. (12) The space-time integration scheme applied to discretize the LOSM equations defines the grid functions of the solution and closure terms in spatial cells and at time layers. At each timestep, these key quantities are computed by Monte Carlo algorithms using an ensemble of particle histories. The scalar flux \phi_{i}^{n-1} , and the functionals for closure terms F_{i}^{n} and F_{i}^{n-1} are cell averages computed at discrete temporal points rather than time averages. The current J_{i\pm 1/2}^{n-1} is similarly computed along the cell edges at the time layer. Monte Carlo simulations use track-length tallies to numerically integrate quantities with greater statistical reliability than census or crossing tallies. Track length tallies capture information by accumulating data along particle paths within cells, whereas census and crossing tallies rely on particles crossing a temporal boundary or spatial face. However track-length tallies produce time-averaged quantities, which do not directly provide the values at discrete time points for the LOSM equations. There exist numerical techniques which use a second order backward difference (BDF-2) formula to extrapolate quantities obtained from Monte Carlo photon transport tallies [ 11 ] . In the developed hybrid algorithm, we use the particle locations at the time layer to compute cell averaged \phi_{i}^{n} and F_{i}^{n} . This

Read the complete text