Implicit Monte Carlo 1

By 0x7df, Thu 15 October 2020, modified Thu 15 October 2020, in category Misc


Implicit Monte Carlo (IMC) is a Monte Carlo method for solving the thermal radiation transport equation, published by Fleck and Cummings in 1971 , according to which:

The method is based upon the concept of effective scattering, wherein a fraction of the radiative energy absorbed is instantaneously and isotropically reradiated in a manner analogous to a scattering process.

The modelled particles in IMC are not photons, but represent 'packets' of photons, whose total energy changes as they interact with the background material in a way that captures the local energy absorption and radiation rates.

One benefit of this 'effective scattering' is variance reduction. Because particles are never absorbed, but simply scatter around until they leave the system (losing energy gradually, to represent absorption), each particle contributes to the solution over a large area of the problem space. If particles were absorbed directly, in accordance with the material opacities, they would contribute only locally to the solution, and many more particles would have to be simulated to get a good, low-variance solution everywhere.

This, however, is possible to achieve in 'standard', as opposed to 'implicit' Monte Carlo methods. The defining aspect is the ability of IMC to better treat the non-linear behaviour of thermal radiation transport, whereby the materials' opacities to radiation and the emission rates are driven by their temperatures, which are of course driven in turn by the radiation field. This non-linearity can be dealt with explicitly, by computing a radiation time-step with constant temperatures and material properties, and then updating the temperatures and opacities at the end of the time-step according to the new radiation fluxes. However, in optically-thick systems, where the radiation field is tightly- coupled with the material, and the material and radiation are in or close to thermal equilibrium, a very short time-step is required that makes the whole process very inefficient. IMC modifies this process by treating only a fraction of the total radiation emission term as an explicit source term, fixed throughout the time-step, and deals with the remainder of the energy emission by way of the 'effective scattering' mechanism: an effective scattering term is introduced which represents the absorption and re-emission (isotropically) of radiation.

To understand this better, consider a computational cell which radiation is entering from the left side (e.g. a plane Marshak wave advancing through a 1D slab). The cell has a fixed temperature and therefore fixed opacities and absorption and emission rates. In a standard Monte Carlo scheme, particles entering from the left boundary are preferentially absorbed in the left part of the cell (in optically thick problems the cells are by definition many mean-free-paths wide). However the emission rate is constant throughout the whole cell, so in effect the emission process is unphysically decoupled from the absorption. In Implicit Monte Carlo, a fraction of both the absorption and emission processes are taken care of through the mechanism of effective scattering, so are very tightly coupled. Effectively, the IMC method allows, to a certain extent, sub-grid-scale information about the absorption and re- emission processes to be accounted for.

Discretisation of the energy equation

Mathematically, IMC is simply the Monte Carlo solution of a time-dependent transport equation which has been implicitly discretised in time, which yields unconditional stability with respect to time-step length, just as an implicitly-discretised finite-difference scheme is unconditionally stable. (The implicitly-discretised equations derived by Fleck and Cummings do not, in fact, need to be solved by the Monte Carlo method to yield this key benefit.)

Consider the basic finite differencing of a simple differential equation

$$ \frac{\partial u}{\partial t} = f{\left(u\right)} $$

An explicit time discretisation is:

$$ u^{n+1} \approx u^n + \Delta t \; f{\left(u^n\right)} $$

and an implicit one is:

$$ u^{n+1} \approx u^n + \Delta t \; f{\left(u^{n+1}\right)} $$

In the implicit scheme, the right-hand-side is evaluated using the end-of-time step values (time level \(n+1\)), the explicit scheme using the values from time level \(n\). A scheme can also be constructed that mixes the two:

$$ u^{n+1} \approx u^n + \Delta t \left[ \alpha \: f{\left(u^{n+1}\right)} + \left(1 - \alpha\right) f{\left(u^n\right)} \right] $$

(with which \(\alpha = 1/2\) yields the well-known Crank-Nicholson scheme). Similarly, in IMC, the radiation energy density equation:

$$ \frac{\partial u_r}{\partial t} = \beta \sigma \left( \phi - cu_r \right) $$

is discretised as:

$$ u_r^{n+1} \approx u_r^n + \Delta t \hat\beta \hat\sigma \left( \phi^\lambda - c\left[ \alpha u_r^{n+1} + \left(1-\alpha\right)u_r^n \right] \right) $$

The value of \(u_r\), the radiation energy density, on the right-hand side has been replaced with an implicit difference \(\alpha u_r^{n+1} + \left(1-\alpha\right) u_r^n\). The value of \(\alpha\) can be left as a free parameter, chosen according to the specifics of the problem to be solved.

