American Nuclear Society
Home

Home / Store / Journals / Electronic Articles / Nuclear Technology / Volume 168 / Number 3 / Pages 793-798

On a New Variance Reduction Technique: Neural Network Biasing - A Study of Two Test Cases with the Monte Carlo Code Tripoli4

Eric Dumonteil

Nuclear Technology / Volume 168 / Number 3 / Pages 793-798

December 2009

Format:

Price:$30.00
Member Price:$27.00
Member Savings:$3.00

Various variance-reduction techniques are used in Monte Carlo particle transport. Most of them rely either on a hypothesis made by the user (parameters of the exponential biasing, mesh and weight bounds for weight windows, etc.) or on a previous calculation of the system with, for example, a deterministic solver. This paper deals with a new acceleration technique, namely, autoadaptative neural network biasing. Indeed, instead of using any a priori knowledge of the system, it is possible, at a given point in a simulation, to use the Monte Carlo histories previously simulated to train a neural network, which, in return, should be able to provide an estimation of the adjoint flux, used then for biasing the simulation. We will describe this method, detail its implementation in the Monte Carlo code Tripoli4, and discuss its results on two test cases.

 
 
 
Questions or comments about the site? Contact the ANS Webmaster.
advertisement