Importance Sampling and Adjoint Hybrid Methods in Monte Carlo Transport with Reflecting Boundaries

TL;DR

This paper introduces a simplified hybrid importance sampling method for Monte Carlo transport simulations with reflecting boundaries, focusing on geometries with limited volume scattering to achieve significant variance reduction and computational speed-up.

Contribution

It proposes a novel approach that approximates the adjoint solution by neglecting volume scattering, reducing computational costs while maintaining unbiased estimators.

Findings

Significant variance reduction in weak scattering geometries

Speed-up in Monte Carlo simulations demonstrated

Effective in remote sensing applications

Abstract

Adjoint methods form a class of importance sampling methods that are used to accelerate Monte Carlo (MC) simulations of transport equations. Ideally, adjoint methods allow for zero-variance MC estimators provided that the solution to an adjoint transport equation is known. Hybrid methods aim at (i) approximately solving the adjoint transport equation with a deterministic method; and (ii) use the solution to construct an unbiased MC sampling algorithm with low variance. The problem with this approach is that both steps can be prohibitively expensive. In this paper, we simplify steps (i) and (ii) by calculating only parts of the adjoint solution. More specifically, in a geometry with limited volume scattering and complicated reflection at the boundary, we consider the situation where the adjoint solution "neglects" volume scattering, whereby significantly reducing the degrees of freedom in steps (i) and (ii). A main application for such a geometry is in remote sensing of the environment using physics-based signal models. Volume scattering is then incorporated using an analog sampling algorithm (or more precisely a simple modification of analog sampling called a heuristic sampling algorithm) in order to obtain unbiased estimators. In geometries with weak volume scattering (with a domain of interest of size comparable to the transport mean free path), we demonstrate numerically significant variance reductions and speed-ups (figures of merit).