The other time-dependent quantities on the RHS are:

In the discretised equation, \(\beta\) and \(\sigma\) have been substituted with values averaged over the time-step, and \(\phi\) by a value \(\phi^\lambda\), where the superscript \(\lambda\) denotes an as-yet unspecified time centering.

Discretisation of the transport equation

Similarly, the transport equation:

$$ \frac{1}{c} \frac{\partial I}{\partial t} + \mu \frac{\partial I}{\partial x} + \sigma I = \frac{1}{2}c\sigma u_r $$

is discretised as:

$$ \frac{1}{c} \frac{\partial I}{\partial t} + \mu \frac{\partial I}{\partial x} + \sigma I = \frac{1}{2}c\sigma \left[\alpha u^{n+1}_r + \left(1-\alpha\right) u^n_r\right] $$

If the discretised energy equation is solved for \(u^{n+1}_r\), it is then possible to write an equation for \(\alpha u^{n+1}_r + \left(1-\alpha\right)u^n_r\), which can be substituted into the transport equation. That is:

$$ u_r^{n+1} \approx u_r^n + \Delta t \hat\beta \hat\sigma \left( \phi^\lambda - c\left[ \alpha u_r^{n+1} + \left(1-\alpha\right)u_r^n \right] \right) $$

is rearranged to give:

$$ u_r^{n+1} \approx \left[ \frac{1-\left(1-\alpha\right)\hat\beta c\Delta t\hat\sigma}{1+\alpha\hat\beta c\Delta t\hat\sigma} \right] u^n_r + \left[ \frac{\hat\beta\hat\sigma\Delta t}{1+\alpha\hat\beta c\Delta t\hat\sigma} \right] \phi^\lambda $$

which allows us to write:

$$ \alpha u^{n+1}_r + \left(1-\alpha\right) u^n_r = \frac{\alpha\hat\beta\hat\sigma\Delta t}{1 + \alpha\hat\beta c\Delta t\hat\sigma} \phi^\lambda + \frac{1}{1 + \alpha\hat\beta c\Delta t\hat\sigma} u^n_r $$

Substituting this into the transport equation gives:

$$ \frac{1}{c} \frac{\partial I}{\partial t} + \mu \frac{\partial I}{\partial x} + \sigma I = \frac{1}{2}\sigma \left[ \frac{\alpha\hat\beta c\hat\sigma\Delta t}{1 + \alpha\hat\beta c\Delta t\hat\sigma} \phi^\lambda + \frac{c}{1 + \alpha\hat\beta c\Delta t\hat\sigma} u^n_r\right] $$

which we can re-write as:

$$ \frac{1}{c} \frac{\partial I}{\partial t} + \mu \frac{\partial I}{\partial x} + \sigma I = \frac{1}{2} \sigma_s \phi^\lambda + \frac{1}{2} c\sigma_a u^n_r $$

where the first term on the right is the equivalent of a scattering term with cross-section \(\sigma_s\) given by:

$$ \sigma_s = \frac{\alpha\hat\beta c\hat\sigma\Delta t}{1 + \alpha\hat\beta c\Delta t\hat\sigma}\sigma $$

and the second term on the right is the 'remaining' source term, where \(\sigma_a\) is the absorption cross-section:

$$ \sigma_a = \frac{1}{1 + \alpha\hat\beta c\Delta t\hat\sigma} \sigma $$

derived from the fact that \(\sigma = \sigma_s + \sigma_a\).

Effectively, a fraction of the original emission term on the RHS as been re-cast as a scattering term, leaving a modified direct emission term.


2020-10-17T23:49:17.368Z | 0x7df:


Add comment